CA2408819C - Machine translation techniques - Google Patents
Machine translation techniques Download PDFInfo
- Publication number
- CA2408819C CA2408819C CA002408819A CA2408819A CA2408819C CA 2408819 C CA2408819 C CA 2408819C CA 002408819 A CA002408819 A CA 002408819A CA 2408819 A CA2408819 A CA 2408819A CA 2408819 C CA2408819 C CA 2408819C
- Authority
- CA
- Canada
- Prior art keywords
- translation
- target language
- word
- current target
- tree
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Lifetime
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/34—Browsing; Visualisation therefor
- G06F16/345—Summarisation for human users
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/253—Grammatical analysis; Style critique
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/40—Processing or translation of natural language
- G06F40/42—Data-driven translation
- G06F40/44—Statistical methods, e.g. probability models
Abstract
Machine translation decoding is accomplished by receiving as input a text segment in a source language to be translated into a target language, generating an initial translation as a current target language translation, applying one or more modification operators to the current target language translation to generate one or more modified target language translations, determining whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, setting a modified target language translation as the current target language translation, and repeating these steps until occurrence of a termination condition. Automatically generating a tree (e. g., either a syntactic tree or a discourse tree) can be accomplished by receiving as input a tree corresponding to a source language text segment, and applying one or more decision rules to the received input to generate a tree corresponding to a target language text segment.
Description
TITLE
MACHINE TRANSLATION TECHNIQUES
Field of the Invention The present application relates to computational linguistics and more particularly to machine translation techniques. More specifically, the present application describes techniques for performing decoding of a source text segment into a target text segment and for rewriting trees from a first linguistic space into another linguistic space.
Background and Summary Machine translation (MT) is the automatic translation, for example, using a computer system, from a first language (e. g., French) into another language (e. g., English).
Systems that perform MT techniques are said to "decode" the source language into the target language. From the end-user's perspective, the MT process is relatively straight-forward. As shown in Fig. 1A, the MT 102 receives as input a source sentence 100, for example, in French (e.g., "ce ne est pas juste"), and after processing the input sentence, outputs the equivalent decoded sentence in the target language - in this example, English ("it is not fair").
One type of conventional MT decoder is the "stack decoder" such as described in U.S. Patent No. 5,477,451 (Brown et al.), entitled "Method and System for Natural Language Translation." In a stack decoder, the universe of possible translations are organized into a graph structure and then exhaustively searched until an optimal solution (translation) is found. Although stack decoders tend to produce good results, they do so at a significant cost -namely, maintaining and searching a large potential solution space such as used by stack decoders is expensive, both computationally and space-wise (e.g., in terms of computer memory). Accordingly, the present inventor recognized that an iterative, incremental decoding technique could produce optimal, or near optimal, results while considerably reducing the computational and space requirements. This decoder is referred to herein as a "greedy" decoder or, equivalently, as a "fast decoder."
The term "greedy" refers to techniques that produce solutions based on myopic optimization - that is, given a partial solution, produce as a next estimate a new solution that improves the objective the most. Put another way, a greedy algorithm typically starts out with an approximate solution and then tries to improve it incrementally until a satisfactory solution is reached.
Implementations of the greedy decoder may include various combinations of the following features.
In one aspect, machine translation (MT) decoding involves receiving as input a text segment (e. g., a clause, a sentence, a paragraph or a treatise) in a source language to be translated into a target language, generating an initial translation (e.g., either a word-for-word or phrase-for-phrase gloss) as a current target language translation, applying one or more modification operators to the current target language translation to generate one or more modified target language translations, determining whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, setting a modified target language translation as the current target language translation, and repeating these steps until occurrence of a termination condition.
Applying one or more modification operators may involve changing the translation of one or two words in the current target language translation the translation.
Alternatively, or in addition, applying one or more modification operators may include (i) changing a translation of a word in the current target language translation and concurrently (ii) inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation. The inserted other word may have a high probability of having a zero-value fertility.
Applying one or more modification operators may include deleting from the current target language translation a word having a zero-value fertility; and/or modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and/or modifying an alignment between the source language text segment and the current target language translation by (i) eliminating a target language word from the current target language translation and (ii) linking words in the source language text segment.
In various embodiments, applying modification operators may include applying two or more of the following: (i) changing the translation of one or two words in the current target language translation; (ii) changing a translation of a word in the current target language translation and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word having a high probability of having a zero-value fertility; (iii) deleting from the current target language translation a word having a zero-value fertility; (iv) modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and/or (v) modifying an alignment between the source language text segment and the current target language translation by eliminating a target language word from the current target language translation and linking words in the source language text segment.
Determining whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation may include calculating a probability of correctness for each of the modified target language translations.
The termination condition may include a determination that a probability of correctness of a modified target language translation is no greater than a probability of correctness of the current target language translation.
The termination condition may be the occurrence of a completion of a predetermined number of iterations and/or the lapse of a predetermined amount of time.
In another aspect, a computer-implemented machine translation decoding method may, for example, implement a greedy decoding algorithm that iteratively modifies a target language translation of a source language text segment (e.g., a clause, a sentence, a paragraph, or a treatise) until an occurrence of a termination condition (e. g., completion of a predetermined number of iterations, lapse of a predetermined period of time, and/or a determination that a probability of correctness of a modified translation is no greater than a probability of correctness of a previous translation.) The MT decoding method may start with an approximate target language translation and iteratively improve the translation with each successive iteration. The approximate target language translation may be, for example, a word-for-word or phrase-for-phrase gloss, or the approximate target language translation may be a predetermined translation selected from among a plurality of predetermined translations.
Iteratively modifying the translation may include incrementally improving the translation with each iteration, for example, by applying one or more modification operations on the translation.
The one or more modification operations comprises one or more of the following operations: (i) changing one or two words in the translation; (ii) changing a translation of a word and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the translation, the inserted other word having a high probability of having a zero-value fertility; (iii) deleting from the translation a word having a zero-value fertility; (iv) modifying an alignment between the source language text segment and the translation by swapping non-overlapping target language word segments in the translation; and (v) modifying an alignment between the source language text segment and the translation by eliminating a target language word from the translation and linking words in the source language text segment.
In another aspect, a machine translation decoder may include a decoding engine comprising one or more modification operators to be applied to a current target language translation to generate one or more modified target language translations; and a process loop to iteratively modify the current target language translation using the one or more modification operators. The process loop may terminate upon occurrence of a termination condition. The process loop may control the decoding engine to incrementally improve the current target language translation with each iteration.
The MT decoder may further include a module (including, for example, a language model and a translation model) for determining a probability of correctness for a translation.
The process loop may terminate upon a determination that a probability of correctness of a modified translation is no greater than a probability of correctness of a previous translation, and/or upon completion of a predetermined number of iterations; and/or after lapse of a predetermined period of time.
One or more of the following advantages may be provided by the greedy decoder as described herein. The techniques and methods described here may result in a MT
decoder that performs with high accuracy, high speed and relatively low computational and space costs. The greedy decoder can be modified as desired to perform a full set of sentence modification operations or any subset thereof.
MACHINE TRANSLATION TECHNIQUES
Field of the Invention The present application relates to computational linguistics and more particularly to machine translation techniques. More specifically, the present application describes techniques for performing decoding of a source text segment into a target text segment and for rewriting trees from a first linguistic space into another linguistic space.
Background and Summary Machine translation (MT) is the automatic translation, for example, using a computer system, from a first language (e. g., French) into another language (e. g., English).
Systems that perform MT techniques are said to "decode" the source language into the target language. From the end-user's perspective, the MT process is relatively straight-forward. As shown in Fig. 1A, the MT 102 receives as input a source sentence 100, for example, in French (e.g., "ce ne est pas juste"), and after processing the input sentence, outputs the equivalent decoded sentence in the target language - in this example, English ("it is not fair").
One type of conventional MT decoder is the "stack decoder" such as described in U.S. Patent No. 5,477,451 (Brown et al.), entitled "Method and System for Natural Language Translation." In a stack decoder, the universe of possible translations are organized into a graph structure and then exhaustively searched until an optimal solution (translation) is found. Although stack decoders tend to produce good results, they do so at a significant cost -namely, maintaining and searching a large potential solution space such as used by stack decoders is expensive, both computationally and space-wise (e.g., in terms of computer memory). Accordingly, the present inventor recognized that an iterative, incremental decoding technique could produce optimal, or near optimal, results while considerably reducing the computational and space requirements. This decoder is referred to herein as a "greedy" decoder or, equivalently, as a "fast decoder."
The term "greedy" refers to techniques that produce solutions based on myopic optimization - that is, given a partial solution, produce as a next estimate a new solution that improves the objective the most. Put another way, a greedy algorithm typically starts out with an approximate solution and then tries to improve it incrementally until a satisfactory solution is reached.
Implementations of the greedy decoder may include various combinations of the following features.
In one aspect, machine translation (MT) decoding involves receiving as input a text segment (e. g., a clause, a sentence, a paragraph or a treatise) in a source language to be translated into a target language, generating an initial translation (e.g., either a word-for-word or phrase-for-phrase gloss) as a current target language translation, applying one or more modification operators to the current target language translation to generate one or more modified target language translations, determining whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, setting a modified target language translation as the current target language translation, and repeating these steps until occurrence of a termination condition.
Applying one or more modification operators may involve changing the translation of one or two words in the current target language translation the translation.
Alternatively, or in addition, applying one or more modification operators may include (i) changing a translation of a word in the current target language translation and concurrently (ii) inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation. The inserted other word may have a high probability of having a zero-value fertility.
Applying one or more modification operators may include deleting from the current target language translation a word having a zero-value fertility; and/or modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and/or modifying an alignment between the source language text segment and the current target language translation by (i) eliminating a target language word from the current target language translation and (ii) linking words in the source language text segment.
In various embodiments, applying modification operators may include applying two or more of the following: (i) changing the translation of one or two words in the current target language translation; (ii) changing a translation of a word in the current target language translation and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word having a high probability of having a zero-value fertility; (iii) deleting from the current target language translation a word having a zero-value fertility; (iv) modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and/or (v) modifying an alignment between the source language text segment and the current target language translation by eliminating a target language word from the current target language translation and linking words in the source language text segment.
Determining whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation may include calculating a probability of correctness for each of the modified target language translations.
The termination condition may include a determination that a probability of correctness of a modified target language translation is no greater than a probability of correctness of the current target language translation.
The termination condition may be the occurrence of a completion of a predetermined number of iterations and/or the lapse of a predetermined amount of time.
In another aspect, a computer-implemented machine translation decoding method may, for example, implement a greedy decoding algorithm that iteratively modifies a target language translation of a source language text segment (e.g., a clause, a sentence, a paragraph, or a treatise) until an occurrence of a termination condition (e. g., completion of a predetermined number of iterations, lapse of a predetermined period of time, and/or a determination that a probability of correctness of a modified translation is no greater than a probability of correctness of a previous translation.) The MT decoding method may start with an approximate target language translation and iteratively improve the translation with each successive iteration. The approximate target language translation may be, for example, a word-for-word or phrase-for-phrase gloss, or the approximate target language translation may be a predetermined translation selected from among a plurality of predetermined translations.
Iteratively modifying the translation may include incrementally improving the translation with each iteration, for example, by applying one or more modification operations on the translation.
The one or more modification operations comprises one or more of the following operations: (i) changing one or two words in the translation; (ii) changing a translation of a word and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the translation, the inserted other word having a high probability of having a zero-value fertility; (iii) deleting from the translation a word having a zero-value fertility; (iv) modifying an alignment between the source language text segment and the translation by swapping non-overlapping target language word segments in the translation; and (v) modifying an alignment between the source language text segment and the translation by eliminating a target language word from the translation and linking words in the source language text segment.
In another aspect, a machine translation decoder may include a decoding engine comprising one or more modification operators to be applied to a current target language translation to generate one or more modified target language translations; and a process loop to iteratively modify the current target language translation using the one or more modification operators. The process loop may terminate upon occurrence of a termination condition. The process loop may control the decoding engine to incrementally improve the current target language translation with each iteration.
The MT decoder may further include a module (including, for example, a language model and a translation model) for determining a probability of correctness for a translation.
The process loop may terminate upon a determination that a probability of correctness of a modified translation is no greater than a probability of correctness of a previous translation, and/or upon completion of a predetermined number of iterations; and/or after lapse of a predetermined period of time.
One or more of the following advantages may be provided by the greedy decoder as described herein. The techniques and methods described here may result in a MT
decoder that performs with high accuracy, high speed and relatively low computational and space costs. The greedy decoder can be modified as desired to perform a full set of sentence modification operations or any subset thereof.
This gives a system designer and/or end-user considerable flexibility to tailor the decoder's speed, accuracy and/or other performance characteristics to match desired objectives or constraints. The use of a set of basic modification operations, each able to be used as a standalone operator or in conjunction with the others, further enhances this flexibility. Moreover, the use of independent standalone operators as constituents of the decoding engine makes the decoder extensible and scalable.
That is, different or additional modification operators can be used to suit the objectives or constraints of the system designer and/or end-user.
In conjunction with MT research and related areas in computational linguistics, researchers have developed and frequently use various types of tree structures to graphically represent the structure of a text segment (e. g., clause, sentence, paragraph or entire treatise).
Two basic tree types include (1) the syntactic tree, which can be used to graphically represent the syntactic relations among components of a text segment, and (2) the rhetorical tree (equivalently, the rhetorical structure tree (RST) or the discourse tree), which can be used to graph the rhetorical relationships among components of a text segment. Rhetorical structure trees (also referred to as discourse trees) are discussed in detail in William C.
That is, different or additional modification operators can be used to suit the objectives or constraints of the system designer and/or end-user.
In conjunction with MT research and related areas in computational linguistics, researchers have developed and frequently use various types of tree structures to graphically represent the structure of a text segment (e. g., clause, sentence, paragraph or entire treatise).
Two basic tree types include (1) the syntactic tree, which can be used to graphically represent the syntactic relations among components of a text segment, and (2) the rhetorical tree (equivalently, the rhetorical structure tree (RST) or the discourse tree), which can be used to graph the rhetorical relationships among components of a text segment. Rhetorical structure trees (also referred to as discourse trees) are discussed in detail in William C.
Mann and Sandra A. Thompson, "Rhetorical structure theory:
Toward a functional theory of text organization," Text, 8 (3) :243-X81 (1988) .
The example shown in Fig. 6 illustrates the types of structures that may be present in a rhetorical structure tree for a text fragment. The leaves of the tree correspond to elementary discourse units ("edus") and the internal nodes correspond to contiguous text spans. Each node in a rhetorical structure tree is characterized by a "status" (i.e., either "nucleus" or "satellite") and a "rhetorical relation," which is a relation that holds between two non-overlapping text spans. In Fig. 6, nuclei are represented by straight lines while satellites are represented by arcs.
The present inventor recognized that significant differences exist between the rhetorical structures of translations of a text in different languages (e. g., Japanese and English). Accordingly, to improve MT quality, and as a component of a larger MT system, the present inventor developed techniques for automatically rewriting (e. g., using a computer system) rhetorical structures from one language into another, for example, rewriting a rhetorical tree for a text segment in Japanese into a rhetorical tree for a counterpart text segment in English.
Implementations of the disclosed tree rewriting techniques may include various combinations of the following features.
In one aspect, automatically generating a tree (e. g., 5~ either a syntactic tree or a discourse tree) involves receiving as input a tree corresponding to a source language text segment, and applying one or more decision rules to the received input to generate a tree corresponding to a target language text segment.
In another aspect, a computer-implemented tree generation method may include receiving as input a tree corresponding to a source language text segment (e. g., clause, sentence, paragraph, or treatise), and applying one or more decision rules (e. g., a sequence of decision rules that collectively represent a transfer function) to the received input to generate a tree corresponding to a target language text segment, which potentially may be a different type of text segment.
The tree generation method further may include automatically determining the one or more decision rules based on a training set, for example, a plurality of input -output tree pairs and a mapping between each of the input-output tree pairs. The mapping between each of the input-output tree pairs may be a mapping between leaves of the input tree and leaves of the paired output tree. Mappings between leaves of input-output tree pairs can be one-to-one, one-to-many, many-to-one, or many-to-many.
Automatically determining the one or more decision rules may include determining a sequence of operations that generates an output tree when applied to the paired input tree. Determining a sequence of operations may include using a plurality of predefined operations that collectively are sufficient to render any input tree into the input tree's paired output tree. The plurality of predefined operations comprise one or more of the following: a shift operation that transfers an elementary discourse tree (edt) from an input list into a stack; a reduce operation that pops two edts from a top of the stack, combines the two popped edts into a new tree, and pushes the new tree on the top of the stack; a break operation that breaks an edt into a predetermined number of units a create-next operation that creates a target language discourse constituent that has no correspondent in the source language tree; a fuse operation that fuses an edt at the top of the stack into the preceding edt; a swap operation that swaps positions of edts in the input list;
and an assignType operation that assigns one or more of the following types to edts: Unit, Multiunit, Sentence, Paragraph, MultiParagraph, and Text.
The plurality of predefined operations may represent a closed set that includes the shift operation, the reduce operation, the break operation, the create-next operation, the fuse operation, the swap operation and the assignType operation.
Determining a sequence of operations may result in a plurality of learning cases, one learning case for each input-output tree pair. In that case, the tree generation method may further include associating one or more features with each of the plurality of learning cases based on context. The associated features may include one or more of the following: operational and discourse features, correspondence-based features, and lexical features.
The tree generation method may further include applying a learning program (e. g., C4.5) to the plurality of learning cases to generate the one or more decision rules In another aspect, a computer-implemented tree generation module may include a predetermined set of decision rules that, when applied to a tree (e. g., syntactic or discourse) corresponding to a source language text segment, generate a tree corresponding to a target language text segment. The predetermined set of decision rules may define a transfer function between source language trees and target language trees.
In another aspect, determining a transfer function between trees (e. g., syntactic or discourse) of different types may include generating a training set comprising a plurality of tree pairs and a mapping between each tree pair, each tree pair comprises a source tree and a corresponding target tree, and generating a plurality of learning cases by determining, for each tree pair, a sequence of operations that result in the target tree when applied to the source tree; and generating a plurality of decision rules by applying a learning algorithm to the plurality of learning cases.
Determining a transfer function between trees of different types further may include, prior to generating the plurality of decision rules, associating one or more features with each of the learning cases based on context.
In another aspect, a computer-implemented discourse-based machine translation system may include a discourse parser that parses the discourse structure of a source language text segment and generates a source language discourse tree for the text segments a discourse-structure transfer module that accepts the source language discourse tree as input and generates as output a target language discourse tree; and a mapping module that maps the target language discourse tree into a target text segment. The discourse-structure transfer module may include a plurality of decision rules generated from a training set of source language-target language tree pairs.
One or more of the following advantages may be provided by tree rewriting as described herein, The techniques and methods described here may result in a tree rewriting capability that allows users (e. g., human end-users such as linguistic researchers or computer processes such as MT systems) to automatically have a tree for a text segment in a source language rewritten, or translated, into a tree for the text segment translated into a target language. This functionality is useful both in its standalone form and as a component of a larger system, such as in a discourse-based machine translation system.
Moreover, because the tree rewriter described here automatically learns how t,o rewrite trees from one language into another, the system is easy and convenient to use.
The mapping scheme used in training the tree rewriter also provides several advantages. For example, by allowing any arbitrary groupings (e. g., one-to-one, one-to-many, many-to-one, many-to-many) between leaves in th,e source and target trees, the flexibility, richness and robustness of the resulting mappings are enhanced.
The enhanced shift-reduce operations used in training the tree rewriter also provide several advantages. For example, the set of basic operations that collectively are sufficient to render any input tree into its paired output tree provides a powerful yet compact tool for rewriting tree structures.
According to one aspect of the present invention, there is provided a machine translation decoding method comprising: receiving as input a text segment in a source language to be translated into a target language; generating an initial translation as a current target language translation using translation information learned in part from text corpora; applying one or more modification operators to the current target language translation to generate one or more modified target language translations;
determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations; setting a modified target language translation as the current target language translation; and repeating said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
According to another aspect of the present invention, there is provided a computer-implemented machine translation decoding method comprising iteratively modifying a target language translation of a source language text segment until an occurrence of a termination condition;
wherein iteratively modifying the translation comprises performing at each iteration modification operations on the translation to form modified translations, the modification operations including changing both word translation and word order, and performing at each iteration an evaluation of the modified translations taking into account both source-to-target-language translation information and target-language structure information learned both in part from text corpora.
According to still another aspect of the present invention, there is provided a machine translation decoder comprising: a decoding engine comprising one or more modification operators to be applied to a current target language translation to generate one or more modified target language translations; and a process loop to iteratively modify the current target language translation using the one or more modification operators, the process loop being configured to effect operations comprising: applying the one or more modification operators to the current target language translation using the decoding engine; determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations; setting a modified target language translation as the current target language translation; and repeating 16a said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
The details of one or more embodiments are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.
Drawing Descriptions These and other aspects of the invention will now be described in detail with reference to the accompanying drawings, wherein:
Figure 1A shows a block diagram of machine translation from a user's perspective.
Figure 1B shows an example of word-level alignment.
Figure 2 shows a flowchart of the operation of an embodiment of the greedy decoder.
Figure 3 shows an example of the greedy decoder producing an English translation of a French sentence.
Figure 4 shows an example of output that a user sees as the greedy decoder produces an English translation of a French sentence.
16b ' 75973-18 Figure 5 is a table showing a comparison between different decoders using a trigram language model.
16c Figure 6 shows an example of a rhetorical structure tree.
Figure 7 is an example of a Japanese source sentence.
Figure 0 is the discourse structure of the Japanese source sentence in Fig. 7.
Figure 9 is the discourse structure of an English target sentence translated from the Japanese source sentence of Fig. 11.
Figure 10 shows a block diagram of the tree rewriter.
Figure 11 shows a block diagram of how a tree rewriter can be used as a subsystem of a larger system.
Figure 12 shows a block diagram of a discourse-based machine translation system with the tree rewriter as a subsystem.
Figure 13 is a flowchart of a procedure for building a tree rewriter.
Figure 14 shows an example of incremental tree reconstruction.
Figure 15 is a graph of the learning curve for the Relation-Reduce classifier.
Detailed Description Greedy Decoder A statistical MT system that translates, for example, French sentences into English, can be divided into three parts: (1) a language model (LM) that assigns a probability P(e) to any English string, (2) a translation model (TM) that assigns a probability P(f~e) to any pair of English and French strings, and (3) a decoder. The decoder takes a previously unseen sentence f and tries to find the a that maximizes P(e~f), or equivalently maximizes P(e) ~ P(f~e).
Brown et al., "The mathematics of statistical machine translation: Parameter estimation", Computational .Linguistics, 19(2), 1993, introduced a series of TMs based on word-for-word substitution and re-ordering, but did not include a decoding algorithm. If the source and target languages are constrained to have the same word order (by choice or through suitable pre-processing), then the linear Viterbi algorithm can be applied as described in Tillmann et al., "A DP-based search using monotone alignments in statistical translation", In Proc. ACL, 1997. If re-ordering is limited to rotations around nodes in a binary tree, then optimal decoding can be carried out by a high-polynomial algorithm (Wu, "A polynomial-time algorithm for statistical machine translation", In Pros. ACL, 1996). For arbitrary word-reordering, the decoding problem is NP-complete (nondeterministiC polynomial time complete) (Knight, "Decoding complexity in word-replacement translation models", Computational Linguistics, 25(4), 1999) .
One strategy (Brown et al., "Method and system for natural language translation," U.S. Patent 5,477,451, 1995;
Wang et al., "Decoding algorithm in statistical machine translation", In Proc. AC.L, 1997) is to examine a large subset of likely decodings and chose just from that. Of course, it is possible to miss a good translation this way.
Thus, while decoding is a clear-cut optimization task in which every problem instance has a right answer, it is hard to come up with good answers quickly. The following sets forth details of a fast greedy decoder and compares its performance with that of a traditional stack decoder.
In developing the greedy decoder IBM Model 4 was used, which revolves around the notion of a word alignment over a pair of sentences (see Figure 1B). A word alignment assigns a single home (English string position) to each French word. If two French words align to the same English word, then that English word is said to have a fertility of two. Likewise, if an English word remains unaligned-to, then it has fertility zero. The word alignment in FigurelB
is shorthand for a hypothetical stochastic process by which an English string gets converted into a French string.
There are several sets of decisions to be made.
First, every English word is assigned a fertility.
These assignments are made stochastically according to a table n(o~e~). Any word with fertility zero is deleted from the strong, any work with fertility two is duplicated, etc.
If a word has fertility greater than one, it is called very fertile.
After each English word in the new string, the fertility of an invisible English NULL element with probability p1 (typically about 0.02) is incremented. The NULL element ultimately produces "spurious" French words.
Next, a word-for-word replacement of English words (including NULL) by French words is performed, according to the table t (f~ ~ e;,) .
Finally, the French words are permuted. In permuting, IBM Model 4 distinguishes between French words that are heads (the leftmost French word generated from a particular English word), non-heads (non-leftmost, generated only by very fertile English words), and NU.LZ-generated.
Heads. The head of one English word is assigned a French string position based on the position assigned to the previous English word. If an English word Ee-~
translates into something at French position j, then the French head word of ei is stochastically placed in French position k with distortion probability dl(k-j~class(ei-i).
class (fk)), where "class" refers to automatically determined word classes for French and English vocabulary items. This relative offset k-j encourages adjacent English words to translate into adjacent French words. If ei_1 is infertile, then j is taken from e;,_~, etc. If ei_1 is very fertile, then j is the average of the positions of its French translations.
Non-heads. If the head of English word ei is placed in French position j, then its first non-head is placed in French position k (>j) according to another table d,~(k-j~class (fk)). The next non-head is placed at position q with probability d,1(q-k~class (fq)), and so forth.
NULh-generated. After heads and non-heads are placed, NULL-generated words are permuted into the remaining vacant slots randomly. If there are ~6o NULL-generated words, then any placement scheme is chosen with probability 1/~0!.
These stochastic decisions, starting with e, result in different choices of f and an alignment of f with e. a is mapped onto a particular <a,f> pair with probability:
p (a~ f ~ e) r r ~, ~n(~i ~ ei)x~~t(~ik ~ ~i)x i=1 i=1 k=1 I
dl (~;1- cp ~ class(e p ), class(~;1))x t=i,~; >o r ~;
d>~ (irk - ~i~k-n I class(zik ))x i=1 k=z m Wo plea (1 _ p1 )"' z~ x ~o t(zok ~LL) k=1 where the factors separated by x symbols denote fertility, translation, head permutation, non-head permutation, null-fertilit6y, and null-translation probabilities. The symbols in this formula are: 1 (the length of e), m (the length of f), ei (the ith English word in e), Eo (the NUZZ
word) , s~i (the fertility of ei~, ~o (the fertility of the NUZZ word) , iik (the kth French word produced by ei in a) , mix (the position of iik in f) , pI (the position of the first fertile word to the left of ei in a) , c,,I (the ceiling of the average of all ~tp;,k for pi, or 0 if pi is undefined) .
In view of the foregoing, given a new sentence f, then an optimal decoder will search for an a that maximizes Pelf) ... p(e) ~ P(f~e) . Here, P(f~e) is the sum of P(a,f~e) over all possible alignments a. Because this sum involves significant computation, typically it is avoided by instead searching for an <e,a> pair that maximizes P~(e,a~f) ~ P(e) P(a,f~e). It is assumed that the language model P(e) is a smoothed n-gram model of English.
Figure 2 is a flowchart of the operation of an embodiment of a greedy decoder for performing MT. As shown therein, the first step 200 is to receive an input sentence to be translated. Although in this example, the text segment being translated is a sentence, virtually any other text segment could be used, for example, clauses, paragraphs or entire treatises.
In step 202, as a first approximation of the translation, the greedy decoder generates a "gloss" of the input sentence, essentially a word-for-word translation.
The gloss is constructed by aligning each French word f~
with its most likely English translation ef~ (ef~ = argmaxe t(e ~ f~)). For example, in translating the French sentence "Bien entendu, i1 parle de une belle victoire", the greedy decoder initially assumes that a good translation is ' Well heard, it talking a beautiful victory " because the best translation of "bien" is "well", the best translation of "entendu" is "heard", and so on. The alignment corresponding to this translation is shown at the top of Figure 3.
In step 204, the decoder estimates the probability of correctness of the current translation, P(c).
After the initial alignment is generated in step 202, the greedy decoder tries to improve it in step 206. That is, the decoder tries to find an alignment (and implicitly, a translation) of higher probability by applying one or more sentence modification operators, described below. The use of a word-level alignment and the particular operators described below were chosen for this particular embodiment.
However, alternative embodiments using different statistical models may benefit from different or additional operations.
The following operators collectively make-up the decoder's translation engine, and include the following:
translateOneOrTwoWoreis ( jz, e1, j2, e~) This operation changes the translation of one or two French words, those located at positions j1 and j~, from ef~l and ef~2 into e1 and e~. If ef~ is a word of fertility 1 and ek is NULL, then ef~ is deleted from the translation. If ef~
is the NULL word, the word ek is inserted into the translation at the position that yields an alignment of highest probability. If ef~~ = ei or ef~~ = e2, then this operation amounts to changing the translation of a single word.
translateAndlnsert ( j, e1, e~) This operation changes the translation of the French word located at position j from ef~ into e1 and simultaneously inserts word e2 at the position that yields the alignment of highest probability. ~nlord e2 is selected from an automatically derived list of 1024 words with high probability of having fertility 0. When ef~ = e1, this operation amounts to inserting a word of fertility 0 into the alignment.
removeWordOfFertility0 (i) This operation deletes the word of fertility 0 at position i in the current alignment.
swapSegments ( iz, i~, This operation creates a new alignment from the old one by swapping non-overlapping English word segments [i1, i~] and [j1, j2]. During the swap operation, all existing links between English and French words are preserved. The segments can be as small as a word or as long as ~e~-1 words, where ~e~ is the length of the English sentence.
joinWords ( i1, ia) This operation eliminates from the alignment the English word at position i1 (or i~) and links the French words generated by eil ( or ei~ ) to ei~ ( or ei1 ) .
At step 208, the decoder estimates the probabilities of correctness, P (M~) ... P (M") , for each of the results of the sentence modification operations. That is, the probability for each new resulting translation is determined At step 210, the decoder determines whether any of the new translations are better than the current translation by comparing their respective probabilities of correctness.
If any of the new translations represents a better solution than the current translation, then the best new translation (that is, the translation solution having the highest probability of correctness) is set as the current translation in step 214 and the decoding process returns to step 206 to perform one or more of the sentence modification operations on the new current translation solution.
Steps 206, 208, 210 and 214 repeat until the sentence modification operations cease (as determined in step 210) to produce translation solutions having higher probabilities of correctness, at which point, the decoding process halts at step 212 and the current translation is output as the final decoding solution. Alternatively, the decoder could cease after a predetermined number of iterations chosen, for example, either by a human end-user or by an application program using the decoder as a translation engine.
Accordingly, in a stepwise fashion, starting from the initial gloss, the greedy decoder uses a process loop (e.g., as shown in Fig. 2, steps 206, 208, 210 and 214) to iterate exhaustively over all alignments that are one operation away from the alignment under consideration. At every step, the decoder chooses the alignment of highest probability, until the probability of the current alignment can no longer be improved. When it starts from the gloss of the French sentence "Bien entendu, i1 parle de une belle victoire.", for example, the greedy decoder alters the initial alignment incrementally as shown in Figure 3, eventually producing the translation "Quite naturally, he talks about a great victory.". In this process, the decoder explores a total of 77421 distinct alignments /
translations, of which "Quite naturally, he talks about a great victory." has the highest probability.
In step 206 of the decoding process, either all of the five sentence modification operations can be used or any subset thereof may be used to the exclusion of the others, depending on the preferences of the system designer and/or end-user. For example, the most time consuming operations in the decoder are swapSegments, translateOneOrTwoWords, and translateAndlnsert. SwapSegments iterates over all possible non-overlapping span pairs that can be built on a sequence of length I a I. TranslateOneOrTwoWords iterates over I f 12 x I t I~ alignments, where I f I is the size of the French sentence and I t I is the number of translations associated with each word (in this implementation, this number is limited to the top 10 translations).
TranslateAndlnsert iterates over I f I x I t I x I z I
alignments, where I z I is the size of the list of words with high probability of having fertility 0 (1024 words in this implementation). Accordingly, the decoder could be designed to omit one or more of these slower operations in order to speed up decoding, but potentially at the cost of accuracy. Alternatively, or in addition, the decoder could be designed to use different or additional sentence modification operations according to the objectives of the system designer and/or end-user.
An advantage of the greedy decoder comes from its speed. As the experiments described below demonstrate, the greedy decoder can produce translations faster than any other decoder. The greedy decoder is an instance of an "anytime algorithm" - the longer it runs, the better the translation it finds. One potential tradeoff of the greedy decoder relates to the size of the solution space that it explores, which is relatively small. The farther away a good translation is from the initial gloss, the less likely the greedy decoder is to find it.
Fig. 4 shows another example of the greedy decoder in action in which an acceptable solution is reached in four iterations. As shown therein, the input sentence to be translated is "ce ne est pas juste." The decoder uses the initial gloss "that not is not fair." and determines that this translation solution (Iteration 1) has a probability of correctness ("Aprob" - the product of LMprob and TMprob) of 1.13162e-22, based on a language model probability (LMprob) of 2.98457e-14 and a translation model probability (TMprob) of 3.79156e-09.
In the second iteration, the decoder changes the first instance of the word "not" in the translation to "is" by applying the translateOneOrTwoWords operation, resulting in a new translation solution "that is is not fair", having the probabilities shown in Fig. 4, Iteration 2. In the third iteration, the decoder applies the removeWordOfFertility0 operation and drops one instance of the word "is" in the translation, resulting in a new translation solution "that is not fair", having the probabilities shown in Fig. 4, Iteration 3. In the fourth and final iteration, the decoder applies the translateOneOrTcvoWords operation again to change the word "that" in the translation to "it", resulting in the final translation solution "it is not fair", having the probabilities shown in Fig. 4, Iteration 4.
To determine the performance of the greedy decoder, a series of experiments was performed. In all experiments, decoding was performed using only the top 10 translations of a word, as determined during training, and a list of 1024 words of fertility 0, which was also extracted automatically from the test corpus.
In experiments to determine the accuracy and compare the speed of the greedy decoder described to a conventional stack decoder (such as described in U.S. Patent No.
5,477,451 to Brown et al., a test collection of 505 sentences was used, uniformly distributed across the lengths 6, S, 10, 15, and 20. Decoders were evaluated with respect to (1) speed, and (2) translation accuracy.
The results in the table shown in Fig. 5, obtained with decoders that use a trigram language model, show that the greedy decoding algorithm is an advantageous alternative to the traditional stack decoding algorithm.
Even when the greedy decoder used an optimized-for-speed set of operations (that is, a subset of the total set of five sentence modification operations discussed above) in which at most one word is translated, moved, or inserted at a time - which is labeled "greedy*" in Fig. 5 - the translation accuracy is affected only slightly. In contrast, the translation speed increases with at least one order of magnitude. Depending on the application of interest, one may choose to use a slow decoder that provides optimal results or a fast, greedy decoder that provides non-optimal, but acceptable results.
Alternative embodiments for the greedy decoder are possible. For example, the greedy decoder could start with multiple different initial translations (e. g., different variations on the gloss used in step 202 in Fig. 2) and then run the greedy decoding algorithm (i.e., steps 204-214 in Fig. 2) on each of the different initial translations in parallel. For example, the greedy decoder code start with an initial, approximate translation selected from among multiple translated phrases stored in a memory. In the end, the best translation could be selected. This parallel translation of different initial solutions could result in more accurate translations.
Tree Rewriter Almost all conventional MT systems process text one sentence at a time. because of this limited focus, MT
systems generally cannot re-group and re-order the clauses and sentences of an input text to achieve the most natural rendering in a target language. Yet, even between languages as close as English and French, there is a 100 mismatch in number of sentences - what is said in two sentences in one language is said in only one, or in three, in another (Gale et al., "A program for aligning sentences in bilingual corpora," Computational .Linguistics, 19(1):75-102, 1993). For distant language pairs, such as Japanese and English, the differences are more significant.
Consider, for example, the Japanese sentence shown in Fig . 7 ( "text ( 1 ) ") . The following ( "text ( 2 ) ") is a word-by-word "gloss" of text (1):
(2) [The Ministry of Health and Welfare last year revealed]
[population of future estimate according toy] [in future 1.499 persons as the lowest3] [that after *SAB* rising to turn that4] [*they*
estimated buts] [already the estimate misses a point ]
[prediction became.']
In contrast, a two-sentence translation of the Japanese sentence produced by a professional translator ("text(3)") reads as follows:
(3) [In its future population estimatesl] [made public last year,] [the Ministry of Health and Welfare predicted that the SAB
would drop to a new low of 1.499 in the future, 3] [but would make a comeback after that,4] [increasing once again.5] [However, it looks as if that prediction will be quickly shattered.6]
The labeled spans of text represent elementary discourse units (edus), i.e., minimal text spans that have an unambiguous discourse function (Mann et al., "Rhetorical structure theory: Toward a functional theory of text organization," Text, 8(3):243-281, 1988). If the text fragments are analyzed closely, one notices that in translating text (1), a professional translator chose to realize the information in Japanese unit 2 first (unit 2 in text (1) corresponds roughly to unit 1 in text (3)); to realize then some of the information in Japanese unit 1 (part of unit 1 in text (1) corresponds to unit 2 in text (3)); to refuse then information given in units 1, 3 and 5 in text (1) and realize it in English as unit 3; and so on.
Also, the translator chose to re-package the information in the original Japanese sentence into two English sentences.
At the elementary unit level, the correspondence between the Japanese sentence in text (1) and its English translation in text (3) can be represented as in mappings (4), below, where j c a denotes the fact that the semantic content of unit j is realized fully in unit e; j ~ a denotes the fact that the semantic content of unit a is realized fully in unit j; j = a denotes the fact that units j and a are semantically equivalent; and j - a denotes the fact that there is a semantic overlap between units j and e, but neither proper inclusion not proper equivalence.
(4) js ~ ez: j1 j2 = ei;
js ~ es;
j4 = en: j4 = es:
j5 = es:
js ~ es:
j~
Hence, the mappings in (4) provide an explicit representation of the way information is re-ordered and re-packaged when translated from Japanese into English.
However, when translating text, it is also the case that the rhetorical rendering changes. What is realized in Japanese using a CONTRAST relation can be realized in English using, for example, a COMPARISON OR A CONCESSION
relation.
Figures 8 and 9 present in the style of Mann, supra, the discourse structures of text fragments (1) and (3), above. Each discourse structure is a tree whose leaves correspond to contiguous text spans. Each node is characterized by a status (NUCLEUS or SATELLITE) and a rhetorical relation, which is a relation that holds between two non-overlapping text spans. The distinction between nuclei and satellites comes from the empirical observation that the nucleus expresses what is more essential to the writer's intention than the satellite; and that the nucleus of a rhetorical relation is comprehensible independent of the satellite, but not vice versa. When spans are equally important, the relation is multinuclear; for example, the CONTRAST relation that holds between unit [3] and span [4,5] in the rhetorical structure of the English text in Figures 8 and 9 is multinuclear. Rhetorical relations that end in the suffix "-e" denote relations that correspond to embedded syntactic constituents. For example, the ELABORATION-OBJECT-ATTRIBUTE-E relation that holds between units 2 and 1 in the English discourse structure corresponds to a restrictive relative.
If one knows the mappings at the edu level, one can determine the mappings at the span (discourse constituent) level as well. For example, using the elementary mappings in (4), one can determine that Japanese span [1,2]
corresponds to the English span [1,2], Japanese unit [4] to English span [4,5], Japanese span [6,7] to English unit [6], Japanese span [1,5] to English span [1,5], and so on.
As Figures 8 and 9 show, the CONCESSION relation that holds between spans [1,5] and [6,7] in the Japanese tree corresponds to a similar relation that holds between span [1,5], and unit [6] in the English tree (modulo the fact that, in Japanese, the relation holds between sentence fragments, while in English it holds between full sentences). However, the TEMPORAL-AFTER relation that holds between units [3] and [4] in the Japanese tree is realized as a CONTRAST relation between unit [3] and span [4,5] in the English tree. And because Japanese units {6]
and [7] are fused into unit [6] in English, the relation ELABORATION-OBJECT-ATTRIBUTE-E is no longer made explicit in the English text.
Some of the differences between the two discourse trees in Figures S and 9 have been traditionally addressed in MT systems at the syntactic level. For example, the re-ordering of units 1 and 2 can be dealt with using only syntactic models. However, as discussed below, there are significant differences between Japanese and English with respect to the way information is packaged and organized rhetorically not only at the sentence level, but also, at the paragraph and text levels. More specifically, as humans translate Japanese into English, they re-order the clauses, sentences, and paragraphs of Japanese texts, they re package the information into clauses, sentences and paragraphs that are not a one-to-one mapping of the original Japanese units, and they rhetorically re-organize the structure of the translated text so as to reflect rhetorical constraints specific to English. If a translation system is to produce text that is not only grammatical but also coherent, it will have to ensure that the discourse structure of the target text reflects the natural renderings of the target language, and not that of the source language.
In the Experiment section below, it is empirically shown that there are significant differences between the rhetorical structure of Japanese texts and their corresponding English translations. These differences further demonstrate the need and desirability of developing computational models for discourse structure rewriting.
Experiment In order to assess the role of discourse structure in MT, a corpus of discourse trees was manually built for 40 Japanese texts and their corresponding translations. The texts were selected randomly from the ARPA corpus (White et al., "Evaluation in the ARPA machine-translation program:
1993 methodology," In Proceedings of the ARPA Human Language Technology Workshop, pages 135-140, Washington, D.C., 1994). On average, each text had about 460 words.
The Japanese texts had a total of 335 paragraphs and 773 sentences. The English texts had a total of 337 paragraphs and 827 sentences.
A discourse annotation protocol was developed for Japanese and English along the lines followed by Marcu et al., "Experiments in constructing a corpus of discourse trees, " In Proc. Of the ~ICL'99 Workshop on Standards and Tools for Discourse Tagging, pages 48-57, Maryland (1999).
Marcus discourse annotation tool (1999) was used in order to manually construct the discourse structure of all Japanese and English texts in the corpus. Ten percent of the Japanese and English texts were rhetorically labeled by two annotators. The tool and the annotation protocol are available at http://www.isi.edu/~marcu/software/
The annotation procedure yielded over the entire corpus 2641 Japanese edus and 2363 English edus.
Corpus k" (3) ks (#) -I- kn (#)-- I kr (#) Japanese 0.856 (80) 0.785 (3377) 0.724 (3377) 0.650 (3377) English 0.925 (60) 0.866 (1826) 0.839 (1826) 0.748 (1826) Table 1: Tagging reliability The reliability of the annotation was computer using Marcu et al. (1999)'s method for computing kappa statistics (Siegel et al., Non parametric Statistics for the Behavioral Sciences, McGraw-Hill, Second edition, 1988) over-hierarchical structures. Table 1 above displays average kappa statistics that reflect the reliability of the annotation of elementary discourse units, k", hierarchical discourse spans, ks, hierarchical nuclearity assignments, k", and hierarchical rhetorical relation assignments, kr. Kappa figures higher than 0.8 correspond to good agreement; kappa figures higher than 0.6 correspond to acceptable agreement. All kappa statistics were statistically significant at levels higher than a = 0.01.
In addition to the kappa statistics, table 1 also displays in parentheses the average number of data points per document, over which the kappa statistics were computed.
For each pair of Japanese-English discourse structures, an alignment file was also built manually, which. specified in the notation discussed on page 1 the correspondence between the edus of the Japanese text and the edus of the English translation.
The similarity between English and Japanese discourse trees was computed using labeled recall and precision figures that reflected the resemblance of the Japanese and English discourse structures with respect to their assignment of edu boundaries, hierarchical spans, nuclearity, and rhetorical relations.
Because the trees compared differ from one language to the other in the number of elementary units, the order of these units, and the way the units are grouped recursively into discourse spans, two types of recall and precision figures were computed. In computing Position-Dependent (P-D) recall and precision figures, a Japanese span was considered to match an English span when the Japanese span contained all the Japanese edus that corresponded to the edus in the English span, and when the Japanese and English spans appeared in the same position with respect to the overall structure. For example, the English tree in Figures 8 and 9 are characterized by 10 subsentential spans: [1], [2], [3], [4], [5], [6], [1,2], [4,5], [3,5], and [1,5]. (Span [1,6] subsumes 2 sentences, so it is not sub-sentential.) The Japanese discourse tree has only 4 spans that could be matched in the same positions with English spans, namely spans [1,2], [4], [5], and [1,5].
Hence the similarity between the Japanese tree and the English tree with respect to their discourse below the sentence level has a recall of 4/10 and a precision of 4/11 (in Figures 8 and 9, there are 11 sub-sentential Japanese spans ) .
In computing Position-Independent (P-I) recall and precision figures, even when a Japanese span "floated"
during the translation to a position in the English tree, the P-I recall and precision figures were not affected.
The Position-Independent figures reflect the intuition that if two trees t1 and t2 both have a subtree t, t1 and t~ are more similar than if they were if they didn't share a tree.
At the sentence level, it is assumed that if, for example, the syntactic structure of a relative clause is translated appropriately (even though it is not appropriately attached), this is better than translating wrongly that clause. The Position-Independent figures offer a more optimistic metric for comparing discourse trees. They span a wider range of values than the Position-Dependent figures, which enables a better characterization of the differences between Japanese and English discourse structures. When one takes an optimistic stance, for the spans at the sub-sentential level in the trees in Table 1 the recall is 6/10 and the precision is 6/11 because in addition to spans [1,2], [4], [5], and [1,5], one can also match Japanese span [1] to English span [2] and Japanese span [ 2 ] to Japanese span [ 1 ] .
In order to provide a better estimate of how close two discourse trees were, Position-Dependent and -Independent reoall and precision figures were computed for the sentential level (where units are given by edus and spans are given by sets of edus or single sentences); paragraph level (where units are given by sentences and spans are given by sets of sentences or single paragraphs); and text level (where units are given by paragraphs and spans are given by sets of paragraphs). These figures offer a detailed picture of how discourse structures and relations are mapped from one language to the other across all discourse levels, from sentence to text. The differences at the sentence level can be explained by differences between the syntactic structures of Japanese and English.
The differences at the paragraph and text levels have a purely rhetorical explanation.
When the recall and precision figures were computed with respect to the nuclearity and relation assignments, the statuses and the rhetorical relations that labeled each pair of spans were also factored in.
Units Spans Status/NUClearityRelations Level P-D P-D P-D P-D P P-D P-D P P-D P-D
R P R R R P
Sentence 29.1 25.0 27.222.7 27..317.7 14.9 12.4 Paragraph 53.9 53.4 46.847.3 38.6 39.0 31.9 32.3 Text 41.3 42.6 31.532.6 28.8 29.9 26.1 27.1 Weighted Average 36.0 32.5 31.828.4 26.0 23.1 20.1 17.9 All 8.2 7.4 5.9 5.3 4.4 3.9 3.3 3.0 P-I P-TP P-I P-IP P-I .__..P-IP P-T P-IP
R R R .._ R
Sentence 71.0 61.0 56.0 46.6 44.3 36.9 30.5 25.4 Paragraph62.1 61.6 53.2 53.8 43.3 43.8 35.1 35.5 Text 74.1 76.5 54.4 56.5 48.5 50.4 41.1 42.7 Weighted 69.6 63.0 55.2 49.2 44.8 39.9 33.1 29.5 Average All 74.5 66.8 50.6 45.8 39.4 35.7 26.8 24.3 Table 2: Similarity of Japanese and English discourse structures Table 2 above summarizes the results (P-D and P-I
(R)ecall and (P)recision figures) for each level (Sentence, Paragraph, and Text). The numbers in the "Weighted Average" line report averages of the Sentence-, Paragraph-, and Text-specific figures, weighted according to the number of units at each level. The numbers in the "All" line reflect recall and precision figures computed across the entire trees, with no attention paid to sentence and paragraph boundaries.
Given the significantly different syntactic structures of Japanese and English the recall and precision results were low, reflecting the similarity between discourse trees built below the sentence level. However, as Table 2 shows, these are significant differences between discourse trees at the paragraph and text levels as well. For example, the Position-Independent figures show that only about 620 of the sentences and only about 530 of the hierarchical spans built across sentences could be matched between the two corpora. when one looks at the status and rhetorical relations associated with the spans built across sentences at the paragraph level, the P-I recall and precision figures drop to about 43% and 35o respectively.
The differences in recall and precision are explained both by differences in the way information is packaged into paragraphs in the two languages and the way it is structured rhetorically both within and about the paragraph level.
These results strongly suggest that if one attempts to translate Japanese into English on a sentence-by-sentence basis, it is likely that the resulting text will be unnatural from a discourse perspective. For example, if some information rendered using a CONTRAST relation in Japanese is rendered using an ELABORATION relation in English, it would be inappropriate to use a discourse marker like "but" in the English translation, although that would be consistent with the Japanese discourse structure.
An inspection of the rhetorical mappings between Japanese and English revealed that some Japanese rhetorical renderings are consistently mapped into one or a few preferred renderings in English. For example, 34 of 115 CONTRAST relations in the Japanese texts are mapped into CONTRAST relations in English; 27 become nuclei of relations such as ANTITHESIS and CONCESSION, 14 are translated as COMPARISON relations, 6 as satellites of CONCESSION relations, 5 as LIST relations, etc.
The discourse-based transfer model Figure 10 is a block diagram of a tree rewriter in the process of being trained. As shown therein, the tree rewriter 700 takes as input two different types of trees, for example, a tree of type A and another tree of type B, and automatically learns how to rewrite type A trees into type B trees. The tree rewriter 700 produces as output a transfer function, H(A ~ B), for rewriting trees of type A
into trees of type B. Accordingly, assuming type A
corresponds to trees in Japanese and type B corresponds to trees in English, H(A -~ B) enables a user (e.g., either a human end-user or a software application invoking the tree rewriter) to automatically convert any tree structure in English into a counterpart tree structure in Japanese.
The tree rewriter works on syntactic trees, rhetorical trees, and virtually any other type of tree structure used in computational linguistics. The tree rewriter has applications not only in machine translation, but also in summarization, discourse parsing, syntactic parsing, information retrieval, automatic test scoring and other applications that generate or use trees. In machine translation, for example, the tree rewriter can be used to rewrite a syntactic / rhetorical tree specific to one language into a syntactic / rhetorical tree for another language. In summarization, the tree rewriter can be used to rewrite the discourse / syntactic structure of a long text or sentence into the discourse / syntactic structure of a short text or sentence.
This high degree of general applicability is shown in Fig. 11, in which the tree rewriter 801, after it has been trained to learn the transfer function H(Tree ~ Tree'), can accept a tree as input from any application 800 that generates trees as output. At the output side, the tree rewriter's output (tree' - a rewritten version of the input tree) can be used as input to any application that uses trees as input.
Figure 12 shows a block diagram of a specific application for the tree rewriter as a component of a larger system - namely, a discourse-based machine translation system. Unlike conventional MT systems which in effect adopt a "tiled" approach to translation, for example, by translating individual sentences of a larger work (e.g., a treatise) separately, the discourse-based MT
system of Figure 12 translates the entire text as whole, potentially resulting in a translation that has a different number and/or arrangement of sentences than the original, but which better captures the underlying discourse, or rhetoric, of the original text.
As shown in Figure 12, the discourse-based MT system 910 receives as input a source language text 900 and produces as output a target language text 908, which is a discourse-based translation of the source language text 900. The system 910 includes three basic components - a discourse parser 902, a discourse-structure transfer module 904 (i.e., a specific instance of a tree rewriter that has been trained to rewrite trees using the transfer function H(Tree -~ Tree')) and a target language tree-text mapper 906.
The discourse parser 902 initially derives the discourse structure of the source language text and produces as output a corresponding discourse tree. Details of a discourse parser that can be used as discourse parser 902 are set forth in Daniel Marcu, "A Decision-Based Approach to Rhetorical Parsing," Proceedings of ACL '99 (1999) .
The target language tree-text mapper 906 is a statistical module that maps the input text into the target language using translation and language models that incorporate discourse-specific features, which are extracted from the outputs of the discourse parser 209 and the discourse-structure transfer module 904. Details of a suitable mapper 906 are set forth in Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, Kenji Yamada, "Fast Decoding and Optimal Decoding for Machine Translation," Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics, July 6-11, 2001.
As noted above, the discourse-structure transfer module 904 is a specific instance of a tree rewriter that has been trained to rewrite trees from the desired input type into the desired output type. More specifically, the discourse-structure transfer module 904 rewrites the discourse structure of the input text so as to reflect a discourse rendering that is natural to the target text.
Fig. 13 is a flow chart illustrating a process 1300 that can be used to train a tree rewriter to automatically learn the transfer function between two different types of tree structures, e.g., a tree of type A
and a tree of type B.
As shown in Fig. 13, the first step 1301 is to generate a training set of input-output pairs of trees [TS, Tt] and a mapping C between leaves each input-output tree pair. The input tree of the pair is of the type from which conversion is desired or, in other words, the source tree type TS. The output tree of the pair is of the type to which conversion is desired or, in other words, the target tree type Tt .
The mapping C between leaves of an input tree and its paired output tree defines a correspondence between a source text segment and its counterpart target language translation. These mappings either can be generated manually as described below, or automatically, as described in Kevin Knight and Daniel Marcu, "Statistics-Based Summarization - Step One: Sentence Compression," The l7tn National Conference on Artificial Intelligence (AAAI-2000), pp. 703-710.
Available types of mappings between leaves of a Japanese-English input-output pair that can be used are shown in equation (4) above, wherein j refers to a Japanese text segment and a refers to an English translation of that text segment. Note that the mappings represented by equation (4) are not limited to one-to-one mappings but rather can be any arbitrary mapping - that is, not only one-to-one, but also one-to-many, many-to-one and many-to-many. This flexibility in the mappings dramatically enhances the richness with which relationships between the input and output trees are defined, and further enhanoes the flexibility of the transfer function H[] that is automatically learned.
After the training set (input-output tree pairs and mappings therebetween) has been generated, the training process next determines in step 1303 the grouping and order of operations that generate a given output tree starting with its paired input tree. This step is performed based on the following seven basic operations, referred to collectively as the "extended shift-reduce" operations -shift, reduce, break, create-next, fuse, swap, and assignType - which are described in detail below under the section heading "The discourse-based transfer model."
These seven operations are sufficient to rewrite any given input tree into its paired output tree.
The output of step 1303 is a set of learning cases -one learning case for each input-output tree pair in the training set. In essence, each learning case is an ordered set of extended shift-reduce operations that when applied to an input tree will generate the paired output tree.
Next, in step 1305, the tree rewriter training process 1300 associates features (e. g., operational and discourse features, correspondence-based features, and lexical features) with the learning cases to reflect the context in which the operations are to be performed. Details of step 1305 are discussed below under the heading "Learning the parameters of the discourse-transfer model."
Next, in step 1307, the tree rewriter training process 1300 applies a learning algorithm, for example, the C4.5 algorithm as described in J. Ross Quinlan, "C4.5: Programs for Machine Learning," Morgan Kaufmann Publishers (1993), to learn a set of decision rules from the learning cases.
Details of step 1307 are discussed below under the heading "Learning the parameters of the discourse-transfer model."
This set of decision rules collectively constitute the transfer function, H (TS -~ Tt) , for rewriting any tree of type TS into a tree of type Tt. This transfer function can then be used by users, applications or other automated processes for rewriting previously unseen trees of type TS
into trees of type Tt.
A more detailed discussion of training a tree rewriter follows.
In order to learn to rewrite discourse structure trees, a related problem, defined below in Definition 3.1, is addressed.
Definition 3.1 Given two trees TS
and Tt and a correspondence Table C
defined between TS and Tt a t the leaf level in terms of =, c, and - reactions, find a sequence of actions that rewrites the tree TS
into T~.
If for any tuple (TS, Tt, C) such a sequence of actions can be derived, it is then possible to use a corpus of (TS, Tt, C) tuples in order to automatically learn to derive from an unseen tree Tsi, which has the same structural properties as the trees TS, a tree Tt~, which has structural properties similar to those of the trees Tt.
Solving the problem in definition 3.1 involves, in part, extending the shift-reduce parsing paradigm applied by Mangerman, "Statistical decision-tree models for parsing," In Proo. Of ACL'95, pages 276-283, Cambridge, Massachusetts (1995), Hermjak~b et al., "Learning parse and translation decisions from examples with rich context," In Proc. Of ACL'97, pages 482-489, Madrid, Spain (1997), and Marcu, "A decision-based approach to rhetorical parsing,"
In Proc. Of ACL'99, pages 365-372, Maryland (1999). In this extended paradigm, the transfer process starts with an empty Stack and an Input List that contains a sequence of elementary discourse trees edts, one edt for each edu in the tree TS given as input. The status and rhetorical relation associated with each edt is undefined. At each step, the transfer module applies an operation that is aimed at building from the units in TS the discourse tree Tt. In the context of the discourse-transfer module, 7 types of operations are implemented:
~ SHIFT operations transfer the first edt from the input list into the stack;
~ REDUCE operations pop the two discourse trees located at the top of the stack; combine them into a new tee updating the statuses and rhetorical relation names of the trees involved in the operation; and push the new tree on the top of the stack. These operations are used to build the structure of the discourse tree in the target language, ~ BREAK operations are used in order to break the edt at the beginning of the input list into a predetermined number of units. These operations are used to ensure that the resulting tree has the same number of edts as Tt. For example, a BREAK operation is used whenever a Japanese edu is mapped into multiple English units.
~ CREATE-NEXT operations are used, for example, in order to create English (target language) discourse constituents that have no correspondent in the Japanese (source language) tree.
~ FUSE operations are used in order to fuse the edt at the top of the stack into the tree that immediately precedes it. These operations are used, for example, whenever multiple Japanese edus are mapped into one English edu.
~ SWAP operations swap the edt at the beginning of the input list with an edt found one or more positions to the right. These operations are used for re-ordering discourse constituents.
~ ASSIGNTYPE operations assign one or more of the following types to the tree t the top of the stack:
Unit, Multiunit, Sentence, Paragraph, MultiParagraph, and Text. These operations are used in order to ensure sentence and paragraph boundaries that are specific to the target language.
For example, the first sentence of the English tree in Figure 9 can be obtained from the original Japanese sequence by following the sequence of actions (5), below, whose effects are shown in Figure 14. For the purpose of compactness, Figure 14 does not illustrate the effect of ASSIGNTYPE actions. For the same purpose, some lines correspond to more than one action.
(5) BREAK 2; SWAP 2; SHIFT; ASSIGNTYPE
UNIT; SHIFT; REDUCE-NS-ELABORATION-OBJECT-ATTRIBUTE-E; ASSIGNTYPE
MULTIUNIT; SHIFT; ASSIGNTYPE UNIT;
SHIFT; ASSIGNTYPE UNIT; FUSE;
ASSIGNTYPE UNIT; SWAP2; SHIFT;
ASSIGNTYPE UNIT; FUSE; BREAK2; SHIFT;
ASSIGNTYPE UNIT; SHIFT; ASSIGNTYPE
UNIT; REDUCE-NS-ELABORATION-ADDITIONAL;
ASSIGNTYPE MULTIUNIT; REDUCE-NS-CONTRAST; ASSIGNTYPE MULTIUNIT; REDUCE-SN-BACKGROUND; ASSIGNTYPE SENTENCE.
For the corpus used, in order to enable a discourse-based transfer module to derive any English discourse tree starting from any Japanese discourse tree, it is sufficient to implement:
~ one SHIFT operation;
~ 3 x 2 x 85 REDUCE operations; (For each of the three possible pairs of nuclearity assignments NUCLEUS-SATELLITE (NS), SATELLITE-NUCLEUS (SN), AND NUCLEUS-NUCLEUS (NN), there are two possible ways to reduce two adjacent trees (one results in a binary tree, the other in a non-binary tree (Marcu, "A decision-based approach to rhetorical parsing," In Proc. Of ACL'99, pages 365-372, Maryland (1999)), and 85 relation names.);
~ three types of BREAK operations; (In the corpus used, a Japanese unit is broken into two, three, or at most four units.);
~ one type of CREATE-NEXT operation;
~ one type of FUSE operation;
~ eleven types of SWAP operations; (In the corpus, Japanese units are at most 11 positions away from their location in an English-specific rendering.) ~ seven types of ASSIGNTYPE operations: Unit, Multiunit, Sentence, MultiSentence, paragraph, MultiParagraph, and Text.
These actions are sufficient for rewriting any tree TS
into any tree Tt, where Tt may have a different number of edus, where the edus of Tt may have a different ordering than the edus of TS, and where the hierarchical structures of the two trees may be different as well.
Learning the parameters of the discourse-transfer model Each configuration of the transfer model is associated with a learning case. The cases were generated by a program that automatically derived the sequence of actions that mapped the Japanese trees in the corpus into the sibling English trees, using the correspondences at the elementary unit level that were constructed manually.
Overall, the 40 pairs of Japanese and English discourse trees yielded 14108 cases, To each learning example, a set of features from the following classes was associated:
Operational and discourse features reflect the number of trees in the stack, the input list, and the types of the last five operations. They encode information pertaining to the types of the partial trees built up to a certain time and the rhetorical relations that hold between these trees.
Correspondence-based features reflect the nuclearity, rhetorical relations, and types of the Japanese trees that correspond to the English-like partial trees derived up to a given time.
Lexical features specify whether the Japanese spans that correspond to the structures derived up to a given time use potential discourse markers, such as dakara (because) and no ni (although) .
The discourse transfer module uses the C4.5 program (Quinlan, C4.5: Programs for Machine Learning, Morgan Ifiaufmann Publishers (1993)) in order to learn decision trees and rules that specify how Japanese discourse trees should be mapped into English-like trees. A ten-fold cross-validation evaluation of the classifier yielded an accuracy of 70.20 (~ 0.21).
In order to better understand the strengths and weaknesses of the classifier, the problem was broken into smaller components. Hence, instead of learning all actions at once, it was determined first whether the rewriting procedure should choose a SHIFT, REDUCE, BREAK, FUSE, SWAP, or ASSIGNTYPE operation (the Domain Action Type" classifier in table 3), and only then to refine this decision by determining what type of reduce operation to perform how many units to break a Japanese units into, how big the distance to the SWAP-ed unit should be, and what type of ASSIGNTYPE operation one should perform. Table 3 below shows the sizes of each data set and the performance of each of these classifiers, as determined using a ten-fold cross-validation procedure. For the purpose of comparison, each classifier is paired with a majority baseline.
Classifier # Cases Accuracy Majority (10-fold baseline cross accuracy validation) General 14108 70.20% (0.21) 22.05%(on ASSIGNTYPE UNIT) (Learns all classes at once) Main Action Type 14108 82.530 (0.25) 45.47%(on ASSIGNTYPE) AssignType 6416 90.46% (0.39) 57.30%(on ASSIGNTYPE Unit) Break 394 82.91% (1.40) 82.91%(on BREAK 2) Nuclearity-Reduce 2388 67.43% (1.03) 50.920(on NS) Relation-Reduce 2388 48.20% (1.01) 17.18%(on ELABORATION-OBJECT-ATTRIBUTE-E) Swap 842 62.98% (1.62) 62.98%(on SWAP1) Table 3: Performance of the classifiers The results in Table 3 show that the most difficult subtasks to learn are that of determining the number of units a Japanese unit should be broken into and that of determining the distance to the unit that is to be swapped.
The features used are not able to refine the baseline classifiers for these action types. The confusion matrix for the "Main Action Type" classifier (see Table 4) shows that the system has trouble mostly identifying BREAK and CREATE-NEXT actions. The system has difficulty learning what type of nuclearity ordering to prefer (the "Nuclearity-Reduce" classifier) and what relation to choose for the English-like structure (the "Relation-Reduce"
classifier).
Action (a) (b) (c) (d) (e) (f) (g) .
ASSIGNTYPE (a) 660 BREAK (b) 1 2 28 l (c) FUSE (d) 69 8 3 REDUCE (e) 4 18 193 30 3 SHIFT (f) 1 4 15 44 243 25 SWAP (g) 3 4 14 43 25 l0 Table 4: Confusion matrix for the Main Action Type classifier Figure 15 shows a typical learning curve, the one that corresponds to the "Reduce Relation" classifier. The learning curves suggest that more training data may improve performance. However, they also suggest that better features may be needed in order to improve performance significantly.
Table 5 below displays some learned rules. The first rule accounts for rhetorical mappings in which the order of the nucleus and satellite of an ATTRIBUTION relation is changed when translated from Japanese into English. The second rule was learned in order to map EXAMPLE Japanese satellites into EVIDENCE English satellites.
if rhetRelOfStack-lInJapTree = ATTRIBUTION
then rhetlOfTopStackInEngTree <-- ATTRIBUTION
if rhetRelOfTopStackInJapTree = EXAMPLE /~
isSentenceTheLastUnitinJapTreeOfTopStack = false then rhetRelOfTopStackInEngTree <- EVIDENCE
Table 5: Rule examples for the Relation-Reduce classifier Evaluation of the discourse-based transfer module By applying the General classifier or the other six classifiers successively, one can map any Japanese discourse tree into a tree whose structure comes closer to the natural rendering of English. To evaluate the discourse-based transfer module, a ten-fold cross-validation experiment was performed. That is, the classifiers were trained on 36 pairs of manually built and aligned discourse structures, and then the learned classifiers were used in order to map 4 unseen Japanese discourse trees into English-like trees. The similarity of the derived trees and the English trees built manually was measured using the metrics discussed above. This procedure was repeated ten times, each time training and testing on different subsets of tree pairs.
The results reported in Table 2 were as a baseline for the model. The baseline corresponds to applying no knowledge of discourse. Table 6 below displays the absolute improvement (in percentage points) in recall and precision figures obtained when the General classifier was used to map Japanese trees into English-looking trees. The General classifier yielded the best results. The results in Table 6 are averaged over a ten-fold cross-validation experiment.
bevel Units Spans StatusNuclearityRelations P-D P-D P P-D P-D P P-D P-D P P-D P-D
R R R R P
Sentence +9.1 +25.5 +2.0 +19.9 +0.4 +13.4 -0.01+8.4 Paragraph -14.7+1.4 -12.5-1.7 -11.0-2.4 -9.9 -3.3 Text -9.6 -13.5 -7.1 -11.1 -6.3 -10.0 -5.2 -8.8 Weighted Average +1.5 +14.1 -2.1 +9.9 -3.1 +6.4 -3.0 +3.9 All -1.2 +2.5 -0.1 +2.9 +0.6 +3.5 +0.7 +2.6 R P R P R P R P
Sentence +13.4 +30.4 +3.1 +36.1 -6.3 +18.6 -10.1 +3.9 Paragraph -15.6 +0.6 -13.5 -0.8 -11.7 -1.8 -10.3 -2.8 Text -15.4 -23.3 -13.0 -20.4 -13.2 -19.5 -11.5 -17.0 Weighted Average +3.6 +15.5 -2.7 +17.1 -8.5 +7.3 -10.5 -0.4 All +12.7 +29.6 +2.0 +28.8 -5.1 +13.0 -7.9 +2.2 Table 6: Relative evaluation of the discourse-based transfer module with respect to the figures in Table 2 The results in Table 6 show that the model described here outperforms the baseline with respect to building English-like discourse structures for sentences, but it under-performs the baseline with respect to building English-like structures at the paragraph and text levels.
One potential shortcoming of this model seems to come from its low performance in assigning paragraph boundaries.
Because the classifier does not learn correctly which spans to consider paragraphs and which spans not, the recall and precision results at the paragraph and text levels are negatively affected. The poorer results at the paragraph and text levels can be also explained by errors whose effect cumulates during the step-by-step tree-reconstruction procedure; and by the fact that, for these levels, there is less data to learn from.
However, if one ignores the sentence and paragraph boundaries and evaluates the discourse structures overall, one can see that this model outperforms the baseline on all accounts according to the Position-Dependent evaluation;
outperforms the baseline with respect to the assignment of elementary units, hierarchical spans, and nuclearity statures according to the Position-Independent evaluation and under-performs the baseline only slightly with respect to the rhetorical relation assignment according to the Position-Independent evaluation. More sophisticated discourse features, such as those discussed by Maynard, Principles of Japanese Discourse: A Handlaook, Cambridge Univ. Press (1998), for example, and a tighter integration with the lexicogrammar of the two languages may yield better cues for learning discourse-based translation models .
Alternative embodiments of the tree rewriter are possible. For example, probabilities could be incorporated into the tree rewriting procedure. Alternatively, or in addition, multiple trees could be rewritten in parallel and the best one selected in the end. In the current embodiment, the target tree Tt is generated in a sequence of deterministic steps with no recursion or branching.
Alternatively, it is possible to associate a probability with each individual step and reconstruct the target tree Tt by exploring multiple alternatives at the same time. The probability of a target tree Tt is given by the product of the probabilities of all steps that led to that tree. In this case, the target tree Tt will be taken to be the resulting tree of maximum probability. An advantage of such an approach is that it enables the learning of probabilistic transfer functions, H[~.
Although only a few embodiments have been described in detail above, those having ordinary skill in the art will certainly understand that many modifications are possible in the preferred embodiment without departing from the teachings thereof. All such modifications are encompassed within the following claims.
Toward a functional theory of text organization," Text, 8 (3) :243-X81 (1988) .
The example shown in Fig. 6 illustrates the types of structures that may be present in a rhetorical structure tree for a text fragment. The leaves of the tree correspond to elementary discourse units ("edus") and the internal nodes correspond to contiguous text spans. Each node in a rhetorical structure tree is characterized by a "status" (i.e., either "nucleus" or "satellite") and a "rhetorical relation," which is a relation that holds between two non-overlapping text spans. In Fig. 6, nuclei are represented by straight lines while satellites are represented by arcs.
The present inventor recognized that significant differences exist between the rhetorical structures of translations of a text in different languages (e. g., Japanese and English). Accordingly, to improve MT quality, and as a component of a larger MT system, the present inventor developed techniques for automatically rewriting (e. g., using a computer system) rhetorical structures from one language into another, for example, rewriting a rhetorical tree for a text segment in Japanese into a rhetorical tree for a counterpart text segment in English.
Implementations of the disclosed tree rewriting techniques may include various combinations of the following features.
In one aspect, automatically generating a tree (e. g., 5~ either a syntactic tree or a discourse tree) involves receiving as input a tree corresponding to a source language text segment, and applying one or more decision rules to the received input to generate a tree corresponding to a target language text segment.
In another aspect, a computer-implemented tree generation method may include receiving as input a tree corresponding to a source language text segment (e. g., clause, sentence, paragraph, or treatise), and applying one or more decision rules (e. g., a sequence of decision rules that collectively represent a transfer function) to the received input to generate a tree corresponding to a target language text segment, which potentially may be a different type of text segment.
The tree generation method further may include automatically determining the one or more decision rules based on a training set, for example, a plurality of input -output tree pairs and a mapping between each of the input-output tree pairs. The mapping between each of the input-output tree pairs may be a mapping between leaves of the input tree and leaves of the paired output tree. Mappings between leaves of input-output tree pairs can be one-to-one, one-to-many, many-to-one, or many-to-many.
Automatically determining the one or more decision rules may include determining a sequence of operations that generates an output tree when applied to the paired input tree. Determining a sequence of operations may include using a plurality of predefined operations that collectively are sufficient to render any input tree into the input tree's paired output tree. The plurality of predefined operations comprise one or more of the following: a shift operation that transfers an elementary discourse tree (edt) from an input list into a stack; a reduce operation that pops two edts from a top of the stack, combines the two popped edts into a new tree, and pushes the new tree on the top of the stack; a break operation that breaks an edt into a predetermined number of units a create-next operation that creates a target language discourse constituent that has no correspondent in the source language tree; a fuse operation that fuses an edt at the top of the stack into the preceding edt; a swap operation that swaps positions of edts in the input list;
and an assignType operation that assigns one or more of the following types to edts: Unit, Multiunit, Sentence, Paragraph, MultiParagraph, and Text.
The plurality of predefined operations may represent a closed set that includes the shift operation, the reduce operation, the break operation, the create-next operation, the fuse operation, the swap operation and the assignType operation.
Determining a sequence of operations may result in a plurality of learning cases, one learning case for each input-output tree pair. In that case, the tree generation method may further include associating one or more features with each of the plurality of learning cases based on context. The associated features may include one or more of the following: operational and discourse features, correspondence-based features, and lexical features.
The tree generation method may further include applying a learning program (e. g., C4.5) to the plurality of learning cases to generate the one or more decision rules In another aspect, a computer-implemented tree generation module may include a predetermined set of decision rules that, when applied to a tree (e. g., syntactic or discourse) corresponding to a source language text segment, generate a tree corresponding to a target language text segment. The predetermined set of decision rules may define a transfer function between source language trees and target language trees.
In another aspect, determining a transfer function between trees (e. g., syntactic or discourse) of different types may include generating a training set comprising a plurality of tree pairs and a mapping between each tree pair, each tree pair comprises a source tree and a corresponding target tree, and generating a plurality of learning cases by determining, for each tree pair, a sequence of operations that result in the target tree when applied to the source tree; and generating a plurality of decision rules by applying a learning algorithm to the plurality of learning cases.
Determining a transfer function between trees of different types further may include, prior to generating the plurality of decision rules, associating one or more features with each of the learning cases based on context.
In another aspect, a computer-implemented discourse-based machine translation system may include a discourse parser that parses the discourse structure of a source language text segment and generates a source language discourse tree for the text segments a discourse-structure transfer module that accepts the source language discourse tree as input and generates as output a target language discourse tree; and a mapping module that maps the target language discourse tree into a target text segment. The discourse-structure transfer module may include a plurality of decision rules generated from a training set of source language-target language tree pairs.
One or more of the following advantages may be provided by tree rewriting as described herein, The techniques and methods described here may result in a tree rewriting capability that allows users (e. g., human end-users such as linguistic researchers or computer processes such as MT systems) to automatically have a tree for a text segment in a source language rewritten, or translated, into a tree for the text segment translated into a target language. This functionality is useful both in its standalone form and as a component of a larger system, such as in a discourse-based machine translation system.
Moreover, because the tree rewriter described here automatically learns how t,o rewrite trees from one language into another, the system is easy and convenient to use.
The mapping scheme used in training the tree rewriter also provides several advantages. For example, by allowing any arbitrary groupings (e. g., one-to-one, one-to-many, many-to-one, many-to-many) between leaves in th,e source and target trees, the flexibility, richness and robustness of the resulting mappings are enhanced.
The enhanced shift-reduce operations used in training the tree rewriter also provide several advantages. For example, the set of basic operations that collectively are sufficient to render any input tree into its paired output tree provides a powerful yet compact tool for rewriting tree structures.
According to one aspect of the present invention, there is provided a machine translation decoding method comprising: receiving as input a text segment in a source language to be translated into a target language; generating an initial translation as a current target language translation using translation information learned in part from text corpora; applying one or more modification operators to the current target language translation to generate one or more modified target language translations;
determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations; setting a modified target language translation as the current target language translation; and repeating said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
According to another aspect of the present invention, there is provided a computer-implemented machine translation decoding method comprising iteratively modifying a target language translation of a source language text segment until an occurrence of a termination condition;
wherein iteratively modifying the translation comprises performing at each iteration modification operations on the translation to form modified translations, the modification operations including changing both word translation and word order, and performing at each iteration an evaluation of the modified translations taking into account both source-to-target-language translation information and target-language structure information learned both in part from text corpora.
According to still another aspect of the present invention, there is provided a machine translation decoder comprising: a decoding engine comprising one or more modification operators to be applied to a current target language translation to generate one or more modified target language translations; and a process loop to iteratively modify the current target language translation using the one or more modification operators, the process loop being configured to effect operations comprising: applying the one or more modification operators to the current target language translation using the decoding engine; determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations; setting a modified target language translation as the current target language translation; and repeating 16a said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
The details of one or more embodiments are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.
Drawing Descriptions These and other aspects of the invention will now be described in detail with reference to the accompanying drawings, wherein:
Figure 1A shows a block diagram of machine translation from a user's perspective.
Figure 1B shows an example of word-level alignment.
Figure 2 shows a flowchart of the operation of an embodiment of the greedy decoder.
Figure 3 shows an example of the greedy decoder producing an English translation of a French sentence.
Figure 4 shows an example of output that a user sees as the greedy decoder produces an English translation of a French sentence.
16b ' 75973-18 Figure 5 is a table showing a comparison between different decoders using a trigram language model.
16c Figure 6 shows an example of a rhetorical structure tree.
Figure 7 is an example of a Japanese source sentence.
Figure 0 is the discourse structure of the Japanese source sentence in Fig. 7.
Figure 9 is the discourse structure of an English target sentence translated from the Japanese source sentence of Fig. 11.
Figure 10 shows a block diagram of the tree rewriter.
Figure 11 shows a block diagram of how a tree rewriter can be used as a subsystem of a larger system.
Figure 12 shows a block diagram of a discourse-based machine translation system with the tree rewriter as a subsystem.
Figure 13 is a flowchart of a procedure for building a tree rewriter.
Figure 14 shows an example of incremental tree reconstruction.
Figure 15 is a graph of the learning curve for the Relation-Reduce classifier.
Detailed Description Greedy Decoder A statistical MT system that translates, for example, French sentences into English, can be divided into three parts: (1) a language model (LM) that assigns a probability P(e) to any English string, (2) a translation model (TM) that assigns a probability P(f~e) to any pair of English and French strings, and (3) a decoder. The decoder takes a previously unseen sentence f and tries to find the a that maximizes P(e~f), or equivalently maximizes P(e) ~ P(f~e).
Brown et al., "The mathematics of statistical machine translation: Parameter estimation", Computational .Linguistics, 19(2), 1993, introduced a series of TMs based on word-for-word substitution and re-ordering, but did not include a decoding algorithm. If the source and target languages are constrained to have the same word order (by choice or through suitable pre-processing), then the linear Viterbi algorithm can be applied as described in Tillmann et al., "A DP-based search using monotone alignments in statistical translation", In Proc. ACL, 1997. If re-ordering is limited to rotations around nodes in a binary tree, then optimal decoding can be carried out by a high-polynomial algorithm (Wu, "A polynomial-time algorithm for statistical machine translation", In Pros. ACL, 1996). For arbitrary word-reordering, the decoding problem is NP-complete (nondeterministiC polynomial time complete) (Knight, "Decoding complexity in word-replacement translation models", Computational Linguistics, 25(4), 1999) .
One strategy (Brown et al., "Method and system for natural language translation," U.S. Patent 5,477,451, 1995;
Wang et al., "Decoding algorithm in statistical machine translation", In Proc. AC.L, 1997) is to examine a large subset of likely decodings and chose just from that. Of course, it is possible to miss a good translation this way.
Thus, while decoding is a clear-cut optimization task in which every problem instance has a right answer, it is hard to come up with good answers quickly. The following sets forth details of a fast greedy decoder and compares its performance with that of a traditional stack decoder.
In developing the greedy decoder IBM Model 4 was used, which revolves around the notion of a word alignment over a pair of sentences (see Figure 1B). A word alignment assigns a single home (English string position) to each French word. If two French words align to the same English word, then that English word is said to have a fertility of two. Likewise, if an English word remains unaligned-to, then it has fertility zero. The word alignment in FigurelB
is shorthand for a hypothetical stochastic process by which an English string gets converted into a French string.
There are several sets of decisions to be made.
First, every English word is assigned a fertility.
These assignments are made stochastically according to a table n(o~e~). Any word with fertility zero is deleted from the strong, any work with fertility two is duplicated, etc.
If a word has fertility greater than one, it is called very fertile.
After each English word in the new string, the fertility of an invisible English NULL element with probability p1 (typically about 0.02) is incremented. The NULL element ultimately produces "spurious" French words.
Next, a word-for-word replacement of English words (including NULL) by French words is performed, according to the table t (f~ ~ e;,) .
Finally, the French words are permuted. In permuting, IBM Model 4 distinguishes between French words that are heads (the leftmost French word generated from a particular English word), non-heads (non-leftmost, generated only by very fertile English words), and NU.LZ-generated.
Heads. The head of one English word is assigned a French string position based on the position assigned to the previous English word. If an English word Ee-~
translates into something at French position j, then the French head word of ei is stochastically placed in French position k with distortion probability dl(k-j~class(ei-i).
class (fk)), where "class" refers to automatically determined word classes for French and English vocabulary items. This relative offset k-j encourages adjacent English words to translate into adjacent French words. If ei_1 is infertile, then j is taken from e;,_~, etc. If ei_1 is very fertile, then j is the average of the positions of its French translations.
Non-heads. If the head of English word ei is placed in French position j, then its first non-head is placed in French position k (>j) according to another table d,~(k-j~class (fk)). The next non-head is placed at position q with probability d,1(q-k~class (fq)), and so forth.
NULh-generated. After heads and non-heads are placed, NULL-generated words are permuted into the remaining vacant slots randomly. If there are ~6o NULL-generated words, then any placement scheme is chosen with probability 1/~0!.
These stochastic decisions, starting with e, result in different choices of f and an alignment of f with e. a is mapped onto a particular <a,f> pair with probability:
p (a~ f ~ e) r r ~, ~n(~i ~ ei)x~~t(~ik ~ ~i)x i=1 i=1 k=1 I
dl (~;1- cp ~ class(e p ), class(~;1))x t=i,~; >o r ~;
d>~ (irk - ~i~k-n I class(zik ))x i=1 k=z m Wo plea (1 _ p1 )"' z~ x ~o t(zok ~LL) k=1 where the factors separated by x symbols denote fertility, translation, head permutation, non-head permutation, null-fertilit6y, and null-translation probabilities. The symbols in this formula are: 1 (the length of e), m (the length of f), ei (the ith English word in e), Eo (the NUZZ
word) , s~i (the fertility of ei~, ~o (the fertility of the NUZZ word) , iik (the kth French word produced by ei in a) , mix (the position of iik in f) , pI (the position of the first fertile word to the left of ei in a) , c,,I (the ceiling of the average of all ~tp;,k for pi, or 0 if pi is undefined) .
In view of the foregoing, given a new sentence f, then an optimal decoder will search for an a that maximizes Pelf) ... p(e) ~ P(f~e) . Here, P(f~e) is the sum of P(a,f~e) over all possible alignments a. Because this sum involves significant computation, typically it is avoided by instead searching for an <e,a> pair that maximizes P~(e,a~f) ~ P(e) P(a,f~e). It is assumed that the language model P(e) is a smoothed n-gram model of English.
Figure 2 is a flowchart of the operation of an embodiment of a greedy decoder for performing MT. As shown therein, the first step 200 is to receive an input sentence to be translated. Although in this example, the text segment being translated is a sentence, virtually any other text segment could be used, for example, clauses, paragraphs or entire treatises.
In step 202, as a first approximation of the translation, the greedy decoder generates a "gloss" of the input sentence, essentially a word-for-word translation.
The gloss is constructed by aligning each French word f~
with its most likely English translation ef~ (ef~ = argmaxe t(e ~ f~)). For example, in translating the French sentence "Bien entendu, i1 parle de une belle victoire", the greedy decoder initially assumes that a good translation is ' Well heard, it talking a beautiful victory " because the best translation of "bien" is "well", the best translation of "entendu" is "heard", and so on. The alignment corresponding to this translation is shown at the top of Figure 3.
In step 204, the decoder estimates the probability of correctness of the current translation, P(c).
After the initial alignment is generated in step 202, the greedy decoder tries to improve it in step 206. That is, the decoder tries to find an alignment (and implicitly, a translation) of higher probability by applying one or more sentence modification operators, described below. The use of a word-level alignment and the particular operators described below were chosen for this particular embodiment.
However, alternative embodiments using different statistical models may benefit from different or additional operations.
The following operators collectively make-up the decoder's translation engine, and include the following:
translateOneOrTwoWoreis ( jz, e1, j2, e~) This operation changes the translation of one or two French words, those located at positions j1 and j~, from ef~l and ef~2 into e1 and e~. If ef~ is a word of fertility 1 and ek is NULL, then ef~ is deleted from the translation. If ef~
is the NULL word, the word ek is inserted into the translation at the position that yields an alignment of highest probability. If ef~~ = ei or ef~~ = e2, then this operation amounts to changing the translation of a single word.
translateAndlnsert ( j, e1, e~) This operation changes the translation of the French word located at position j from ef~ into e1 and simultaneously inserts word e2 at the position that yields the alignment of highest probability. ~nlord e2 is selected from an automatically derived list of 1024 words with high probability of having fertility 0. When ef~ = e1, this operation amounts to inserting a word of fertility 0 into the alignment.
removeWordOfFertility0 (i) This operation deletes the word of fertility 0 at position i in the current alignment.
swapSegments ( iz, i~, This operation creates a new alignment from the old one by swapping non-overlapping English word segments [i1, i~] and [j1, j2]. During the swap operation, all existing links between English and French words are preserved. The segments can be as small as a word or as long as ~e~-1 words, where ~e~ is the length of the English sentence.
joinWords ( i1, ia) This operation eliminates from the alignment the English word at position i1 (or i~) and links the French words generated by eil ( or ei~ ) to ei~ ( or ei1 ) .
At step 208, the decoder estimates the probabilities of correctness, P (M~) ... P (M") , for each of the results of the sentence modification operations. That is, the probability for each new resulting translation is determined At step 210, the decoder determines whether any of the new translations are better than the current translation by comparing their respective probabilities of correctness.
If any of the new translations represents a better solution than the current translation, then the best new translation (that is, the translation solution having the highest probability of correctness) is set as the current translation in step 214 and the decoding process returns to step 206 to perform one or more of the sentence modification operations on the new current translation solution.
Steps 206, 208, 210 and 214 repeat until the sentence modification operations cease (as determined in step 210) to produce translation solutions having higher probabilities of correctness, at which point, the decoding process halts at step 212 and the current translation is output as the final decoding solution. Alternatively, the decoder could cease after a predetermined number of iterations chosen, for example, either by a human end-user or by an application program using the decoder as a translation engine.
Accordingly, in a stepwise fashion, starting from the initial gloss, the greedy decoder uses a process loop (e.g., as shown in Fig. 2, steps 206, 208, 210 and 214) to iterate exhaustively over all alignments that are one operation away from the alignment under consideration. At every step, the decoder chooses the alignment of highest probability, until the probability of the current alignment can no longer be improved. When it starts from the gloss of the French sentence "Bien entendu, i1 parle de une belle victoire.", for example, the greedy decoder alters the initial alignment incrementally as shown in Figure 3, eventually producing the translation "Quite naturally, he talks about a great victory.". In this process, the decoder explores a total of 77421 distinct alignments /
translations, of which "Quite naturally, he talks about a great victory." has the highest probability.
In step 206 of the decoding process, either all of the five sentence modification operations can be used or any subset thereof may be used to the exclusion of the others, depending on the preferences of the system designer and/or end-user. For example, the most time consuming operations in the decoder are swapSegments, translateOneOrTwoWords, and translateAndlnsert. SwapSegments iterates over all possible non-overlapping span pairs that can be built on a sequence of length I a I. TranslateOneOrTwoWords iterates over I f 12 x I t I~ alignments, where I f I is the size of the French sentence and I t I is the number of translations associated with each word (in this implementation, this number is limited to the top 10 translations).
TranslateAndlnsert iterates over I f I x I t I x I z I
alignments, where I z I is the size of the list of words with high probability of having fertility 0 (1024 words in this implementation). Accordingly, the decoder could be designed to omit one or more of these slower operations in order to speed up decoding, but potentially at the cost of accuracy. Alternatively, or in addition, the decoder could be designed to use different or additional sentence modification operations according to the objectives of the system designer and/or end-user.
An advantage of the greedy decoder comes from its speed. As the experiments described below demonstrate, the greedy decoder can produce translations faster than any other decoder. The greedy decoder is an instance of an "anytime algorithm" - the longer it runs, the better the translation it finds. One potential tradeoff of the greedy decoder relates to the size of the solution space that it explores, which is relatively small. The farther away a good translation is from the initial gloss, the less likely the greedy decoder is to find it.
Fig. 4 shows another example of the greedy decoder in action in which an acceptable solution is reached in four iterations. As shown therein, the input sentence to be translated is "ce ne est pas juste." The decoder uses the initial gloss "that not is not fair." and determines that this translation solution (Iteration 1) has a probability of correctness ("Aprob" - the product of LMprob and TMprob) of 1.13162e-22, based on a language model probability (LMprob) of 2.98457e-14 and a translation model probability (TMprob) of 3.79156e-09.
In the second iteration, the decoder changes the first instance of the word "not" in the translation to "is" by applying the translateOneOrTwoWords operation, resulting in a new translation solution "that is is not fair", having the probabilities shown in Fig. 4, Iteration 2. In the third iteration, the decoder applies the removeWordOfFertility0 operation and drops one instance of the word "is" in the translation, resulting in a new translation solution "that is not fair", having the probabilities shown in Fig. 4, Iteration 3. In the fourth and final iteration, the decoder applies the translateOneOrTcvoWords operation again to change the word "that" in the translation to "it", resulting in the final translation solution "it is not fair", having the probabilities shown in Fig. 4, Iteration 4.
To determine the performance of the greedy decoder, a series of experiments was performed. In all experiments, decoding was performed using only the top 10 translations of a word, as determined during training, and a list of 1024 words of fertility 0, which was also extracted automatically from the test corpus.
In experiments to determine the accuracy and compare the speed of the greedy decoder described to a conventional stack decoder (such as described in U.S. Patent No.
5,477,451 to Brown et al., a test collection of 505 sentences was used, uniformly distributed across the lengths 6, S, 10, 15, and 20. Decoders were evaluated with respect to (1) speed, and (2) translation accuracy.
The results in the table shown in Fig. 5, obtained with decoders that use a trigram language model, show that the greedy decoding algorithm is an advantageous alternative to the traditional stack decoding algorithm.
Even when the greedy decoder used an optimized-for-speed set of operations (that is, a subset of the total set of five sentence modification operations discussed above) in which at most one word is translated, moved, or inserted at a time - which is labeled "greedy*" in Fig. 5 - the translation accuracy is affected only slightly. In contrast, the translation speed increases with at least one order of magnitude. Depending on the application of interest, one may choose to use a slow decoder that provides optimal results or a fast, greedy decoder that provides non-optimal, but acceptable results.
Alternative embodiments for the greedy decoder are possible. For example, the greedy decoder could start with multiple different initial translations (e. g., different variations on the gloss used in step 202 in Fig. 2) and then run the greedy decoding algorithm (i.e., steps 204-214 in Fig. 2) on each of the different initial translations in parallel. For example, the greedy decoder code start with an initial, approximate translation selected from among multiple translated phrases stored in a memory. In the end, the best translation could be selected. This parallel translation of different initial solutions could result in more accurate translations.
Tree Rewriter Almost all conventional MT systems process text one sentence at a time. because of this limited focus, MT
systems generally cannot re-group and re-order the clauses and sentences of an input text to achieve the most natural rendering in a target language. Yet, even between languages as close as English and French, there is a 100 mismatch in number of sentences - what is said in two sentences in one language is said in only one, or in three, in another (Gale et al., "A program for aligning sentences in bilingual corpora," Computational .Linguistics, 19(1):75-102, 1993). For distant language pairs, such as Japanese and English, the differences are more significant.
Consider, for example, the Japanese sentence shown in Fig . 7 ( "text ( 1 ) ") . The following ( "text ( 2 ) ") is a word-by-word "gloss" of text (1):
(2) [The Ministry of Health and Welfare last year revealed]
[population of future estimate according toy] [in future 1.499 persons as the lowest3] [that after *SAB* rising to turn that4] [*they*
estimated buts] [already the estimate misses a point ]
[prediction became.']
In contrast, a two-sentence translation of the Japanese sentence produced by a professional translator ("text(3)") reads as follows:
(3) [In its future population estimatesl] [made public last year,] [the Ministry of Health and Welfare predicted that the SAB
would drop to a new low of 1.499 in the future, 3] [but would make a comeback after that,4] [increasing once again.5] [However, it looks as if that prediction will be quickly shattered.6]
The labeled spans of text represent elementary discourse units (edus), i.e., minimal text spans that have an unambiguous discourse function (Mann et al., "Rhetorical structure theory: Toward a functional theory of text organization," Text, 8(3):243-281, 1988). If the text fragments are analyzed closely, one notices that in translating text (1), a professional translator chose to realize the information in Japanese unit 2 first (unit 2 in text (1) corresponds roughly to unit 1 in text (3)); to realize then some of the information in Japanese unit 1 (part of unit 1 in text (1) corresponds to unit 2 in text (3)); to refuse then information given in units 1, 3 and 5 in text (1) and realize it in English as unit 3; and so on.
Also, the translator chose to re-package the information in the original Japanese sentence into two English sentences.
At the elementary unit level, the correspondence between the Japanese sentence in text (1) and its English translation in text (3) can be represented as in mappings (4), below, where j c a denotes the fact that the semantic content of unit j is realized fully in unit e; j ~ a denotes the fact that the semantic content of unit a is realized fully in unit j; j = a denotes the fact that units j and a are semantically equivalent; and j - a denotes the fact that there is a semantic overlap between units j and e, but neither proper inclusion not proper equivalence.
(4) js ~ ez: j1 j2 = ei;
js ~ es;
j4 = en: j4 = es:
j5 = es:
js ~ es:
j~
Hence, the mappings in (4) provide an explicit representation of the way information is re-ordered and re-packaged when translated from Japanese into English.
However, when translating text, it is also the case that the rhetorical rendering changes. What is realized in Japanese using a CONTRAST relation can be realized in English using, for example, a COMPARISON OR A CONCESSION
relation.
Figures 8 and 9 present in the style of Mann, supra, the discourse structures of text fragments (1) and (3), above. Each discourse structure is a tree whose leaves correspond to contiguous text spans. Each node is characterized by a status (NUCLEUS or SATELLITE) and a rhetorical relation, which is a relation that holds between two non-overlapping text spans. The distinction between nuclei and satellites comes from the empirical observation that the nucleus expresses what is more essential to the writer's intention than the satellite; and that the nucleus of a rhetorical relation is comprehensible independent of the satellite, but not vice versa. When spans are equally important, the relation is multinuclear; for example, the CONTRAST relation that holds between unit [3] and span [4,5] in the rhetorical structure of the English text in Figures 8 and 9 is multinuclear. Rhetorical relations that end in the suffix "-e" denote relations that correspond to embedded syntactic constituents. For example, the ELABORATION-OBJECT-ATTRIBUTE-E relation that holds between units 2 and 1 in the English discourse structure corresponds to a restrictive relative.
If one knows the mappings at the edu level, one can determine the mappings at the span (discourse constituent) level as well. For example, using the elementary mappings in (4), one can determine that Japanese span [1,2]
corresponds to the English span [1,2], Japanese unit [4] to English span [4,5], Japanese span [6,7] to English unit [6], Japanese span [1,5] to English span [1,5], and so on.
As Figures 8 and 9 show, the CONCESSION relation that holds between spans [1,5] and [6,7] in the Japanese tree corresponds to a similar relation that holds between span [1,5], and unit [6] in the English tree (modulo the fact that, in Japanese, the relation holds between sentence fragments, while in English it holds between full sentences). However, the TEMPORAL-AFTER relation that holds between units [3] and [4] in the Japanese tree is realized as a CONTRAST relation between unit [3] and span [4,5] in the English tree. And because Japanese units {6]
and [7] are fused into unit [6] in English, the relation ELABORATION-OBJECT-ATTRIBUTE-E is no longer made explicit in the English text.
Some of the differences between the two discourse trees in Figures S and 9 have been traditionally addressed in MT systems at the syntactic level. For example, the re-ordering of units 1 and 2 can be dealt with using only syntactic models. However, as discussed below, there are significant differences between Japanese and English with respect to the way information is packaged and organized rhetorically not only at the sentence level, but also, at the paragraph and text levels. More specifically, as humans translate Japanese into English, they re-order the clauses, sentences, and paragraphs of Japanese texts, they re package the information into clauses, sentences and paragraphs that are not a one-to-one mapping of the original Japanese units, and they rhetorically re-organize the structure of the translated text so as to reflect rhetorical constraints specific to English. If a translation system is to produce text that is not only grammatical but also coherent, it will have to ensure that the discourse structure of the target text reflects the natural renderings of the target language, and not that of the source language.
In the Experiment section below, it is empirically shown that there are significant differences between the rhetorical structure of Japanese texts and their corresponding English translations. These differences further demonstrate the need and desirability of developing computational models for discourse structure rewriting.
Experiment In order to assess the role of discourse structure in MT, a corpus of discourse trees was manually built for 40 Japanese texts and their corresponding translations. The texts were selected randomly from the ARPA corpus (White et al., "Evaluation in the ARPA machine-translation program:
1993 methodology," In Proceedings of the ARPA Human Language Technology Workshop, pages 135-140, Washington, D.C., 1994). On average, each text had about 460 words.
The Japanese texts had a total of 335 paragraphs and 773 sentences. The English texts had a total of 337 paragraphs and 827 sentences.
A discourse annotation protocol was developed for Japanese and English along the lines followed by Marcu et al., "Experiments in constructing a corpus of discourse trees, " In Proc. Of the ~ICL'99 Workshop on Standards and Tools for Discourse Tagging, pages 48-57, Maryland (1999).
Marcus discourse annotation tool (1999) was used in order to manually construct the discourse structure of all Japanese and English texts in the corpus. Ten percent of the Japanese and English texts were rhetorically labeled by two annotators. The tool and the annotation protocol are available at http://www.isi.edu/~marcu/software/
The annotation procedure yielded over the entire corpus 2641 Japanese edus and 2363 English edus.
Corpus k" (3) ks (#) -I- kn (#)-- I kr (#) Japanese 0.856 (80) 0.785 (3377) 0.724 (3377) 0.650 (3377) English 0.925 (60) 0.866 (1826) 0.839 (1826) 0.748 (1826) Table 1: Tagging reliability The reliability of the annotation was computer using Marcu et al. (1999)'s method for computing kappa statistics (Siegel et al., Non parametric Statistics for the Behavioral Sciences, McGraw-Hill, Second edition, 1988) over-hierarchical structures. Table 1 above displays average kappa statistics that reflect the reliability of the annotation of elementary discourse units, k", hierarchical discourse spans, ks, hierarchical nuclearity assignments, k", and hierarchical rhetorical relation assignments, kr. Kappa figures higher than 0.8 correspond to good agreement; kappa figures higher than 0.6 correspond to acceptable agreement. All kappa statistics were statistically significant at levels higher than a = 0.01.
In addition to the kappa statistics, table 1 also displays in parentheses the average number of data points per document, over which the kappa statistics were computed.
For each pair of Japanese-English discourse structures, an alignment file was also built manually, which. specified in the notation discussed on page 1 the correspondence between the edus of the Japanese text and the edus of the English translation.
The similarity between English and Japanese discourse trees was computed using labeled recall and precision figures that reflected the resemblance of the Japanese and English discourse structures with respect to their assignment of edu boundaries, hierarchical spans, nuclearity, and rhetorical relations.
Because the trees compared differ from one language to the other in the number of elementary units, the order of these units, and the way the units are grouped recursively into discourse spans, two types of recall and precision figures were computed. In computing Position-Dependent (P-D) recall and precision figures, a Japanese span was considered to match an English span when the Japanese span contained all the Japanese edus that corresponded to the edus in the English span, and when the Japanese and English spans appeared in the same position with respect to the overall structure. For example, the English tree in Figures 8 and 9 are characterized by 10 subsentential spans: [1], [2], [3], [4], [5], [6], [1,2], [4,5], [3,5], and [1,5]. (Span [1,6] subsumes 2 sentences, so it is not sub-sentential.) The Japanese discourse tree has only 4 spans that could be matched in the same positions with English spans, namely spans [1,2], [4], [5], and [1,5].
Hence the similarity between the Japanese tree and the English tree with respect to their discourse below the sentence level has a recall of 4/10 and a precision of 4/11 (in Figures 8 and 9, there are 11 sub-sentential Japanese spans ) .
In computing Position-Independent (P-I) recall and precision figures, even when a Japanese span "floated"
during the translation to a position in the English tree, the P-I recall and precision figures were not affected.
The Position-Independent figures reflect the intuition that if two trees t1 and t2 both have a subtree t, t1 and t~ are more similar than if they were if they didn't share a tree.
At the sentence level, it is assumed that if, for example, the syntactic structure of a relative clause is translated appropriately (even though it is not appropriately attached), this is better than translating wrongly that clause. The Position-Independent figures offer a more optimistic metric for comparing discourse trees. They span a wider range of values than the Position-Dependent figures, which enables a better characterization of the differences between Japanese and English discourse structures. When one takes an optimistic stance, for the spans at the sub-sentential level in the trees in Table 1 the recall is 6/10 and the precision is 6/11 because in addition to spans [1,2], [4], [5], and [1,5], one can also match Japanese span [1] to English span [2] and Japanese span [ 2 ] to Japanese span [ 1 ] .
In order to provide a better estimate of how close two discourse trees were, Position-Dependent and -Independent reoall and precision figures were computed for the sentential level (where units are given by edus and spans are given by sets of edus or single sentences); paragraph level (where units are given by sentences and spans are given by sets of sentences or single paragraphs); and text level (where units are given by paragraphs and spans are given by sets of paragraphs). These figures offer a detailed picture of how discourse structures and relations are mapped from one language to the other across all discourse levels, from sentence to text. The differences at the sentence level can be explained by differences between the syntactic structures of Japanese and English.
The differences at the paragraph and text levels have a purely rhetorical explanation.
When the recall and precision figures were computed with respect to the nuclearity and relation assignments, the statuses and the rhetorical relations that labeled each pair of spans were also factored in.
Units Spans Status/NUClearityRelations Level P-D P-D P-D P-D P P-D P-D P P-D P-D
R P R R R P
Sentence 29.1 25.0 27.222.7 27..317.7 14.9 12.4 Paragraph 53.9 53.4 46.847.3 38.6 39.0 31.9 32.3 Text 41.3 42.6 31.532.6 28.8 29.9 26.1 27.1 Weighted Average 36.0 32.5 31.828.4 26.0 23.1 20.1 17.9 All 8.2 7.4 5.9 5.3 4.4 3.9 3.3 3.0 P-I P-TP P-I P-IP P-I .__..P-IP P-T P-IP
R R R .._ R
Sentence 71.0 61.0 56.0 46.6 44.3 36.9 30.5 25.4 Paragraph62.1 61.6 53.2 53.8 43.3 43.8 35.1 35.5 Text 74.1 76.5 54.4 56.5 48.5 50.4 41.1 42.7 Weighted 69.6 63.0 55.2 49.2 44.8 39.9 33.1 29.5 Average All 74.5 66.8 50.6 45.8 39.4 35.7 26.8 24.3 Table 2: Similarity of Japanese and English discourse structures Table 2 above summarizes the results (P-D and P-I
(R)ecall and (P)recision figures) for each level (Sentence, Paragraph, and Text). The numbers in the "Weighted Average" line report averages of the Sentence-, Paragraph-, and Text-specific figures, weighted according to the number of units at each level. The numbers in the "All" line reflect recall and precision figures computed across the entire trees, with no attention paid to sentence and paragraph boundaries.
Given the significantly different syntactic structures of Japanese and English the recall and precision results were low, reflecting the similarity between discourse trees built below the sentence level. However, as Table 2 shows, these are significant differences between discourse trees at the paragraph and text levels as well. For example, the Position-Independent figures show that only about 620 of the sentences and only about 530 of the hierarchical spans built across sentences could be matched between the two corpora. when one looks at the status and rhetorical relations associated with the spans built across sentences at the paragraph level, the P-I recall and precision figures drop to about 43% and 35o respectively.
The differences in recall and precision are explained both by differences in the way information is packaged into paragraphs in the two languages and the way it is structured rhetorically both within and about the paragraph level.
These results strongly suggest that if one attempts to translate Japanese into English on a sentence-by-sentence basis, it is likely that the resulting text will be unnatural from a discourse perspective. For example, if some information rendered using a CONTRAST relation in Japanese is rendered using an ELABORATION relation in English, it would be inappropriate to use a discourse marker like "but" in the English translation, although that would be consistent with the Japanese discourse structure.
An inspection of the rhetorical mappings between Japanese and English revealed that some Japanese rhetorical renderings are consistently mapped into one or a few preferred renderings in English. For example, 34 of 115 CONTRAST relations in the Japanese texts are mapped into CONTRAST relations in English; 27 become nuclei of relations such as ANTITHESIS and CONCESSION, 14 are translated as COMPARISON relations, 6 as satellites of CONCESSION relations, 5 as LIST relations, etc.
The discourse-based transfer model Figure 10 is a block diagram of a tree rewriter in the process of being trained. As shown therein, the tree rewriter 700 takes as input two different types of trees, for example, a tree of type A and another tree of type B, and automatically learns how to rewrite type A trees into type B trees. The tree rewriter 700 produces as output a transfer function, H(A ~ B), for rewriting trees of type A
into trees of type B. Accordingly, assuming type A
corresponds to trees in Japanese and type B corresponds to trees in English, H(A -~ B) enables a user (e.g., either a human end-user or a software application invoking the tree rewriter) to automatically convert any tree structure in English into a counterpart tree structure in Japanese.
The tree rewriter works on syntactic trees, rhetorical trees, and virtually any other type of tree structure used in computational linguistics. The tree rewriter has applications not only in machine translation, but also in summarization, discourse parsing, syntactic parsing, information retrieval, automatic test scoring and other applications that generate or use trees. In machine translation, for example, the tree rewriter can be used to rewrite a syntactic / rhetorical tree specific to one language into a syntactic / rhetorical tree for another language. In summarization, the tree rewriter can be used to rewrite the discourse / syntactic structure of a long text or sentence into the discourse / syntactic structure of a short text or sentence.
This high degree of general applicability is shown in Fig. 11, in which the tree rewriter 801, after it has been trained to learn the transfer function H(Tree ~ Tree'), can accept a tree as input from any application 800 that generates trees as output. At the output side, the tree rewriter's output (tree' - a rewritten version of the input tree) can be used as input to any application that uses trees as input.
Figure 12 shows a block diagram of a specific application for the tree rewriter as a component of a larger system - namely, a discourse-based machine translation system. Unlike conventional MT systems which in effect adopt a "tiled" approach to translation, for example, by translating individual sentences of a larger work (e.g., a treatise) separately, the discourse-based MT
system of Figure 12 translates the entire text as whole, potentially resulting in a translation that has a different number and/or arrangement of sentences than the original, but which better captures the underlying discourse, or rhetoric, of the original text.
As shown in Figure 12, the discourse-based MT system 910 receives as input a source language text 900 and produces as output a target language text 908, which is a discourse-based translation of the source language text 900. The system 910 includes three basic components - a discourse parser 902, a discourse-structure transfer module 904 (i.e., a specific instance of a tree rewriter that has been trained to rewrite trees using the transfer function H(Tree -~ Tree')) and a target language tree-text mapper 906.
The discourse parser 902 initially derives the discourse structure of the source language text and produces as output a corresponding discourse tree. Details of a discourse parser that can be used as discourse parser 902 are set forth in Daniel Marcu, "A Decision-Based Approach to Rhetorical Parsing," Proceedings of ACL '99 (1999) .
The target language tree-text mapper 906 is a statistical module that maps the input text into the target language using translation and language models that incorporate discourse-specific features, which are extracted from the outputs of the discourse parser 209 and the discourse-structure transfer module 904. Details of a suitable mapper 906 are set forth in Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, Kenji Yamada, "Fast Decoding and Optimal Decoding for Machine Translation," Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics, July 6-11, 2001.
As noted above, the discourse-structure transfer module 904 is a specific instance of a tree rewriter that has been trained to rewrite trees from the desired input type into the desired output type. More specifically, the discourse-structure transfer module 904 rewrites the discourse structure of the input text so as to reflect a discourse rendering that is natural to the target text.
Fig. 13 is a flow chart illustrating a process 1300 that can be used to train a tree rewriter to automatically learn the transfer function between two different types of tree structures, e.g., a tree of type A
and a tree of type B.
As shown in Fig. 13, the first step 1301 is to generate a training set of input-output pairs of trees [TS, Tt] and a mapping C between leaves each input-output tree pair. The input tree of the pair is of the type from which conversion is desired or, in other words, the source tree type TS. The output tree of the pair is of the type to which conversion is desired or, in other words, the target tree type Tt .
The mapping C between leaves of an input tree and its paired output tree defines a correspondence between a source text segment and its counterpart target language translation. These mappings either can be generated manually as described below, or automatically, as described in Kevin Knight and Daniel Marcu, "Statistics-Based Summarization - Step One: Sentence Compression," The l7tn National Conference on Artificial Intelligence (AAAI-2000), pp. 703-710.
Available types of mappings between leaves of a Japanese-English input-output pair that can be used are shown in equation (4) above, wherein j refers to a Japanese text segment and a refers to an English translation of that text segment. Note that the mappings represented by equation (4) are not limited to one-to-one mappings but rather can be any arbitrary mapping - that is, not only one-to-one, but also one-to-many, many-to-one and many-to-many. This flexibility in the mappings dramatically enhances the richness with which relationships between the input and output trees are defined, and further enhanoes the flexibility of the transfer function H[] that is automatically learned.
After the training set (input-output tree pairs and mappings therebetween) has been generated, the training process next determines in step 1303 the grouping and order of operations that generate a given output tree starting with its paired input tree. This step is performed based on the following seven basic operations, referred to collectively as the "extended shift-reduce" operations -shift, reduce, break, create-next, fuse, swap, and assignType - which are described in detail below under the section heading "The discourse-based transfer model."
These seven operations are sufficient to rewrite any given input tree into its paired output tree.
The output of step 1303 is a set of learning cases -one learning case for each input-output tree pair in the training set. In essence, each learning case is an ordered set of extended shift-reduce operations that when applied to an input tree will generate the paired output tree.
Next, in step 1305, the tree rewriter training process 1300 associates features (e. g., operational and discourse features, correspondence-based features, and lexical features) with the learning cases to reflect the context in which the operations are to be performed. Details of step 1305 are discussed below under the heading "Learning the parameters of the discourse-transfer model."
Next, in step 1307, the tree rewriter training process 1300 applies a learning algorithm, for example, the C4.5 algorithm as described in J. Ross Quinlan, "C4.5: Programs for Machine Learning," Morgan Kaufmann Publishers (1993), to learn a set of decision rules from the learning cases.
Details of step 1307 are discussed below under the heading "Learning the parameters of the discourse-transfer model."
This set of decision rules collectively constitute the transfer function, H (TS -~ Tt) , for rewriting any tree of type TS into a tree of type Tt. This transfer function can then be used by users, applications or other automated processes for rewriting previously unseen trees of type TS
into trees of type Tt.
A more detailed discussion of training a tree rewriter follows.
In order to learn to rewrite discourse structure trees, a related problem, defined below in Definition 3.1, is addressed.
Definition 3.1 Given two trees TS
and Tt and a correspondence Table C
defined between TS and Tt a t the leaf level in terms of =, c, and - reactions, find a sequence of actions that rewrites the tree TS
into T~.
If for any tuple (TS, Tt, C) such a sequence of actions can be derived, it is then possible to use a corpus of (TS, Tt, C) tuples in order to automatically learn to derive from an unseen tree Tsi, which has the same structural properties as the trees TS, a tree Tt~, which has structural properties similar to those of the trees Tt.
Solving the problem in definition 3.1 involves, in part, extending the shift-reduce parsing paradigm applied by Mangerman, "Statistical decision-tree models for parsing," In Proo. Of ACL'95, pages 276-283, Cambridge, Massachusetts (1995), Hermjak~b et al., "Learning parse and translation decisions from examples with rich context," In Proc. Of ACL'97, pages 482-489, Madrid, Spain (1997), and Marcu, "A decision-based approach to rhetorical parsing,"
In Proc. Of ACL'99, pages 365-372, Maryland (1999). In this extended paradigm, the transfer process starts with an empty Stack and an Input List that contains a sequence of elementary discourse trees edts, one edt for each edu in the tree TS given as input. The status and rhetorical relation associated with each edt is undefined. At each step, the transfer module applies an operation that is aimed at building from the units in TS the discourse tree Tt. In the context of the discourse-transfer module, 7 types of operations are implemented:
~ SHIFT operations transfer the first edt from the input list into the stack;
~ REDUCE operations pop the two discourse trees located at the top of the stack; combine them into a new tee updating the statuses and rhetorical relation names of the trees involved in the operation; and push the new tree on the top of the stack. These operations are used to build the structure of the discourse tree in the target language, ~ BREAK operations are used in order to break the edt at the beginning of the input list into a predetermined number of units. These operations are used to ensure that the resulting tree has the same number of edts as Tt. For example, a BREAK operation is used whenever a Japanese edu is mapped into multiple English units.
~ CREATE-NEXT operations are used, for example, in order to create English (target language) discourse constituents that have no correspondent in the Japanese (source language) tree.
~ FUSE operations are used in order to fuse the edt at the top of the stack into the tree that immediately precedes it. These operations are used, for example, whenever multiple Japanese edus are mapped into one English edu.
~ SWAP operations swap the edt at the beginning of the input list with an edt found one or more positions to the right. These operations are used for re-ordering discourse constituents.
~ ASSIGNTYPE operations assign one or more of the following types to the tree t the top of the stack:
Unit, Multiunit, Sentence, Paragraph, MultiParagraph, and Text. These operations are used in order to ensure sentence and paragraph boundaries that are specific to the target language.
For example, the first sentence of the English tree in Figure 9 can be obtained from the original Japanese sequence by following the sequence of actions (5), below, whose effects are shown in Figure 14. For the purpose of compactness, Figure 14 does not illustrate the effect of ASSIGNTYPE actions. For the same purpose, some lines correspond to more than one action.
(5) BREAK 2; SWAP 2; SHIFT; ASSIGNTYPE
UNIT; SHIFT; REDUCE-NS-ELABORATION-OBJECT-ATTRIBUTE-E; ASSIGNTYPE
MULTIUNIT; SHIFT; ASSIGNTYPE UNIT;
SHIFT; ASSIGNTYPE UNIT; FUSE;
ASSIGNTYPE UNIT; SWAP2; SHIFT;
ASSIGNTYPE UNIT; FUSE; BREAK2; SHIFT;
ASSIGNTYPE UNIT; SHIFT; ASSIGNTYPE
UNIT; REDUCE-NS-ELABORATION-ADDITIONAL;
ASSIGNTYPE MULTIUNIT; REDUCE-NS-CONTRAST; ASSIGNTYPE MULTIUNIT; REDUCE-SN-BACKGROUND; ASSIGNTYPE SENTENCE.
For the corpus used, in order to enable a discourse-based transfer module to derive any English discourse tree starting from any Japanese discourse tree, it is sufficient to implement:
~ one SHIFT operation;
~ 3 x 2 x 85 REDUCE operations; (For each of the three possible pairs of nuclearity assignments NUCLEUS-SATELLITE (NS), SATELLITE-NUCLEUS (SN), AND NUCLEUS-NUCLEUS (NN), there are two possible ways to reduce two adjacent trees (one results in a binary tree, the other in a non-binary tree (Marcu, "A decision-based approach to rhetorical parsing," In Proc. Of ACL'99, pages 365-372, Maryland (1999)), and 85 relation names.);
~ three types of BREAK operations; (In the corpus used, a Japanese unit is broken into two, three, or at most four units.);
~ one type of CREATE-NEXT operation;
~ one type of FUSE operation;
~ eleven types of SWAP operations; (In the corpus, Japanese units are at most 11 positions away from their location in an English-specific rendering.) ~ seven types of ASSIGNTYPE operations: Unit, Multiunit, Sentence, MultiSentence, paragraph, MultiParagraph, and Text.
These actions are sufficient for rewriting any tree TS
into any tree Tt, where Tt may have a different number of edus, where the edus of Tt may have a different ordering than the edus of TS, and where the hierarchical structures of the two trees may be different as well.
Learning the parameters of the discourse-transfer model Each configuration of the transfer model is associated with a learning case. The cases were generated by a program that automatically derived the sequence of actions that mapped the Japanese trees in the corpus into the sibling English trees, using the correspondences at the elementary unit level that were constructed manually.
Overall, the 40 pairs of Japanese and English discourse trees yielded 14108 cases, To each learning example, a set of features from the following classes was associated:
Operational and discourse features reflect the number of trees in the stack, the input list, and the types of the last five operations. They encode information pertaining to the types of the partial trees built up to a certain time and the rhetorical relations that hold between these trees.
Correspondence-based features reflect the nuclearity, rhetorical relations, and types of the Japanese trees that correspond to the English-like partial trees derived up to a given time.
Lexical features specify whether the Japanese spans that correspond to the structures derived up to a given time use potential discourse markers, such as dakara (because) and no ni (although) .
The discourse transfer module uses the C4.5 program (Quinlan, C4.5: Programs for Machine Learning, Morgan Ifiaufmann Publishers (1993)) in order to learn decision trees and rules that specify how Japanese discourse trees should be mapped into English-like trees. A ten-fold cross-validation evaluation of the classifier yielded an accuracy of 70.20 (~ 0.21).
In order to better understand the strengths and weaknesses of the classifier, the problem was broken into smaller components. Hence, instead of learning all actions at once, it was determined first whether the rewriting procedure should choose a SHIFT, REDUCE, BREAK, FUSE, SWAP, or ASSIGNTYPE operation (the Domain Action Type" classifier in table 3), and only then to refine this decision by determining what type of reduce operation to perform how many units to break a Japanese units into, how big the distance to the SWAP-ed unit should be, and what type of ASSIGNTYPE operation one should perform. Table 3 below shows the sizes of each data set and the performance of each of these classifiers, as determined using a ten-fold cross-validation procedure. For the purpose of comparison, each classifier is paired with a majority baseline.
Classifier # Cases Accuracy Majority (10-fold baseline cross accuracy validation) General 14108 70.20% (0.21) 22.05%(on ASSIGNTYPE UNIT) (Learns all classes at once) Main Action Type 14108 82.530 (0.25) 45.47%(on ASSIGNTYPE) AssignType 6416 90.46% (0.39) 57.30%(on ASSIGNTYPE Unit) Break 394 82.91% (1.40) 82.91%(on BREAK 2) Nuclearity-Reduce 2388 67.43% (1.03) 50.920(on NS) Relation-Reduce 2388 48.20% (1.01) 17.18%(on ELABORATION-OBJECT-ATTRIBUTE-E) Swap 842 62.98% (1.62) 62.98%(on SWAP1) Table 3: Performance of the classifiers The results in Table 3 show that the most difficult subtasks to learn are that of determining the number of units a Japanese unit should be broken into and that of determining the distance to the unit that is to be swapped.
The features used are not able to refine the baseline classifiers for these action types. The confusion matrix for the "Main Action Type" classifier (see Table 4) shows that the system has trouble mostly identifying BREAK and CREATE-NEXT actions. The system has difficulty learning what type of nuclearity ordering to prefer (the "Nuclearity-Reduce" classifier) and what relation to choose for the English-like structure (the "Relation-Reduce"
classifier).
Action (a) (b) (c) (d) (e) (f) (g) .
ASSIGNTYPE (a) 660 BREAK (b) 1 2 28 l (c) FUSE (d) 69 8 3 REDUCE (e) 4 18 193 30 3 SHIFT (f) 1 4 15 44 243 25 SWAP (g) 3 4 14 43 25 l0 Table 4: Confusion matrix for the Main Action Type classifier Figure 15 shows a typical learning curve, the one that corresponds to the "Reduce Relation" classifier. The learning curves suggest that more training data may improve performance. However, they also suggest that better features may be needed in order to improve performance significantly.
Table 5 below displays some learned rules. The first rule accounts for rhetorical mappings in which the order of the nucleus and satellite of an ATTRIBUTION relation is changed when translated from Japanese into English. The second rule was learned in order to map EXAMPLE Japanese satellites into EVIDENCE English satellites.
if rhetRelOfStack-lInJapTree = ATTRIBUTION
then rhetlOfTopStackInEngTree <-- ATTRIBUTION
if rhetRelOfTopStackInJapTree = EXAMPLE /~
isSentenceTheLastUnitinJapTreeOfTopStack = false then rhetRelOfTopStackInEngTree <- EVIDENCE
Table 5: Rule examples for the Relation-Reduce classifier Evaluation of the discourse-based transfer module By applying the General classifier or the other six classifiers successively, one can map any Japanese discourse tree into a tree whose structure comes closer to the natural rendering of English. To evaluate the discourse-based transfer module, a ten-fold cross-validation experiment was performed. That is, the classifiers were trained on 36 pairs of manually built and aligned discourse structures, and then the learned classifiers were used in order to map 4 unseen Japanese discourse trees into English-like trees. The similarity of the derived trees and the English trees built manually was measured using the metrics discussed above. This procedure was repeated ten times, each time training and testing on different subsets of tree pairs.
The results reported in Table 2 were as a baseline for the model. The baseline corresponds to applying no knowledge of discourse. Table 6 below displays the absolute improvement (in percentage points) in recall and precision figures obtained when the General classifier was used to map Japanese trees into English-looking trees. The General classifier yielded the best results. The results in Table 6 are averaged over a ten-fold cross-validation experiment.
bevel Units Spans StatusNuclearityRelations P-D P-D P P-D P-D P P-D P-D P P-D P-D
R R R R P
Sentence +9.1 +25.5 +2.0 +19.9 +0.4 +13.4 -0.01+8.4 Paragraph -14.7+1.4 -12.5-1.7 -11.0-2.4 -9.9 -3.3 Text -9.6 -13.5 -7.1 -11.1 -6.3 -10.0 -5.2 -8.8 Weighted Average +1.5 +14.1 -2.1 +9.9 -3.1 +6.4 -3.0 +3.9 All -1.2 +2.5 -0.1 +2.9 +0.6 +3.5 +0.7 +2.6 R P R P R P R P
Sentence +13.4 +30.4 +3.1 +36.1 -6.3 +18.6 -10.1 +3.9 Paragraph -15.6 +0.6 -13.5 -0.8 -11.7 -1.8 -10.3 -2.8 Text -15.4 -23.3 -13.0 -20.4 -13.2 -19.5 -11.5 -17.0 Weighted Average +3.6 +15.5 -2.7 +17.1 -8.5 +7.3 -10.5 -0.4 All +12.7 +29.6 +2.0 +28.8 -5.1 +13.0 -7.9 +2.2 Table 6: Relative evaluation of the discourse-based transfer module with respect to the figures in Table 2 The results in Table 6 show that the model described here outperforms the baseline with respect to building English-like discourse structures for sentences, but it under-performs the baseline with respect to building English-like structures at the paragraph and text levels.
One potential shortcoming of this model seems to come from its low performance in assigning paragraph boundaries.
Because the classifier does not learn correctly which spans to consider paragraphs and which spans not, the recall and precision results at the paragraph and text levels are negatively affected. The poorer results at the paragraph and text levels can be also explained by errors whose effect cumulates during the step-by-step tree-reconstruction procedure; and by the fact that, for these levels, there is less data to learn from.
However, if one ignores the sentence and paragraph boundaries and evaluates the discourse structures overall, one can see that this model outperforms the baseline on all accounts according to the Position-Dependent evaluation;
outperforms the baseline with respect to the assignment of elementary units, hierarchical spans, and nuclearity statures according to the Position-Independent evaluation and under-performs the baseline only slightly with respect to the rhetorical relation assignment according to the Position-Independent evaluation. More sophisticated discourse features, such as those discussed by Maynard, Principles of Japanese Discourse: A Handlaook, Cambridge Univ. Press (1998), for example, and a tighter integration with the lexicogrammar of the two languages may yield better cues for learning discourse-based translation models .
Alternative embodiments of the tree rewriter are possible. For example, probabilities could be incorporated into the tree rewriting procedure. Alternatively, or in addition, multiple trees could be rewritten in parallel and the best one selected in the end. In the current embodiment, the target tree Tt is generated in a sequence of deterministic steps with no recursion or branching.
Alternatively, it is possible to associate a probability with each individual step and reconstruct the target tree Tt by exploring multiple alternatives at the same time. The probability of a target tree Tt is given by the product of the probabilities of all steps that led to that tree. In this case, the target tree Tt will be taken to be the resulting tree of maximum probability. An advantage of such an approach is that it enables the learning of probabilistic transfer functions, H[~.
Although only a few embodiments have been described in detail above, those having ordinary skill in the art will certainly understand that many modifications are possible in the preferred embodiment without departing from the teachings thereof. All such modifications are encompassed within the following claims.
Claims (24)
1. A machine translation decoding method comprising:
receiving as input a text segment in a source language to be translated into a target language;
generating an initial translation as a current target language translation using translation information learned in part from text corpora;
applying one or more modification operators to the current target language translation to generate one or more modified target language translations;
determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations;
setting a modified target language translation as the current target language translation; and repeating said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
receiving as input a text segment in a source language to be translated into a target language;
generating an initial translation as a current target language translation using translation information learned in part from text corpora;
applying one or more modification operators to the current target language translation to generate one or more modified target language translations;
determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations;
setting a modified target language translation as the current target language translation; and repeating said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
2. The method of claim 1 wherein the text segment comprises a clause, a sentence, a paragraph or a treatise.
3. The method of claim 1 wherein generating an initial translation comprises generating a gloss.
4. The method of claim 3 wherein the gloss is a word-for-word gloss or a phrase-for-phrase gloss.
5. The method of claim 1 wherein applying one or more modification operators comprises changing in the current target language translation the translation of one or two words.
6. The method of claim 1 wherein applying one or more modification operators comprises (i) changing in the current target language translation a translation of a word and concurrently (ii) inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word being selected from a list of words derived according to their probabilities having a zero-value fertility.
7. The method of claim 1 wherein applying one or more modification operators comprises deleting from the current target language translation a word having a zero-value fertility.
8. The method of claim 1 wherein applying one or more modification operators comprises modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation.
9. The method of claim 1 wherein applying one or more modification operators comprises modifying an alignment between the source language text segment and the current target language translation by (i) eliminating a target language word from the current target language translation and (ii) linking words in the source language text segment.
10. The method of claim 1 wherein applying one or more modification operators comprises applying two or more of the following:
(i) changing in the current target language translation the translation of one or two words;
(ii) changing in the current target language translation a translation of a word and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word having a high probability of having a zero-value fertility;
(iii) deleting from the current target language translation a word having a zero-value fertility;
(iv) modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and (v) modifying an alignment between the source language text segment and the current target language translation by eliminating a target language word from the current target language translation and linking words in the source language text segment.
(i) changing in the current target language translation the translation of one or two words;
(ii) changing in the current target language translation a translation of a word and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word having a high probability of having a zero-value fertility;
(iii) deleting from the current target language translation a word having a zero-value fertility;
(iv) modifying an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and (v) modifying an alignment between the source language text segment and the current target language translation by eliminating a target language word from the current target language translation and linking words in the source language text segment.
11. A computer-implemented machine translation decoding method comprising iteratively modifying a target language translation of a source language text segment until an occurrence of a termination condition; wherein iteratively modifying the translation comprises performing at each iteration modification operations on the translation to form modified translations, the modification operations including changing both word translation and word order, and performing at each iteration an evaluation of the modified translations taking into account both source-to-target-language translation information and target-language structure information learned both in part from text corpora.
12. The method of claim 11 wherein the termination condition comprises a determination that a probability of correctness of a modified translation is no greater than a probability of correctness of a previous translation.
13. The method of claim 11 wherein the termination condition comprises a completion of a predetermined number of iterations.
14. The method of claim 11 wherein the source language text segment comprises a clause, a sentence, a paragraph or a treatise.
15. The method of claim 11 wherein the method starts with an approximate target language translation and iteratively improves the translation with each successive iteration.
16. The method of claim 15 wherein the approximate target language translation comprises a gloss.
17. The method of claim 16 wherein the gloss comprises a word-for-word gloss or a phrase-for-phrase gloss.
18. The method of claim 15 wherein the approximate target language translation comprises a predetermined translation selected from among a plurality of predetermined translations.
19. The method of claim 11 wherein said iteratively modifying comprises iteratively changing output such that at every iterative step the change of the output corresponds to a best of available options for the translation in that iterative step.
20. The method of claim 11 wherein iteratively modifying the translation comprises incrementally improving the translation with each iteration.
21. The method of claim 11 wherein the modification operations comprises one or more of the following operations:
(i) changing one or two words in the translation;
(ii) changing a translation of a word and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the translation, the inserted other word having a high probability of having a zero-value fertility;
(iii) deleting from the translation a word having a zero-value fertility;
(iv) modifying an alignment between the source language text segment and the translation by swapping non-overlapping target language word segments in the translation; and (v) modifying an alignment between the source language text segment and the translation by eliminating a target language word from the translation and linking words in the source language text segment.
(i) changing one or two words in the translation;
(ii) changing a translation of a word and concurrently inserting another word at a position that yields an alignment of highest probability between the source language text segment and the translation, the inserted other word having a high probability of having a zero-value fertility;
(iii) deleting from the translation a word having a zero-value fertility;
(iv) modifying an alignment between the source language text segment and the translation by swapping non-overlapping target language word segments in the translation; and (v) modifying an alignment between the source language text segment and the translation by eliminating a target language word from the translation and linking words in the source language text segment.
22. A machine translation decoder comprising:
a decoding engine comprising one or more modification operators to be applied to a current target language translation to generate one or more modified target language translations; and a process loop to iteratively modify the current target language translation using the one or more modification operators, the process loop being configured to effect operations comprising:
applying the one or more modification operators to the current target language translation using the decoding engine;
determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations;
setting a modified target language translation as the current target language translation; and repeating said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
a decoding engine comprising one or more modification operators to be applied to a current target language translation to generate one or more modified target language translations; and a process loop to iteratively modify the current target language translation using the one or more modification operators, the process loop being configured to effect operations comprising:
applying the one or more modification operators to the current target language translation using the decoding engine;
determining, according to a statistical model of machine translation, whether one or more of the modified target language translations represents an improved translation in comparison with the current target language translation, said determining comprising calculating a probability of correctness for each of the modified target language translations;
setting a modified target language translation as the current target language translation; and repeating said applying, said determining and said setting until occurrence of a termination condition comprising a completion of a predetermined number of iterations, a lapse of a predetermined amount of time, or a determination that a probability of correctness of the current target language translation is greater than each of one or more probabilities of correctness of the one or more modified target language translations.
23. The decoder of claim 22, further comprising a language model and a translation module employed in said determining.
24. The decoder of claim 22 wherein the one or more modification operators comprise one or more of the following:
(i) an operator to change in the current target language translation the translation of one or two words;
(ii) an operator to change in the current target language translation a translation of a word and to concurrently insert another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word having a high probability of having a zero-value fertility;
(iii) an operator to delete from the current target language translation a word having a zero-value fertility;
(iv) an operator to modify an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and (v) an operator to modify an alignment between the source language text segment and the current target language translation by eliminating a target language word from the current target language translation and linking words in the source language text segment.
(i) an operator to change in the current target language translation the translation of one or two words;
(ii) an operator to change in the current target language translation a translation of a word and to concurrently insert another word at a position that yields an alignment of highest probability between the source language text segment and the current target language translation, the inserted other word having a high probability of having a zero-value fertility;
(iii) an operator to delete from the current target language translation a word having a zero-value fertility;
(iv) an operator to modify an alignment between the source language text segment and the current target language translation by swapping non-overlapping target language word segments in the current target language translation; and (v) an operator to modify an alignment between the source language text segment and the current target language translation by eliminating a target language word from the current target language translation and linking words in the source language text segment.
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US20364300P | 2000-05-11 | 2000-05-11 | |
US60/203,643 | 2000-05-11 | ||
PCT/US2001/015379 WO2001086491A2 (en) | 2000-05-11 | 2001-05-11 | Machine translation techniques |
Publications (2)
Publication Number | Publication Date |
---|---|
CA2408819A1 CA2408819A1 (en) | 2001-11-15 |
CA2408819C true CA2408819C (en) | 2006-11-07 |
Family
ID=22754752
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CA002408819A Expired - Lifetime CA2408819C (en) | 2000-05-11 | 2001-05-11 | Machine translation techniques |
Country Status (7)
Country | Link |
---|---|
US (2) | US20020046018A1 (en) |
EP (1) | EP1352338A2 (en) |
JP (1) | JP2004501429A (en) |
CN (1) | CN1465018A (en) |
AU (2) | AU2001261505A1 (en) |
CA (1) | CA2408819C (en) |
WO (2) | WO2001086489A2 (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8825466B1 (en) | 2007-06-08 | 2014-09-02 | Language Weaver, Inc. | Modification of annotated bilingual segment pairs in syntax-based machine translation |
US8942973B2 (en) | 2012-03-09 | 2015-01-27 | Language Weaver, Inc. | Content page URL translation |
US8990064B2 (en) | 2009-07-28 | 2015-03-24 | Language Weaver, Inc. | Translating documents based on content |
US9122674B1 (en) | 2006-12-15 | 2015-09-01 | Language Weaver, Inc. | Use of annotations in statistical machine translation |
US9152622B2 (en) | 2012-11-26 | 2015-10-06 | Language Weaver, Inc. | Personalized machine translation via online adaptation |
US9213694B2 (en) | 2013-10-10 | 2015-12-15 | Language Weaver, Inc. | Efficient online domain adaptation |
US10261994B2 (en) | 2012-05-25 | 2019-04-16 | Sdl Inc. | Method and system for automatic management of reputation of translators |
US10319252B2 (en) | 2005-11-09 | 2019-06-11 | Sdl Inc. | Language capability assessment and training apparatus and techniques |
Families Citing this family (292)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060116865A1 (en) | 1999-09-17 | 2006-06-01 | Www.Uniscape.Com | E-services translation utilizing machine translation and translation memory |
US6961692B1 (en) * | 2000-08-01 | 2005-11-01 | Fuji Xerox Co, Ltd. | System and method for writing analysis using the linguistic discourse model |
WO2002027542A1 (en) * | 2000-09-28 | 2002-04-04 | Intel Corporation (A Corporation Of Delaware) | A method and apparatus for extracting entity names and their relations |
US7054803B2 (en) * | 2000-12-19 | 2006-05-30 | Xerox Corporation | Extracting sentence translations from translated documents |
US6990439B2 (en) * | 2001-01-10 | 2006-01-24 | Microsoft Corporation | Method and apparatus for performing machine translation using a unified language model and translation model |
US7904595B2 (en) | 2001-01-18 | 2011-03-08 | Sdl International America Incorporated | Globalization management system and method therefor |
US7069207B2 (en) * | 2001-01-26 | 2006-06-27 | Microsoft Corporation | Linguistically intelligent text compression |
JP3916124B2 (en) * | 2001-02-15 | 2007-05-16 | インターナショナル・ビジネス・マシーンズ・コーポレーション | Digital document browsing system, browser, digital document display method, program, and storage medium |
US7734459B2 (en) * | 2001-06-01 | 2010-06-08 | Microsoft Corporation | Automatic extraction of transfer mappings from bilingual corpora |
US8214196B2 (en) | 2001-07-03 | 2012-07-03 | University Of Southern California | Syntax-based statistical translation model |
US7146358B1 (en) * | 2001-08-28 | 2006-12-05 | Google Inc. | Systems and methods for using anchor text as parallel corpora for cross-language information retrieval |
US7447623B2 (en) * | 2001-10-29 | 2008-11-04 | British Telecommunications Public Limited Company | Machine translation |
US7491690B2 (en) | 2001-11-14 | 2009-02-17 | Northwestern University | Self-assembly and mineralization of peptide-amphiphile nanofibers |
US7295966B2 (en) * | 2002-01-14 | 2007-11-13 | Microsoft Corporation | System for normalizing a discourse representation structure and normalized data structure |
US7127208B2 (en) * | 2002-01-23 | 2006-10-24 | Educational Testing Service | Automated annotation |
US7371719B2 (en) * | 2002-02-15 | 2008-05-13 | Northwestern University | Self-assembly of peptide-amphiphile nanofibers under physiological conditions |
US20040076930A1 (en) * | 2002-02-22 | 2004-04-22 | Steinberg Linda S. | Partal assessment design system for educational testing |
US7340466B2 (en) * | 2002-02-26 | 2008-03-04 | Kang Jo Mgmt. Limited Liability Company | Topic identification and use thereof in information retrieval systems |
US7716207B2 (en) * | 2002-02-26 | 2010-05-11 | Odom Paul S | Search engine methods and systems for displaying relevant topics |
US20060004732A1 (en) * | 2002-02-26 | 2006-01-05 | Odom Paul S | Search engine methods and systems for generating relevant search results and advertisements |
JP3959453B2 (en) * | 2002-03-14 | 2007-08-15 | 沖電気工業株式会社 | Translation mediation system and translation mediation server |
WO2004001623A2 (en) * | 2002-03-26 | 2003-12-31 | University Of Southern California | Constructing a translation lexicon from comparable, non-parallel corpora |
JP4088131B2 (en) * | 2002-03-28 | 2008-05-21 | 富士通株式会社 | Synchronous content information generation program, synchronous content information generation device, and synchronous content information generation method |
US7634398B2 (en) * | 2002-05-16 | 2009-12-15 | Microsoft Corporation | Method and apparatus for reattaching nodes in a parse structure |
US7805302B2 (en) * | 2002-05-20 | 2010-09-28 | Microsoft Corporation | Applying a structured language model to information extraction |
US7534761B1 (en) | 2002-08-21 | 2009-05-19 | North Western University | Charged peptide-amphiphile solutions and self-assembled peptide nanofiber networks formed therefrom |
US7305336B2 (en) * | 2002-08-30 | 2007-12-04 | Fuji Xerox Co., Ltd. | System and method for summarization combining natural language generation with structural analysis |
DE10337934A1 (en) * | 2002-09-30 | 2004-04-08 | Siemens Ag | Unstructured text conversion method in which the text is structured using structuring rules that operate on text fragments and sort them using terminology and subject dependent structuring rules |
US7554021B2 (en) * | 2002-11-12 | 2009-06-30 | Northwestern University | Composition and method for self-assembly and mineralization of peptide amphiphiles |
WO2004046167A2 (en) | 2002-11-14 | 2004-06-03 | Northwestern University | Synthesis and self-assembly of abc triblock bola peptide |
US20040194036A1 (en) * | 2002-11-14 | 2004-09-30 | Magdalena Wolska | Automated evaluation of overly repetitive word use in an essay |
US20050108256A1 (en) * | 2002-12-06 | 2005-05-19 | Attensity Corporation | Visualization of integrated structured and unstructured data |
WO2004072104A2 (en) * | 2003-02-11 | 2004-08-26 | Northwestern University | Methods and materials for nanocrystalline surface coatings and attachment of peptide amphiphile nanofibers thereon |
US20040230415A1 (en) * | 2003-05-12 | 2004-11-18 | Stefan Riezler | Systems and methods for grammatical text condensation |
US8548794B2 (en) | 2003-07-02 | 2013-10-01 | University Of Southern California | Statistical noun phrase translation |
US7711545B2 (en) * | 2003-07-02 | 2010-05-04 | Language Weaver, Inc. | Empirical methods for splitting compound words with application to machine translation |
JP2005100335A (en) * | 2003-09-01 | 2005-04-14 | Advanced Telecommunication Research Institute International | Machine translation apparatus, machine translation computer program, and computer |
JP3919771B2 (en) * | 2003-09-09 | 2007-05-30 | 株式会社国際電気通信基礎技術研究所 | Machine translation system, control device thereof, and computer program |
US7610190B2 (en) * | 2003-10-15 | 2009-10-27 | Fuji Xerox Co., Ltd. | Systems and methods for hybrid text summarization |
JP4741510B2 (en) * | 2003-12-05 | 2011-08-03 | ノースウエスタン ユニバーシティ | Branched peptide amphiphiles, related epitope compounds and their self-assembled structures |
KR20070004560A (en) * | 2003-12-05 | 2007-01-09 | 노오쓰웨스턴 유니버시티 | Self-assembling peptide amphiphiles and related methods for growth factor delivery |
US20050138556A1 (en) * | 2003-12-18 | 2005-06-23 | Xerox Corporation | Creation of normalized summaries using common domain models for input text analysis and output text generation |
US7657420B2 (en) * | 2003-12-19 | 2010-02-02 | Palo Alto Research Center Incorporated | Systems and methods for the generation of alternate phrases from packed meaning |
US8037102B2 (en) | 2004-02-09 | 2011-10-11 | Robert T. and Virginia T. Jenkins | Manipulating sets of hierarchical data |
US20050187772A1 (en) * | 2004-02-25 | 2005-08-25 | Fuji Xerox Co., Ltd. | Systems and methods for synthesizing speech using discourse function level prosodic features |
US7983896B2 (en) | 2004-03-05 | 2011-07-19 | SDL Language Technology | In-context exact (ICE) matching |
US8296127B2 (en) | 2004-03-23 | 2012-10-23 | University Of Southern California | Discovery of parallel text portions in comparable collections of corpora and training using comparable texts |
WO2005098788A2 (en) * | 2004-04-02 | 2005-10-20 | Sri International | System and method for assessment design |
US8666725B2 (en) | 2004-04-16 | 2014-03-04 | University Of Southern California | Selection and use of nonstatistical translation components in a statistical machine translation framework |
US20050256848A1 (en) * | 2004-05-13 | 2005-11-17 | International Business Machines Corporation | System and method for user rank search |
US9646107B2 (en) | 2004-05-28 | 2017-05-09 | Robert T. and Virginia T. Jenkins as Trustee of the Jenkins Family Trust | Method and/or system for simplifying tree expressions such as for query reduction |
US7562008B2 (en) * | 2004-06-23 | 2009-07-14 | Ning-Ping Chan | Machine translation method and system that decomposes complex sentences into two or more sentences |
US7620632B2 (en) * | 2004-06-30 | 2009-11-17 | Skyler Technology, Inc. | Method and/or system for performing tree matching |
US7882147B2 (en) * | 2004-06-30 | 2011-02-01 | Robert T. and Virginia T. Jenkins | File location naming hierarchy |
US8768969B2 (en) * | 2004-07-09 | 2014-07-01 | Nuance Communications, Inc. | Method and system for efficient representation, manipulation, communication, and search of hierarchical composite named entities |
US7599914B2 (en) * | 2004-07-26 | 2009-10-06 | Google Inc. | Phrase-based searching in an information retrieval system |
US7567959B2 (en) | 2004-07-26 | 2009-07-28 | Google Inc. | Multiple index based information retrieval system |
US7702618B1 (en) | 2004-07-26 | 2010-04-20 | Google Inc. | Information retrieval system for archiving multiple document versions |
US7426507B1 (en) | 2004-07-26 | 2008-09-16 | Google, Inc. | Automatic taxonomy generation in search results using phrases |
US7584175B2 (en) * | 2004-07-26 | 2009-09-01 | Google Inc. | Phrase-based generation of document descriptions |
US7580921B2 (en) * | 2004-07-26 | 2009-08-25 | Google Inc. | Phrase identification in an information retrieval system |
US7711679B2 (en) | 2004-07-26 | 2010-05-04 | Google Inc. | Phrase-based detection of duplicate documents in an information retrieval system |
US7580929B2 (en) * | 2004-07-26 | 2009-08-25 | Google Inc. | Phrase-based personalization of searches in an information retrieval system |
US7536408B2 (en) * | 2004-07-26 | 2009-05-19 | Google Inc. | Phrase-based indexing in an information retrieval system |
JP2006039120A (en) * | 2004-07-26 | 2006-02-09 | Sony Corp | Interactive device and interactive method, program and recording medium |
JP5452868B2 (en) * | 2004-10-12 | 2014-03-26 | ユニヴァーシティー オブ サザン カリフォルニア | Training for text-to-text applications that use string-to-tree conversion for training and decoding |
US7801923B2 (en) | 2004-10-29 | 2010-09-21 | Robert T. and Virginia T. Jenkins as Trustees of the Jenkins Family Trust | Method and/or system for tagging trees |
US7627591B2 (en) * | 2004-10-29 | 2009-12-01 | Skyler Technology, Inc. | Method and/or system for manipulating tree expressions |
US7970600B2 (en) * | 2004-11-03 | 2011-06-28 | Microsoft Corporation | Using a first natural language parser to train a second parser |
US7801723B2 (en) * | 2004-11-30 | 2010-09-21 | Palo Alto Research Center Incorporated | Systems and methods for user-interest sensitive condensation |
US7630995B2 (en) * | 2004-11-30 | 2009-12-08 | Skyler Technology, Inc. | Method and/or system for transmitting and/or receiving data |
US7636727B2 (en) | 2004-12-06 | 2009-12-22 | Skyler Technology, Inc. | Enumeration of trees from finite number of nodes |
US7827029B2 (en) * | 2004-11-30 | 2010-11-02 | Palo Alto Research Center Incorporated | Systems and methods for user-interest sensitive note-taking |
US20060155530A1 (en) * | 2004-12-14 | 2006-07-13 | International Business Machines Corporation | Method and apparatus for generation of text documents |
US7401077B2 (en) * | 2004-12-21 | 2008-07-15 | Palo Alto Research Center Incorporated | Systems and methods for using and constructing user-interest sensitive indicators of search results |
US8316059B1 (en) | 2004-12-30 | 2012-11-20 | Robert T. and Virginia T. Jenkins | Enumeration of rooted partial subtrees |
JP4301515B2 (en) * | 2005-01-04 | 2009-07-22 | インターナショナル・ビジネス・マシーンズ・コーポレーション | Text display method, information processing apparatus, information processing system, and program |
WO2006079036A2 (en) * | 2005-01-21 | 2006-07-27 | Northwestern University | Methods and compositions for encapsulation of cells |
US8615530B1 (en) | 2005-01-31 | 2013-12-24 | Robert T. and Virginia T. Jenkins as Trustees for the Jenkins Family Trust | Method and/or system for tree transformation |
US7681177B2 (en) | 2005-02-28 | 2010-03-16 | Skyler Technology, Inc. | Method and/or system for transforming between trees and strings |
WO2006093928A2 (en) * | 2005-02-28 | 2006-09-08 | Educational Testing Service | Method of model scaling for an automated essay scoring system |
US7851445B2 (en) * | 2005-03-04 | 2010-12-14 | Northwestern University | Angiogenic heparin-binding epitopes, peptide amphiphiles, self-assembled compositions and related methods of use |
JP4050755B2 (en) * | 2005-03-30 | 2008-02-20 | 株式会社東芝 | Communication support device, communication support method, and communication support program |
US8356040B2 (en) | 2005-03-31 | 2013-01-15 | Robert T. and Virginia T. Jenkins | Method and/or system for transforming between trees and arrays |
US7899821B1 (en) | 2005-04-29 | 2011-03-01 | Karl Schiffmann | Manipulation and/or analysis of hierarchical data |
US20060277028A1 (en) * | 2005-06-01 | 2006-12-07 | Microsoft Corporation | Training a statistical parser on noisy data by filtering |
US8886517B2 (en) | 2005-06-17 | 2014-11-11 | Language Weaver, Inc. | Trust scoring for language translation systems |
WO2006133571A1 (en) * | 2005-06-17 | 2006-12-21 | National Research Council Of Canada | Means and method for adapted language translation |
US8676563B2 (en) | 2009-10-01 | 2014-03-18 | Language Weaver, Inc. | Providing human-generated and machine-generated trusted translations |
US20070010989A1 (en) * | 2005-07-07 | 2007-01-11 | International Business Machines Corporation | Decoding procedure for statistical machine translation |
US7779396B2 (en) * | 2005-08-10 | 2010-08-17 | Microsoft Corporation | Syntactic program language translation |
US8924212B1 (en) * | 2005-08-26 | 2014-12-30 | At&T Intellectual Property Ii, L.P. | System and method for robust access and entry to large structured data using voice form-filling |
US20070192309A1 (en) * | 2005-10-12 | 2007-08-16 | Gordon Fischer | Method and system for identifying sentence boundaries |
US9165039B2 (en) * | 2005-11-29 | 2015-10-20 | Kang Jo Mgmt, Limited Liability Company | Methods and systems for providing personalized contextual search results |
EP2511833B1 (en) | 2006-02-17 | 2020-02-05 | Google LLC | Encoding and adaptive, scalable accessing of distributed translation models |
CN101401488B (en) * | 2006-03-06 | 2012-07-04 | 皇家飞利浦电子股份有限公司 | Use of decision trees for automatic commissioning |
US8943080B2 (en) | 2006-04-07 | 2015-01-27 | University Of Southern California | Systems and methods for identifying parallel documents and sentence fragments in multilingual document collections |
US7827028B2 (en) | 2006-04-07 | 2010-11-02 | Basis Technology Corporation | Method and system of machine translation |
US7552047B2 (en) * | 2006-05-02 | 2009-06-23 | International Business Machines Corporation | Instance-based sentence boundary determination by optimization |
US8886518B1 (en) | 2006-08-07 | 2014-11-11 | Language Weaver, Inc. | System and method for capitalizing machine translated text |
US8548795B2 (en) * | 2006-10-10 | 2013-10-01 | Abbyy Software Ltd. | Method for translating documents from one language into another using a database of translations, a terminology dictionary, a translation dictionary, and a machine translation system |
US9984071B2 (en) | 2006-10-10 | 2018-05-29 | Abbyy Production Llc | Language ambiguity detection of text |
US20080086298A1 (en) * | 2006-10-10 | 2008-04-10 | Anisimovich Konstantin | Method and system for translating sentences between langauges |
US9645993B2 (en) | 2006-10-10 | 2017-05-09 | Abbyy Infopoisk Llc | Method and system for semantic searching |
US8195447B2 (en) * | 2006-10-10 | 2012-06-05 | Abbyy Software Ltd. | Translating sentences between languages using language-independent semantic structures and ratings of syntactic constructions |
US8214199B2 (en) * | 2006-10-10 | 2012-07-03 | Abbyy Software, Ltd. | Systems for translating sentences between languages using language-independent semantic structures and ratings of syntactic constructions |
US9047275B2 (en) | 2006-10-10 | 2015-06-02 | Abbyy Infopoisk Llc | Methods and systems for alignment of parallel text corpora |
US9633005B2 (en) | 2006-10-10 | 2017-04-25 | Abbyy Infopoisk Llc | Exhaustive automatic processing of textual information |
US8145473B2 (en) | 2006-10-10 | 2012-03-27 | Abbyy Software Ltd. | Deep model statistics method for machine translation |
US9235573B2 (en) | 2006-10-10 | 2016-01-12 | Abbyy Infopoisk Llc | Universal difference measure |
JP5082374B2 (en) * | 2006-10-19 | 2012-11-28 | 富士通株式会社 | Phrase alignment program, translation program, phrase alignment device, and phrase alignment method |
US8433556B2 (en) | 2006-11-02 | 2013-04-30 | University Of Southern California | Semi-supervised training for statistical word alignment |
JP4997966B2 (en) * | 2006-12-28 | 2012-08-15 | 富士通株式会社 | Parallel translation example sentence search program, parallel translation example sentence search device, and parallel translation example sentence search method |
US8468149B1 (en) | 2007-01-26 | 2013-06-18 | Language Weaver, Inc. | Multi-lingual online community |
US7895030B2 (en) * | 2007-03-16 | 2011-02-22 | International Business Machines Corporation | Visualization method for machine translation |
US8615389B1 (en) | 2007-03-16 | 2013-12-24 | Language Weaver, Inc. | Generation and exploitation of an approximate language model |
US8959011B2 (en) | 2007-03-22 | 2015-02-17 | Abbyy Infopoisk Llc | Indicating and correcting errors in machine translation systems |
US8166045B1 (en) | 2007-03-30 | 2012-04-24 | Google Inc. | Phrase extraction using subphrase scoring |
US8166021B1 (en) | 2007-03-30 | 2012-04-24 | Google Inc. | Query phrasification |
US8086594B1 (en) | 2007-03-30 | 2011-12-27 | Google Inc. | Bifurcated document relevance scoring |
US7702614B1 (en) | 2007-03-30 | 2010-04-20 | Google Inc. | Index updating using segment swapping |
US7925655B1 (en) | 2007-03-30 | 2011-04-12 | Google Inc. | Query scheduling using hierarchical tiers of index servers |
US7693813B1 (en) | 2007-03-30 | 2010-04-06 | Google Inc. | Index server architecture using tiered and sharded phrase posting lists |
US8831928B2 (en) | 2007-04-04 | 2014-09-09 | Language Weaver, Inc. | Customizable machine translation service |
US7908552B2 (en) * | 2007-04-13 | 2011-03-15 | A-Life Medical Inc. | Mere-parsing with boundary and semantic driven scoping |
US8076295B2 (en) * | 2007-04-17 | 2011-12-13 | Nanotope, Inc. | Peptide amphiphiles having improved solubility and methods of using same |
US7925496B1 (en) * | 2007-04-23 | 2011-04-12 | The United States Of America As Represented By The Secretary Of The Navy | Method for summarizing natural language text |
US7877251B2 (en) * | 2007-05-07 | 2011-01-25 | Microsoft Corporation | Document translation system |
US9779079B2 (en) * | 2007-06-01 | 2017-10-03 | Xerox Corporation | Authoring system |
US8452585B2 (en) * | 2007-06-21 | 2013-05-28 | Microsoft Corporation | Discriminative syntactic word order model for machine translation |
US8812296B2 (en) | 2007-06-27 | 2014-08-19 | Abbyy Infopoisk Llc | Method and system for natural language dictionary generation |
JP5256654B2 (en) * | 2007-06-29 | 2013-08-07 | 富士通株式会社 | Sentence division program, sentence division apparatus, and sentence division method |
US8103498B2 (en) * | 2007-08-10 | 2012-01-24 | Microsoft Corporation | Progressive display rendering of processed text |
US8117223B2 (en) | 2007-09-07 | 2012-02-14 | Google Inc. | Integrating external related phrase information into a phrase-based indexing information retrieval system |
US8229728B2 (en) * | 2008-01-04 | 2012-07-24 | Fluential, Llc | Methods for using manual phrase alignment data to generate translation models for statistical machine translation |
US20120284015A1 (en) * | 2008-01-28 | 2012-11-08 | William Drewes | Method for Increasing the Accuracy of Subject-Specific Statistical Machine Translation (SMT) |
US8666729B1 (en) * | 2010-02-10 | 2014-03-04 | West Corporation | Processing natural language grammar |
US8738360B2 (en) * | 2008-06-06 | 2014-05-27 | Apple Inc. | Data detection of a character sequence having multiple possible data types |
US9411800B2 (en) * | 2008-06-27 | 2016-08-09 | Microsoft Technology Licensing, Llc | Adaptive generation of out-of-dictionary personalized long words |
US9262409B2 (en) | 2008-08-06 | 2016-02-16 | Abbyy Infopoisk Llc | Translation of a selected text fragment of a screen |
US9129601B2 (en) * | 2008-11-26 | 2015-09-08 | At&T Intellectual Property I, L.P. | System and method for dialog modeling |
US8244519B2 (en) * | 2008-12-03 | 2012-08-14 | Xerox Corporation | Dynamic translation memory using statistical machine translation |
US20100169359A1 (en) * | 2008-12-30 | 2010-07-01 | Barrett Leslie A | System, Method, and Apparatus for Information Extraction of Textual Documents |
WO2010120830A1 (en) * | 2009-04-13 | 2010-10-21 | Northwestern University | Novel peptide-based scaffolds for cartilage regeneration and methods for their use |
CN101996166B (en) * | 2009-08-14 | 2015-08-05 | 张龙哺 | Bilingual sentence is to medelling recording method and interpretation method and translation system |
US8380486B2 (en) | 2009-10-01 | 2013-02-19 | Language Weaver, Inc. | Providing machine-generated translations and corresponding trust levels |
US10417646B2 (en) | 2010-03-09 | 2019-09-17 | Sdl Inc. | Predicting the cost associated with translating textual content |
US8788260B2 (en) * | 2010-05-11 | 2014-07-22 | Microsoft Corporation | Generating snippets based on content features |
US8594998B2 (en) * | 2010-07-30 | 2013-11-26 | Ben-Gurion University Of The Negev Research And Development Authority | Multilingual sentence extractor |
JP2012063868A (en) * | 2010-09-14 | 2012-03-29 | Internatl Business Mach Corp <Ibm> | Method to generate combined parser by combining language processing parsers, and its computer and computer program |
US20120109945A1 (en) * | 2010-10-29 | 2012-05-03 | Emilia Maria Lapko | Method and system of improving navigation within a set of electronic documents |
US9710429B1 (en) * | 2010-11-12 | 2017-07-18 | Google Inc. | Providing text resources updated with translation input from multiple users |
US9317595B2 (en) * | 2010-12-06 | 2016-04-19 | Yahoo! Inc. | Fast title/summary extraction from long descriptions |
US20120143593A1 (en) * | 2010-12-07 | 2012-06-07 | Microsoft Corporation | Fuzzy matching and scoring based on direct alignment |
US9720899B1 (en) | 2011-01-07 | 2017-08-01 | Narrative Science, Inc. | Automatic generation of narratives from data using communication goals and narrative analytics |
US10185477B1 (en) | 2013-03-15 | 2019-01-22 | Narrative Science Inc. | Method and system for configuring automatic generation of narratives from data |
US10657540B2 (en) | 2011-01-29 | 2020-05-19 | Sdl Netherlands B.V. | Systems, methods, and media for web content management |
US9547626B2 (en) | 2011-01-29 | 2017-01-17 | Sdl Plc | Systems, methods, and media for managing ambient adaptability of web applications and web services |
US10580015B2 (en) | 2011-02-25 | 2020-03-03 | Sdl Netherlands B.V. | Systems, methods, and media for executing and optimizing online marketing initiatives |
US10140320B2 (en) | 2011-02-28 | 2018-11-27 | Sdl Inc. | Systems, methods, and media for generating analytical data |
US11003838B2 (en) | 2011-04-18 | 2021-05-11 | Sdl Inc. | Systems and methods for monitoring post translation editing |
US8694303B2 (en) | 2011-06-15 | 2014-04-08 | Language Weaver, Inc. | Systems and methods for tuning parameters in statistical machine translation |
US9984054B2 (en) | 2011-08-24 | 2018-05-29 | Sdl Inc. | Web interface including the review and manipulation of a web document and utilizing permission based control |
US8914279B1 (en) * | 2011-09-23 | 2014-12-16 | Google Inc. | Efficient parsing with structured prediction cascades |
US8886515B2 (en) | 2011-10-19 | 2014-11-11 | Language Weaver, Inc. | Systems and methods for enhancing machine translation post edit review processes |
KR101475284B1 (en) * | 2011-11-29 | 2014-12-23 | 에스케이텔레콤 주식회사 | Error detection apparatus and method based on shallow parser for estimating writing automatically |
US8903707B2 (en) | 2012-01-12 | 2014-12-02 | International Business Machines Corporation | Predicting pronouns of dropped pronoun style languages for natural language translation |
US20150161109A1 (en) * | 2012-01-13 | 2015-06-11 | Google Inc. | Reordering words for machine translation |
CN102662935A (en) * | 2012-04-08 | 2012-09-12 | 北京语智云帆科技有限公司 | Interactive machine translation method and machine translation system |
US8989485B2 (en) | 2012-04-27 | 2015-03-24 | Abbyy Development Llc | Detecting a junction in a text line of CJK characters |
US8971630B2 (en) | 2012-04-27 | 2015-03-03 | Abbyy Development Llc | Fast CJK character recognition |
US9773270B2 (en) | 2012-05-11 | 2017-09-26 | Fredhopper B.V. | Method and system for recommending products based on a ranking cocktail |
WO2014021074A1 (en) * | 2012-07-31 | 2014-02-06 | 日本電気株式会社 | Problem situation detection device, problem situation detection method and problem situation detection-use program |
US11308528B2 (en) | 2012-09-14 | 2022-04-19 | Sdl Netherlands B.V. | Blueprinting of multimedia assets |
US11386186B2 (en) | 2012-09-14 | 2022-07-12 | Sdl Netherlands B.V. | External content library connector systems and methods |
US10452740B2 (en) | 2012-09-14 | 2019-10-22 | Sdl Netherlands B.V. | External content libraries |
US9336185B1 (en) * | 2012-09-18 | 2016-05-10 | Amazon Technologies, Inc. | Generating an electronic publication sample |
US9916306B2 (en) | 2012-10-19 | 2018-03-13 | Sdl Inc. | Statistical linguistic analysis of source content |
CN102999486B (en) * | 2012-11-16 | 2016-12-21 | 沈阳雅译网络技术有限公司 | Phrase rule abstracting method based on combination |
CN103870494A (en) * | 2012-12-14 | 2014-06-18 | 中兴通讯股份有限公司 | Method, device and terminal for setting bookmark of browser |
US9298703B2 (en) | 2013-02-08 | 2016-03-29 | Machine Zone, Inc. | Systems and methods for incentivizing user feedback for translation processing |
US9031829B2 (en) | 2013-02-08 | 2015-05-12 | Machine Zone, Inc. | Systems and methods for multi-user multi-lingual communications |
US10650103B2 (en) | 2013-02-08 | 2020-05-12 | Mz Ip Holdings, Llc | Systems and methods for incentivizing user feedback for translation processing |
US9231898B2 (en) | 2013-02-08 | 2016-01-05 | Machine Zone, Inc. | Systems and methods for multi-user multi-lingual communications |
US9600473B2 (en) | 2013-02-08 | 2017-03-21 | Machine Zone, Inc. | Systems and methods for multi-user multi-lingual communications |
US8996352B2 (en) | 2013-02-08 | 2015-03-31 | Machine Zone, Inc. | Systems and methods for correcting translations in multi-user multi-lingual communications |
US9501506B1 (en) | 2013-03-15 | 2016-11-22 | Google Inc. | Indexing system |
US9483568B1 (en) | 2013-06-05 | 2016-11-01 | Google Inc. | Indexing system |
US9355372B2 (en) | 2013-07-03 | 2016-05-31 | Thomson Reuters Global Resources | Method and system for simplifying implicit rhetorical relation prediction in large scale annotated corpus |
CA2917153C (en) * | 2013-07-03 | 2022-05-17 | Thomson Reuters Global Resources | Method and system for simplifying implicit rhetorical relation prediction in large scale annotated corpus |
JP6058513B2 (en) * | 2013-10-01 | 2017-01-11 | 日本電信電話株式会社 | Word order rearrangement device, translation device, method, and program |
US9336186B1 (en) * | 2013-10-10 | 2016-05-10 | Google Inc. | Methods and apparatus related to sentence compression |
JP6226321B2 (en) * | 2013-10-23 | 2017-11-08 | 株式会社サン・フレア | Translation support system, translation support system server, translation support system client, translation support system control method, and program thereof |
KR102256291B1 (en) * | 2013-11-15 | 2021-05-27 | 삼성전자 주식회사 | Method for recognizing a translatable situation and performancing a translatable function and electronic device implementing the same |
RU2592395C2 (en) | 2013-12-19 | 2016-07-20 | Общество с ограниченной ответственностью "Аби ИнфоПоиск" | Resolution semantic ambiguity by statistical analysis |
CN103645931B (en) * | 2013-12-25 | 2016-06-22 | 盛杰 | The method of code conversion and device |
JP6323828B2 (en) * | 2013-12-27 | 2018-05-16 | インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Machines Corporation | Support device, information processing method, and program |
RU2586577C2 (en) | 2014-01-15 | 2016-06-10 | Общество с ограниченной ответственностью "Аби ИнфоПоиск" | Filtering arcs parser graph |
US11334720B2 (en) | 2019-04-17 | 2022-05-17 | International Business Machines Corporation | Machine learned sentence span inclusion judgments |
EP3155542A4 (en) * | 2014-06-11 | 2017-10-18 | Hewlett-Packard Development Company, L.P. | Functional summarization of non-textual content based on a meta-algorithmic pattern |
US9582501B1 (en) * | 2014-06-16 | 2017-02-28 | Yseop Sa | Techniques for automatic generation of natural language text |
US9524293B2 (en) | 2014-08-15 | 2016-12-20 | Google Inc. | Techniques for automatically swapping languages and/or content for machine translation |
RU2596600C2 (en) | 2014-09-02 | 2016-09-10 | Общество с ограниченной ответственностью "Аби Девелопмент" | Methods and systems for processing images of mathematical expressions |
US9372848B2 (en) | 2014-10-17 | 2016-06-21 | Machine Zone, Inc. | Systems and methods for language detection |
US10162811B2 (en) | 2014-10-17 | 2018-12-25 | Mz Ip Holdings, Llc | Systems and methods for language detection |
US11238090B1 (en) | 2015-11-02 | 2022-02-01 | Narrative Science Inc. | Applied artificial intelligence technology for using narrative analytics to automatically generate narratives from visualization data |
US11922344B2 (en) | 2014-10-22 | 2024-03-05 | Narrative Science Llc | Automatic generation of narratives from data using communication goals and narrative analytics |
US9626358B2 (en) | 2014-11-26 | 2017-04-18 | Abbyy Infopoisk Llc | Creating ontologies by analyzing natural language texts |
US10333696B2 (en) | 2015-01-12 | 2019-06-25 | X-Prime, Inc. | Systems and methods for implementing an efficient, scalable homomorphic transformation of encrypted data with minimal data expansion and improved processing efficiency |
JP6607482B2 (en) * | 2015-02-02 | 2019-11-20 | 国立研究開発法人情報通信研究機構 | Syntax analysis device, learning device, machine translation device, and program |
US9767193B2 (en) * | 2015-03-27 | 2017-09-19 | Fujitsu Limited | Generation apparatus and method |
JP6565262B2 (en) * | 2015-03-27 | 2019-08-28 | 富士通株式会社 | Abbreviated sentence generation apparatus, method, and program |
CN105117389B (en) * | 2015-07-28 | 2018-01-19 | 百度在线网络技术(北京)有限公司 | Interpretation method and device |
CN106484682B (en) | 2015-08-25 | 2019-06-25 | 阿里巴巴集团控股有限公司 | Machine translation method, device and electronic equipment based on statistics |
CN106484681B (en) | 2015-08-25 | 2019-07-09 | 阿里巴巴集团控股有限公司 | A kind of method, apparatus and electronic equipment generating candidate translation |
US9990361B2 (en) * | 2015-10-08 | 2018-06-05 | Facebook, Inc. | Language independent representations |
US10586168B2 (en) | 2015-10-08 | 2020-03-10 | Facebook, Inc. | Deep translations |
US10614167B2 (en) | 2015-10-30 | 2020-04-07 | Sdl Plc | Translation review workflow systems and methods |
US11222184B1 (en) | 2015-11-02 | 2022-01-11 | Narrative Science Inc. | Applied artificial intelligence technology for using narrative analytics to automatically generate narratives from bar charts |
US11188588B1 (en) | 2015-11-02 | 2021-11-30 | Narrative Science Inc. | Applied artificial intelligence technology for using narrative analytics to interactively generate narratives from visualization data |
US11232268B1 (en) | 2015-11-02 | 2022-01-25 | Narrative Science Inc. | Applied artificial intelligence technology for using narrative analytics to automatically generate narratives from line charts |
US10765956B2 (en) | 2016-01-07 | 2020-09-08 | Machine Zone Inc. | Named entity recognition on chat data |
CN106021239B (en) * | 2016-04-29 | 2018-10-26 | 北京创鑫旅程网络技术有限公司 | A kind of translation quality real-time estimating method |
US10853583B1 (en) | 2016-08-31 | 2020-12-01 | Narrative Science Inc. | Applied artificial intelligence technology for selective control over narrative generation from visualizations of data |
US10346548B1 (en) * | 2016-09-26 | 2019-07-09 | Lilt, Inc. | Apparatus and method for prefix-constrained decoding in a neural machine translation system |
US10235362B1 (en) * | 2016-09-28 | 2019-03-19 | Amazon Technologies, Inc. | Continuous translation refinement with automated delivery of re-translated content |
US10275459B1 (en) | 2016-09-28 | 2019-04-30 | Amazon Technologies, Inc. | Source language content scoring for localizability |
US10261995B1 (en) | 2016-09-28 | 2019-04-16 | Amazon Technologies, Inc. | Semantic and natural language processing for content categorization and routing |
KR102130429B1 (en) * | 2016-11-07 | 2020-07-07 | 한화테크윈 주식회사 | Method and device for decoding multimedia file |
US10685189B2 (en) | 2016-11-17 | 2020-06-16 | Goldman Sachs & Co. LLC | System and method for coupled detection of syntax and semantics for natural language understanding and generation |
US11954445B2 (en) | 2017-02-17 | 2024-04-09 | Narrative Science Llc | Applied artificial intelligence technology for narrative generation based on explanation communication goals |
US11068661B1 (en) | 2017-02-17 | 2021-07-20 | Narrative Science Inc. | Applied artificial intelligence technology for narrative generation based on smart attributes |
US10943069B1 (en) | 2017-02-17 | 2021-03-09 | Narrative Science Inc. | Applied artificial intelligence technology for narrative generation based on a conditional outcome framework |
US11568148B1 (en) | 2017-02-17 | 2023-01-31 | Narrative Science Inc. | Applied artificial intelligence technology for narrative generation based on explanation communication goals |
US10762304B1 (en) | 2017-02-17 | 2020-09-01 | Narrative Science | Applied artificial intelligence technology for performing natural language generation (NLG) using composable communication goals and ontologies to generate narrative stories |
US11386274B2 (en) * | 2017-05-10 | 2022-07-12 | Oracle International Corporation | Using communicative discourse trees to detect distributed incompetence |
US10599885B2 (en) * | 2017-05-10 | 2020-03-24 | Oracle International Corporation | Utilizing discourse structure of noisy user-generated content for chatbot learning |
US11373632B2 (en) * | 2017-05-10 | 2022-06-28 | Oracle International Corporation | Using communicative discourse trees to create a virtual persuasive dialogue |
US10679011B2 (en) * | 2017-05-10 | 2020-06-09 | Oracle International Corporation | Enabling chatbots by detecting and supporting argumentation |
US20220284194A1 (en) * | 2017-05-10 | 2022-09-08 | Oracle International Corporation | Using communicative discourse trees to detect distributed incompetence |
US11586827B2 (en) * | 2017-05-10 | 2023-02-21 | Oracle International Corporation | Generating desired discourse structure from an arbitrary text |
US10796102B2 (en) * | 2017-05-10 | 2020-10-06 | Oracle International Corporation | Enabling rhetorical analysis via the use of communicative discourse trees |
US11615145B2 (en) | 2017-05-10 | 2023-03-28 | Oracle International Corporation | Converting a document into a chatbot-accessible form via the use of communicative discourse trees |
US10817670B2 (en) * | 2017-05-10 | 2020-10-27 | Oracle International Corporation | Enabling chatbots by validating argumentation |
US10839154B2 (en) * | 2017-05-10 | 2020-11-17 | Oracle International Corporation | Enabling chatbots by detecting and supporting affective argumentation |
US10839161B2 (en) | 2017-06-15 | 2020-11-17 | Oracle International Corporation | Tree kernel learning for text classification into classes of intent |
US11100144B2 (en) | 2017-06-15 | 2021-08-24 | Oracle International Corporation | Data loss prevention system for cloud security based on document discourse analysis |
US10417350B1 (en) | 2017-08-28 | 2019-09-17 | Amazon Technologies, Inc. | Artificial intelligence system for automated adaptation of text-based classification models for multiple languages |
WO2019060353A1 (en) | 2017-09-21 | 2019-03-28 | Mz Ip Holdings, Llc | System and method for translating chat messages |
US11182412B2 (en) * | 2017-09-27 | 2021-11-23 | Oracle International Corporation | Search indexing using discourse trees |
EP3688609A1 (en) | 2017-09-28 | 2020-08-05 | Oracle International Corporation | Determining cross-document rhetorical relationships based on parsing and identification of named entities |
CN116992859A (en) | 2017-09-28 | 2023-11-03 | 甲骨文国际公司 | Enabling autonomous agents to differentiate between questions and requests |
US11809825B2 (en) | 2017-09-28 | 2023-11-07 | Oracle International Corporation | Management of a focused information sharing dialogue based on discourse trees |
US10635863B2 (en) | 2017-10-30 | 2020-04-28 | Sdl Inc. | Fragment recall and adaptive automated translation |
WO2019107623A1 (en) * | 2017-11-30 | 2019-06-06 | 주식회사 시스트란인터내셔널 | Machine translation method and apparatus therefor |
US10817676B2 (en) | 2017-12-27 | 2020-10-27 | Sdl Inc. | Intelligent routing services and systems |
US11042708B1 (en) | 2018-01-02 | 2021-06-22 | Narrative Science Inc. | Context saliency-based deictic parser for natural language generation |
US10963649B1 (en) | 2018-01-17 | 2021-03-30 | Narrative Science Inc. | Applied artificial intelligence technology for narrative generation using an invocable analysis service and configuration-driven analytics |
US11537645B2 (en) * | 2018-01-30 | 2022-12-27 | Oracle International Corporation | Building dialogue structure by using communicative discourse trees |
CN111670435A (en) | 2018-01-30 | 2020-09-15 | 甲骨文国际公司 | Detecting requests for interpretation using a communication utterance tree |
WO2019183144A1 (en) * | 2018-03-19 | 2019-09-26 | Coffing Daniel L | Processing natural language arguments and propositions |
US11328016B2 (en) | 2018-05-09 | 2022-05-10 | Oracle International Corporation | Constructing imaginary discourse trees to improve answering convergent questions |
US11455494B2 (en) | 2018-05-30 | 2022-09-27 | Oracle International Corporation | Automated building of expanded datasets for training of autonomous agents |
US11334726B1 (en) | 2018-06-28 | 2022-05-17 | Narrative Science Inc. | Applied artificial intelligence technology for using natural language processing to train a natural language generation system with respect to date and number textual features |
US11645459B2 (en) * | 2018-07-02 | 2023-05-09 | Oracle International Corporation | Social autonomous agent implementation using lattice queries and relevancy detection |
US11429794B2 (en) | 2018-09-06 | 2022-08-30 | Daniel L. Coffing | System for providing dialogue guidance |
EP3850781A4 (en) | 2018-09-14 | 2022-05-04 | Coffing, Daniel L. | Fact management system |
US11256867B2 (en) | 2018-10-09 | 2022-02-22 | Sdl Inc. | Systems and methods of machine learning for digital assets and message creation |
US11562135B2 (en) | 2018-10-16 | 2023-01-24 | Oracle International Corporation | Constructing conclusive answers for autonomous agents |
CN109710952B (en) * | 2018-12-27 | 2023-06-16 | 北京百度网讯科技有限公司 | Translation history retrieval method, device, equipment and medium based on artificial intelligence |
US10990767B1 (en) | 2019-01-28 | 2021-04-27 | Narrative Science Inc. | Applied artificial intelligence technology for adaptive natural language understanding |
US11321536B2 (en) * | 2019-02-13 | 2022-05-03 | Oracle International Corporation | Chatbot conducting a virtual social dialogue |
CN110069624B (en) | 2019-04-28 | 2021-05-04 | 北京小米智能科技有限公司 | Text processing method and device |
US11275892B2 (en) | 2019-04-29 | 2022-03-15 | International Business Machines Corporation | Traversal-based sentence span judgements |
US11263394B2 (en) * | 2019-08-02 | 2022-03-01 | Adobe Inc. | Low-resource sentence compression system |
US11449682B2 (en) | 2019-08-29 | 2022-09-20 | Oracle International Corporation | Adjusting chatbot conversation to user personality and mood |
US11599731B2 (en) * | 2019-10-02 | 2023-03-07 | Oracle International Corporation | Generating recommendations by using communicative discourse trees of conversations |
US11556698B2 (en) * | 2019-10-22 | 2023-01-17 | Oracle International Corporation | Augmenting textual explanations with complete discourse trees |
US11580298B2 (en) | 2019-11-14 | 2023-02-14 | Oracle International Corporation | Detecting hypocrisy in text |
US11501085B2 (en) | 2019-11-20 | 2022-11-15 | Oracle International Corporation | Employing abstract meaning representation to lay the last mile towards reading comprehension |
US11775772B2 (en) | 2019-12-05 | 2023-10-03 | Oracle International Corporation | Chatbot providing a defeating reply |
CN111104807A (en) * | 2019-12-06 | 2020-05-05 | 北京搜狗科技发展有限公司 | Data processing method and device and electronic equipment |
US11847420B2 (en) | 2020-03-05 | 2023-12-19 | Oracle International Corporation | Conversational explainability |
US11074402B1 (en) * | 2020-04-07 | 2021-07-27 | International Business Machines Corporation | Linguistically consistent document annotation |
US11475210B2 (en) * | 2020-08-31 | 2022-10-18 | Twilio Inc. | Language model for abstractive summarization |
US11822892B2 (en) * | 2020-12-16 | 2023-11-21 | International Business Machines Corporation | Automated natural language splitting for generation of knowledge graphs |
KR102562920B1 (en) * | 2020-12-29 | 2023-08-02 | 엑스엘에이트 아이앤씨 | Apparatus and method for machine translation |
US11765267B2 (en) | 2020-12-31 | 2023-09-19 | Twilio Inc. | Tool for annotating and reviewing audio conversations |
CN112784612B (en) * | 2021-01-26 | 2023-12-22 | 浙江香侬慧语科技有限责任公司 | Method, device, medium and equipment for synchronous machine translation based on iterative modification |
US11809804B2 (en) | 2021-05-26 | 2023-11-07 | Twilio Inc. | Text formatter |
CN113705158A (en) * | 2021-09-26 | 2021-11-26 | 上海一者信息科技有限公司 | Method for intelligently restoring original text style in document translation |
CN115795039B (en) * | 2023-02-08 | 2023-06-02 | 成都索贝数码科技股份有限公司 | Style title generation method, equipment and medium based on deep learning |
Family Cites Families (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS61217871A (en) * | 1985-03-25 | 1986-09-27 | Toshiba Corp | Translation processor |
DE3616751A1 (en) * | 1985-05-20 | 1986-11-20 | Sharp K.K., Osaka | TRANSLATION SYSTEM |
JPH02301869A (en) * | 1989-05-17 | 1990-12-13 | Hitachi Ltd | Method for maintaining and supporting natural language processing system |
US5369574A (en) * | 1990-08-01 | 1994-11-29 | Canon Kabushiki Kaisha | Sentence generating system |
US5477451A (en) * | 1991-07-25 | 1995-12-19 | International Business Machines Corp. | Method and system for natural language translation |
GB2279164A (en) * | 1993-06-18 | 1994-12-21 | Canon Res Ct Europe Ltd | Processing a bilingual database. |
US6304841B1 (en) * | 1993-10-28 | 2001-10-16 | International Business Machines Corporation | Automatic construction of conditional exponential models from elementary features |
US5510981A (en) * | 1993-10-28 | 1996-04-23 | International Business Machines Corporation | Language translation apparatus and method using context-based translation models |
US5642520A (en) * | 1993-12-07 | 1997-06-24 | Nippon Telegraph And Telephone Corporation | Method and apparatus for recognizing topic structure of language data |
JP3377290B2 (en) * | 1994-04-27 | 2003-02-17 | シャープ株式会社 | Machine translation device with idiom processing function |
JP2855409B2 (en) * | 1994-11-17 | 1999-02-10 | 日本アイ・ビー・エム株式会社 | Natural language processing method and system |
GB2295470A (en) | 1994-11-28 | 1996-05-29 | Sharp Kk | Machine translation system |
US5903858A (en) | 1995-06-23 | 1999-05-11 | Saraki; Masashi | Translation machine for editing a original text by rewriting the same and translating the rewrote one |
JP3579204B2 (en) * | 1997-01-17 | 2004-10-20 | 富士通株式会社 | Document summarizing apparatus and method |
US5991710A (en) * | 1997-05-20 | 1999-11-23 | International Business Machines Corporation | Statistical translation system with features based on phrases or groups of words |
DE69837979T2 (en) * | 1997-06-27 | 2008-03-06 | International Business Machines Corp. | System for extracting multilingual terminology |
US6112168A (en) * | 1997-10-20 | 2000-08-29 | Microsoft Corporation | Automatically recognizing the discourse structure of a body of text |
US6533822B2 (en) * | 1998-01-30 | 2003-03-18 | Xerox Corporation | Creating summaries along with indicators, and automatically positioned tabs |
GB2337611A (en) * | 1998-05-20 | 1999-11-24 | Sharp Kk | Multilingual document retrieval system |
GB2338089A (en) * | 1998-06-02 | 1999-12-08 | Sharp Kk | Indexing method |
US6092034A (en) * | 1998-07-27 | 2000-07-18 | International Business Machines Corporation | Statistical translation system and method for fast sense disambiguation and translation of large corpora using fertility models and sense models |
JP2000132550A (en) * | 1998-10-26 | 2000-05-12 | Matsushita Electric Ind Co Ltd | Chinese generating device for machine translation |
US6393389B1 (en) * | 1999-09-23 | 2002-05-21 | Xerox Corporation | Using ranked translation choices to obtain sequences indicating meaning of multi-token expressions |
-
2001
- 2001-05-11 JP JP2001583366A patent/JP2004501429A/en active Pending
- 2001-05-11 AU AU2001261505A patent/AU2001261505A1/en not_active Abandoned
- 2001-05-11 EP EP01935406A patent/EP1352338A2/en not_active Withdrawn
- 2001-05-11 AU AU2001261506A patent/AU2001261506A1/en not_active Abandoned
- 2001-05-11 CN CN01812317A patent/CN1465018A/en active Pending
- 2001-05-11 WO PCT/US2001/015380 patent/WO2001086489A2/en active Search and Examination
- 2001-05-11 US US09/854,301 patent/US20020046018A1/en not_active Abandoned
- 2001-05-11 CA CA002408819A patent/CA2408819C/en not_active Expired - Lifetime
- 2001-05-11 US US09/854,327 patent/US7533013B2/en not_active Expired - Lifetime
- 2001-05-11 WO PCT/US2001/015379 patent/WO2001086491A2/en active Application Filing
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10319252B2 (en) | 2005-11-09 | 2019-06-11 | Sdl Inc. | Language capability assessment and training apparatus and techniques |
US9122674B1 (en) | 2006-12-15 | 2015-09-01 | Language Weaver, Inc. | Use of annotations in statistical machine translation |
US8825466B1 (en) | 2007-06-08 | 2014-09-02 | Language Weaver, Inc. | Modification of annotated bilingual segment pairs in syntax-based machine translation |
US8990064B2 (en) | 2009-07-28 | 2015-03-24 | Language Weaver, Inc. | Translating documents based on content |
US8942973B2 (en) | 2012-03-09 | 2015-01-27 | Language Weaver, Inc. | Content page URL translation |
US10261994B2 (en) | 2012-05-25 | 2019-04-16 | Sdl Inc. | Method and system for automatic management of reputation of translators |
US9152622B2 (en) | 2012-11-26 | 2015-10-06 | Language Weaver, Inc. | Personalized machine translation via online adaptation |
US9213694B2 (en) | 2013-10-10 | 2015-12-15 | Language Weaver, Inc. | Efficient online domain adaptation |
Also Published As
Publication number | Publication date |
---|---|
WO2001086489A3 (en) | 2003-07-24 |
WO2001086491A3 (en) | 2003-08-14 |
EP1352338A2 (en) | 2003-10-15 |
AU2001261506A1 (en) | 2001-11-20 |
US20020046018A1 (en) | 2002-04-18 |
WO2001086489A2 (en) | 2001-11-15 |
CA2408819A1 (en) | 2001-11-15 |
JP2004501429A (en) | 2004-01-15 |
US20020040292A1 (en) | 2002-04-04 |
WO2001086491A2 (en) | 2001-11-15 |
AU2001261505A1 (en) | 2001-11-20 |
CN1465018A (en) | 2003-12-31 |
US7533013B2 (en) | 2009-05-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CA2408819C (en) | Machine translation techniques | |
JP5774751B2 (en) | Extracting treelet translation pairs | |
Al-Onaizan et al. | Statistical machine translation | |
Quirk et al. | Dependency treelet translation: Syntactically informed phrasal SMT | |
Marcu et al. | The automatic translation of discourse structures | |
Koehn | Noun phrase translation | |
EP0568319B1 (en) | Machine translation system | |
US8249856B2 (en) | Machine translation | |
US9047275B2 (en) | Methods and systems for alignment of parallel text corpora | |
JP5586817B2 (en) | Extracting treelet translation pairs | |
KR20050045822A (en) | System for identifying paraphrases using machine translation techniques | |
EP1390868A1 (en) | Statistical memory-based translation system | |
Callison-Burch et al. | Co-training for statistical machine translation | |
Yuan | Grammatical error correction in non-native English | |
Lavie et al. | Experiments with a Hindi-to-English transfer-based MT system under a miserly data scenario | |
Menezes et al. | Syntactic models for structural word insertion and deletion during translation | |
Horvat | Hierarchical statistical semantic translation and realization | |
Huang | Forest-based algorithms in natural language processing | |
Jones et al. | A probabilistic parser and its applications | |
DeNeefe | Tree-adjoining machine translation | |
Williams | Unification-based constraints for statistical machine translation | |
Quirk et al. | Dependency tree translation: Syntactically informed phrasal smt | |
Islam | Towards achieving a delicate blending between rule-based translator and neural machine translator for Bengali to English translation | |
Markantonatou et al. | Hybrid machine translation for low-and middle-density languages | |
Levinboim | Invertibility and Transitivity in Low-resource Machine Translation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
EEER | Examination request | ||
MKEX | Expiry |
Effective date: 20210511 |