WO1996010795A1 - A database accessing system - Google Patents
A database accessing system Download PDFInfo
- Publication number
- WO1996010795A1 WO1996010795A1 PCT/US1995/012600 US9512600W WO9610795A1 WO 1996010795 A1 WO1996010795 A1 WO 1996010795A1 US 9512600 W US9512600 W US 9512600W WO 9610795 A1 WO9610795 A1 WO 9610795A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- database
- word
- request
- operative
- entry
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/90335—Query processing
-
- 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
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- 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
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/3332—Query translation
- G06F16/3338—Query expansion
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y10—TECHNICAL SUBJECTS COVERED BY FORMER USPC
- Y10S—TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y10S707/00—Data processing: database and file management or data structures
- Y10S707/99931—Database or file accessing
- Y10S707/99933—Query processing, i.e. searching
- Y10S707/99935—Query augmenting and refining, e.g. inexact access
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y10—TECHNICAL SUBJECTS COVERED BY FORMER USPC
- Y10S—TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y10S707/00—Data processing: database and file management or data structures
- Y10S707/99931—Database or file accessing
- Y10S707/99933—Query processing, i.e. searching
- Y10S707/99936—Pattern matching access
Definitions
- the present invention relates to methods and apparatus for accessing databases generally and for accessing databases via the telephone, in particular.
- the present invention seeks to provide an improved system for accessing a computerized database from a location which is not in computer communication with the database.
- a database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including a sequence of representations of possibly erroneous user inputs, the system including a similar word finder opera ⁇ tive, for at least one interpretation of each representa ⁇ tion, to find at least one database word which is at least similar to that interpretation, and a database entry evaluator operative, for each database word found by the similar word finder, to assign similarity values for relevant entries in the database, the values repre ⁇ senting the degree of similarity between each database entry and the request.
- the request includes a sequence of unambiguous representations of possibly erroneous user inputs.
- the relevant entries include all database entries having the database word in an individual context.
- the request includes a sequence of ambiguous representations of possibly erroneous user inputs and also including an ambiguous input evaluator operative to identify at least one inter- pretation of each ambiguous representation and to provide the interpretation to the similar word finder.
- the ambiguous input evalua ⁇ tor includes a database analyzer operative to extract database characteristics from the database.
- system also includes a session manager receiving similarity values from the database entry evaluator and operative to inter ⁇ actively provide a user sensible message based on the similarity values.
- the session manager is operative to prompt the user to provide the request by orally spelling at least one database word.
- the session manager is operative to prompt the user to provide the request by pronouncing at least one database word.
- the session manager is operative to prompt the user to provide the request by keying at least one database word on the keypad.
- a database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including at least one ambiguous representation of at least one possibly erroneous user input, the system including an ambiguous input evaluator operative to identify at least one interpretation of each ambiguous representation, a similar word finder operative, for at least one interpretation of each representation, to find at least one database word which is at least similar to that interpretation, and a database entry evaluator
- the relevant entries in ⁇ clude all database entries having the database word in an individual context.
- the ambiguous input evaluator includes a database analyzer operative to extract database characteristics from the database.
- system also includes a session manager receiving similarity values from the database entry evaluator and operative to inter ⁇ actively provide a user sensible message based on the similarity values.
- the session manager is operative to prompt the user to provide the request by orally spelling at least one database word.
- the session manager is operative to prompt the user to provide the request by pronouncing at least one database word.
- the session manager is operative to prompt the user to provide the request by keying at least one database word on the keypad.
- a database accessing method for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including a sequence of representations of possibly erroneous user inputs, the method including finding, for at least one interpretation of each representation, at least one database word which is at least similar to that interpretation, and assigning similarity values for relevant entries in the database, for each database word found, the values representing the degree of similarity between each database entry and the request.
- a database accessing method for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including at least one ambiguous representation of at least one possibly erroneous user input, the method including identifying at least one interpretation of each ambiguous representation, finding, for at least one interpretation of each representation, at least one database word which is at least similar to that interpretation, and for each database word found by the similar word finder, assigning similarity values for relevant entries in the database, the values representing the degree of similarity between each database entry and the request.
- database as used herein is intended to include any type of database including but not limited to data ⁇ bases including financial and commercial data, databases including administrative data such as personnel data, transportation vehicle flight schedule databases, databases including technological data such as weather data, and databases including population data such as zipcode directories, classified and unclassified tele ⁇ phone directories.
- Fig. 1 is a high level functional block diagram of a database access system constructed and operative in accordance with a preferred embodiment of the present invention
- Fig. 2 is a flowchart illustrating a preferred method of operation for the spellguess unit of Fig. 1;
- Fig. 3 is a flowchart illustrating a preferred method for generating the phonetic translator of Fig. 1;
- Fig. 4 is a flowchart illustrating a preferred method of operation for the dictionary lookup unit of Fig. 1;
- Fig. 5 is a flowchart illustrating a preferred method of operation for database lookup and candidate sorter units of Fig. 1;
- Fig. 6 is a flowchart illustrating a preferred method of operation for the conversation manager of Fig.
- Fig. 8 is a flowchart illustrating a preferred method for loading the database into memory and for updating the concordance prepared by the method of Fig. 7;
- Fig. 9 is a simplified block diagram of one implementation of the apparatus of Fig. 1 in which the ambiguous input comprises telephone keypad input;
- Fig. 10 is a simplified block diagram of one implementation of the apparatus of Fig. 1 in which the ambiguous input comprises output of a probabilistic spoken character recognition device
- Fig. 11 is a simplified block diagram of one implementation of the apparatus of Fig. 1 in which the ambiguous input comprises output of a probabilistic spoken phoneme recognition device;
- Fig. 12 is a diagram of the computer memory data structures serving the apparatus of Fig. 1, includ ⁇ ing a dictionary, concordance and database storing typi ⁇ cal data and the internal connections between these structures;
- Fig. 13 is a diagram of a similarity search index, storing typical data, which is useful in perform ⁇ ing the dictionary look-up method of Fig. 4.
- FIG. 1 is a high level functional block diagram of a database acce& ⁇ system constructed and operative in accordance with a preferred embodiment of the present invention.
- a phonetic translator 40 receives the possible strings generated by spellguess unit 30 and translates each of them into phonetic form.
- a preferred method for generating translator 40 is described below with refer ⁇ ence to Fig. 3. If translator 40 is generated by the method of Fig. 3, the complexity of phonetic translator 40 is independent of the number of phonetic rules as described below with reference to Fig. 3.
- conventional phonetic translation methods may be employed such as the methods described in the following publication:
- the phonetic translator converts strings in the English alphabet to strings in a phonetic alphabet such as the DECvoice phonetic alphabet developed by Digital Equipment Corpora ⁇ tion. More generally, however, the phonetic translator may be operative to translate strings from any first alphabet into any second alphabet.
- the DECvoice phonetic alphabet is described in the following publication:
- the phonetic translator may be omitted in which case the access to the dictionary may be based directly on the guessed strings rather than on the translation of the guessed strings into phonetic form.
- a dictionary look up unit 50 receives each phonetic or alphabetical string and conducts a similarity search within a dictionary 60.
- the dictionary 60 comprises an alphabetical or phonetically ordered or otherwise ordered list of some or all words or strings appearing in the database 70 which is to be accessed.
- a concordance 72 associates the dictionary 60 with the database 70.
- a similarity search index 74 allows the dictionary 60 to be searched for entries which are similar to a desired user-selected string.
- a numeric index 76 is provided if a telephone keypad is used which extracts from the dictionary all entries having a given numerical code.
- a trigram stochas ⁇ tic finite state automaton 78 stores trigrams and their probabilities and is accessed by spellguess unit 30, as described in more detail below.
- Dictionary 60, database 70, concordance 71, numeric index 76 and trigram stochastic finite state automaton 78 may be stored in any suitable electronic medium such as a hard disc or RAM.
- a preferred method for loading the dictionary, similarity index, numeric index and trigram automaton into memory, as well as for prepar ⁇ ing the concordance, is described below with reference to Fig. 7.
- a preferred method for loading the database into memory and for updating the concordance prepared by the method of Fig. 7 is described below with reference to Fig. 8.
- the output of the dictionary lookup unit 50 typically comprises, for each string received from unit 30 or 40: a. A possibly empty list of dictionary entries whose similarity to the string is better than or equal to a predetermined similarity threshold; and b. A similarity score representing the degree of similarity between that entry and the corresponding string.
- Spellguess unit 30 guesses various spellings, among which, for the purposes of this example, the word CATHERINE is assumed to appear. This word is input to the phonetic translator 40, which may translate it into (K Ae Th Eh R lh N) . The phonetically translated string is sent to the dictionary look up unit 50, which may produce the following list of dictionary words and corresponding similarity scores:
- the database 70 is now accessed by a database look up unit 90 so as to identify, for each dictionary entry, all database entries which include that dictionary entry.
- the database entries identified by database lookup 90 for each dictionary entry include only those database entries which include the dictionary entry in the correct context.
- a conversation manag ⁇ er 80 may direct the user to enter the name of the city of interest and subsequently to enter the name of the street of interest.
- the database entries identified for a dictionary entry comprising a putative street name comprise only those database entries having that street name.
- Database lookup unit 90 generates and continu ⁇ ally updates, as the call or database access session progresses, a similarity vector 92.
- the similarity vector 92 has a number of vector components which correspond to the number of database entries. Each similarity vector component stores a similarity score of the corresponding database entry. This similarity score is provided by the dictionary lookup 50 for each user-supplied word as the database access session proceeds.
- the database lookup 90 accumulates these similarity scores in the appropriate locations within the similarity vector.
- a candi ⁇ date sorter 100 extracts a predetermined number of the highest scoring similarity vector components and stores these components in a best candidates box 94.
- the simi ⁇ larity vector 92 and best candidates box 94 are prefera ⁇ bly stored in computer memory and are initialized at the beginning of each database access session.
- the conversation manager 80 receives the output of unit 90 or 100 and determines a course of action.
- the considerations of the conversation manager may, for example, include the following: a. If one database entry has a similarity score which significantly exceeds the similarity score of all other candidates and which exceeds a threshold similarity score, this entry may be deemed correct and therefore is presented to the user for confirmation.
- each entry in database 70 comprises a plurality of words or strings.
- each entry in a zipcode directory typically includes some or all of the following words: state name, street name, street type, house number, locality.
- the conversation manager may decide to prompt the user to enter information identifying another word within the desired database entry (such as "street name"). c. If a small number of entries have similar or identical scores, the conversation manager may decide to present all of these to the user and prompt the user to select one of them about which s/he wishes to receive more information. d. If a large number of entries have similar or identical scores and all strings within the entry have been processed, the conversation manager may wish to refer the user to the services of a human operator. In this case, the conversation manager 80 provides a suit ⁇ able routing instruction to conventional telephone equip- ment which routes the user to a human operator.
- the output of the conversation manager is a suitable message to the user which is received by a text- to-speech voice synthesizer 110, such as DECvoice, com ⁇ pitchally available from Digital Equipment Corporation, Maynard, MA, USA or Text Assist, commercially available from Creative Labs, Inc., Milpitas, CA, USA.
- a text- to-speech voice synthesizer 110 such as DECvoice, com ⁇ pitchally available from Digital Equipment Corporation, Maynard, MA, USA or Text Assist, commercially available from Creative Labs, Inc., Milpitas, CA, USA.
- the output of the conversation manager may be received by a pre-recorded voice player 115 which plays a selected message from among a plurality of re ⁇ corded messages. Examples of a commercially available voice player 115 include the DECvoice or Sound Blaster, commercially available from Creative Labs, Inc., Milpi ⁇ tas, CA, USA.
- Fig. 1 utilizes the following data structures in order to perform a similarity dictionary lookup, rather than an exact match lookup, and to support spell guess functions: a. Word dictionary 60 - A list of entries which includes a single entry for each uniquely spelled database word for which direct access is required.
- the words in the database which are to be accessed directly may include only the localities or alternatively, the state as well as the street name and locality.
- the dictionary includes a count for every word appearance in the database in the specific context, and a pointer to the concordance. For example, if the word MARYLAND appears in the database as a state name, a city name and a street name, the dictionary will contain only a single entry for the word MARYLAND, with separate counters for the number of times the word appears as state name, city name and street name, and a single pointer to the concordance, which points to the first occurrence of the word MARYLAND in the concordance, b.
- Concordance 72 - A list of entries including a single entry for each occurrence of each dictionary word in the database.Concordance entries relating to the same dictionary word are always arranged consecutively and sorted, within the dictionary word, according to the entry context (for example, appearances of the dictionary word in a locality context, followed by ap ⁇ pearances of the dictionary word in a street name con ⁇ text). Each database entry is pointed to by all concord ⁇ ance entries corresponding to that database entry.
- database entries residing in consecutive addresses correspond to a plurality of concordance en ⁇ tries residing in consecutive addresses
- the plurality of concordance entries may be replaced by a single entry pointing to the range of database entries.
- Similarity search index 74 An index into the word dictionary that enables access to the word dictionary by word similarity rather than by exact match of words.
- the similarity search index is preferably organized as an N-ary tree in which each node points to its left son and its right brother.
- An N-ary tree is a well-known data structure which is described in Tremblay, J. P.; Sorenson, P. G., "An introduction to data struc- tures with applications", McGraw-Hill, 1976, pp. 316 318.
- the similarity search index is built in several tree layers, such as 7 or 9 layers. Each layer corre ⁇ sponds to a different "reduced alphabet" of the diction ⁇ ary alphabet, which may be either graphic or phonetic.
- a reduced alphabet for the English lan ⁇ guage alphabet is an alphabet including only two symbols, c for consonants and v for vowels.
- the correspondence between the English alphabet and the reduced alphabet is that all consonants are replaced by c and all vowels are replaced by v.
- a reduced alphabet is an alphabet that contains less symbols than an original alphabet and complies with the following criteria:
- Each layer of the similarity search index contains the same dictionary words but in a different reduced alphabet.
- Each word in the similarity search index is represented in vector format with a reduced alphabet.
- a word in vector format is represented by a vector V of length L, where L is the number of symbols in the alpha ⁇ bet.
- Each element Vi of the vector contains the number of occurrences of the ith alphabetic symbol in the word. For example, given an alphabet ⁇ a,b,c ⁇ and a word "abbac", the vector representation is "2,2,1". To continue the above example of the ⁇ c,v ⁇ alphabet, the words BLAKE and GRACE will both be expressed by the pair "2,3".
- Each layer of the similarity search index is built with a different reduced alphabet adhering to the above guide ⁇ lines.
- Fig. 13 illustrates an example of a "grapheme based" similarity search index for an English language dictionary. It is appreciated that, alternatively, a phonetic-based similarity search index may be employed.
- the index of Fig. 13 is an N-ary tree which includes the following three layers:
- Layer 1 is based on a single symbol reduced alphabet which transforms each English language word into its length.
- Layer 3 is based on three symbols, which trans ⁇ forms each English word into an ordered triplet (a,b,c) where a is the number of letters from among the set (P, F, B, V, W, T, D) in the word, b is the number of letters from among the remaining consonants in the word and c is the number of vowels in the word.
- Each dictionary word is mapped to one node in each index layer.
- the word ADKINS appearing in Fig. 13 is mapped into node (6) in layer 1 since ADKINS contains 6 letters, into node (4,2) in layer 2 , since ADKINS contains 4 consonants and 2 vowels, and into node (1,3,2) in layer 3 since only the letter D belongs to the set (P, F, B, V, W, T, D), and there are 3 other consonants, and 2 vowels.
- Numeric index 76 This index is organized as a balanced tree, and is employed when the ambiguous input generator is a keypad in which one number is used to represent a plurality of letters, such as a tele ⁇ phone keypad. The numeric index uses the ambiguous numer ⁇ ic input value keyed in by the user of the keypad to point directly to all dictionary entries having that numeric input value.
- Trigram stochastic finite state automaton 78 uses the ambiguous numer ⁇ ic input value keyed in by the user of
- An automaton that stores all trigrams appearing in the dictionary along with their probabilities which are computed based on dictionary appearance frequencies.
- the trigram stochastic finite state automaton is employed by spell guess unit 30 to anticipate the probable spelling of any given ambiguous string.
- Fig. 7 which illus ⁇ trates a preferred method for loading the dictionary, similarity index, numeric index and trigram automaton into computer memory, as well as for preparing the con ⁇ cordance in the computer memory.
- the method of Fig. 7 preferably includes the following processes: a. Allocate data structures (process 570) - This process creates the initial structures for the word dictionary, the concordance, similarity search index, the numeric index and the trigram stochastic finite state automaton. b. Scan dictionary words (process 580) - This loop reads all dictionary words from a dictionary organized as a sequential disk file. For each word the following proc ⁇ esses are performed: i. Update word dictionary (process 590) - This process copies the word into the dictionary. ii. Update similarity search index (process 600) - The word is translated into the reduced alphabet of every layer of the index. All index layers are searched, from the highest to the lowest level, and any missing words are added.
- iii Update numeric index (process 610) - If the numeric code of the word does not exist in the index, it is added. Once the numeric code of the word does exist, a pointer is added to it which points at the newly entered word.
- iv. Extract trigrams (process 620) All trigrams contained in the new word are stored in a temporary list. Each time a trigram on the list appears, an associated frequency value is increment- ed. The trigram list including frequency values are employed to construct the automaton in process 640, once all diction- 20
- Each dictionary word occupies one or more entries in the con ⁇ cordance.
- the concordance entries are created at this time, but are not associated with database entries pointers until the database itself is loaded into memory, c.
- Create trigram stochastic finite state automa ⁇ ton process 640) - Based on the trigram information (population and frequencies) accumulated by process 620, a stochastic finite state automaton is built. This auto ⁇ maton will be later used by spellguess unit 30 to guess the spelling of the ambiguous input string.
- a stochastic finite state automaton is described in the above-refer ⁇ enced publication by Lorenz.
- Fig. 8 which illus ⁇ trates a preferred method for loading the database into memory and for updating the concordance prepared by the method of Fig. 7.
- the method of Fig. 8 preferably includes the following processes: a. Allocate database (process 650) - Prepares the memory required to hold the entire database. b. Database scan loop 660 - Reads a plurality of database entries stored in sequential format and residing on magnetic media, such as hard disk drive. For every database entry, the following processes are performed:
- Inner database entry word scan loop (process 680) - Scans all the database entry words that are de ⁇ fined as requiring direct access through the dictionary and concordance. For every such word, the process per ⁇ forms the following two processes: i. Locate word in dictionary (process 690) Since the dictionary is loaded before the database, and since every word requiring direct access has to be present in the dictionary, this process locates the word entry in the dictionary, in order to locate the right entry in the concordance, that corresponds to this data ⁇ base entry. ii. Update concordance (process 700) - Adds a pointer from the concordance entry to the database entry. This pointer establishes a link between a dictionary word and the corresponding database entries.
- spellguess unit 30 is preferably operative to select a plurality of strings of characters to which ambiguous input might correspond, each associated with a probability.
- the output of the spellguess unit is the N most probable strings and their probabilities, where N is any suitable number such as 20.
- the method of Fig. 2 preferably includes the following processes: a. An initialization process 140 in which an initial string is constructed. The process then invokes a recursive string lengthening process 144, until all possible spelling alternatives have been checked. b. A recursive string lengthening process 144 including blocks 150 - 220 which generates a list of the N most probable strings and their probabilities, based on pre-known frequencies of various "trigrams" or 3-letter strings in the dictionary. Process 144 receives as input a partial string of length L whose probability has al ⁇ ready been computed and computes the probability of at least one string of length L+l which includes the input partial string.
- the partial string is taken to include two "dummy" characters such as two # ' s . c.
- Output process 230 the N most probable strings are copied onto a digital storage medium such as computer memory.
- the N most probable strings are retained as a heap sorted by their probabilities.
- Spellguess unit 30 uses the method of Fig. 2 in order to produce a list of the most probable strings represented by the numeric string "2874", along with their probabil ⁇ ities.
- the probability of an individual candidate string is based on the proba ⁇ bilities of each trigram included in the string.
- the probability assigned to BURG might be rough ⁇ ly the product of the known frequencies of occurrence of each of the following trigrams: BUR, URG.
- Block 150 "Minimum probability” is typically 0 if less than N strings have been accumulated. If N strings have already been accumulated, “minimum probabil ⁇ ity” is equal to the probability of the least probable of the N strings.
- the process back ⁇ tracks to the previous recursive level because the proba ⁇ bility of the input partial string will only decline further as its length increases.
- Block 160 If the partial string is, in fact, a "final" string, i.e. a string of the length which the user entered, then it is entered onto a list, typically organized as a heap, of most probable strings (block 190). Otherwise, the process advances directly to block 200.
- a "final” string i.e. a string of the length which the user entered
- the lowest probability string is discarded (block 180) before the new string is inserted (block 190).
- the string is lengthened by considering the possible characters which might fill the (L+l)'th position in the string. For example, if keypad input is employed and the user punched the #2 key for the L+l'th position, the candidate characters are A, B and C. For each candidate character, blocks 210 and 220 are performed.
- Block 210 The candidate character is concate ⁇ nated with the partial string and the probability of the new, longer partial string is defined by generating the product of the probability of the previous partial string and the probability of the trigram consisting of the last three characters of the new, longer partial string.
- Block 220 The next recursive level of the string lengthening process begins, thereby to add yet another character to the current partial string.
- phonetic translator unit 40 is preferably operative to convert strings from one alphabet (such as English), into another alphabet (such as the DECvoice phonetic alphabet).
- the complex ⁇ ity of the phonetic translator is n, where n is the length of the string to be converted, regardless of the number of phonetic translation rules.
- the method of Fig. 3 preferably includes the following processes:
- Process 260 Load translation rules: read an external file that describes phonetic translation rules for translating between the alphabets employed. Prefera ⁇ bly, the syntax of the external file is checked. The structure of a rule is as follows: grapheme->phoneme/"preceding substring"_"following substring"
- the above rule instructs the phonetic transla ⁇ tor to translate a certain grapheme to a certain phoneme if the grapheme is preceded by a certain "preceding substring” and is followed by a certain "following sub ⁇ string". Any of the phoneme, preceding substring and following substring can be empty.
- Additional states may be defined similarly if it is desired to accommodate preceding strings which are more than 2 characters long.
- Process 274 Build a subautomaton for each phonetic rule.
- the subautomaton for an individual phonet ⁇ ic rule has n states, where n is the total number of characters in the "preceding substring", the grapheme and the "following substring".
- the last, n'th, state is also termed herein the Accepting State. Add transitions from each state to the next state which are respectively determined by the characters of the preceding substring, grapheme and following substring.
- the output generated by the subautomaton is the phoneme which the rule in ques ⁇ tion associates with the preceding substring, grapheme and following substring.
- the rule 0->Uh/W_L has 3 states, such as for example (S10,S11,S12) and 3 transitions: W as preceding substring causes a transition from the initial state to state S10, 0 as grapheme causes a transition from S10 to Sll, and L as following substring causes a transition from Sll to the accepting state S12.
- the output of the subautomaton for the above sample rule is the phoneme (Uh) . Since this rule has one character, W, in the "preceding substring", then state SI serves as its initial state. Therefore, the subautomaton for this rule comprises the following transitions: (S1,W,S10), (S10,0,S11), (S11,L,S12).
- Process 280 Convert the non-deterministic automaton (NDFSA) constructed in processes 270-274 into a deterministic finite state automaton (DFSA) including all the phonetic translation rules.
- the conversion process may be based on well known methods, such as the conver ⁇ sion method described in Salomaa, A. , Computation and automata, Cambridge University Press, 1985, pp. 44 - 55.
- Process 290 Build states matrix: Convert the DFSA created by process 280 into a matrix, where the rows correspond respectively to the states in the DFSA and the columns correspond respectively to the possible inputs i.e. to the grapheme alphabet.
- the contents of each matrix cell includes: the next state, whether the state is an accepting state or not, and the phoneme 26
- the phonetic translator 40 is operative to translate a graphic word into a phonetic word as follows:
- a graphic word such as a word in Eng ⁇ lish
- the fol ⁇ lowing examples illustrate the outputs of the phonetic translator unit 40 for a variety of inputs: grapheme input word: BUBBLE phonetic output word: B Ah B L grapheme input word: PHONE phonetic output word: F Ao N grapheme input word: BYTE phonetic output word: B Ay T grapheme input word: WOLF phonetic output word: W Uh L F
- Fig. 4 which illus ⁇ trates a preferred method of operation for dictionary lookup unit 50 of Fig. 1.
- the input to the dictionary look up unit 50 are the strings that were guessed by spellguess unit 30, if the phonetic translator is omitted from the system of Fig. 1, or the strings that were generated by the phonet ⁇ ic translator unit 40.
- Dictionary look up unit 50 does not search for an exact match with the input strings, but rather performs a similarity search in order to find all dictionary words that are similar, to a predetermined extent, to the input strings, if not necessarily identi ⁇ cal.
- Dictionary look up 50 uses a similarity threshold in order to define the predetermined extent of similarity.
- the method of Fig. 4 preferably includes the following processes: a. Initialization (process 310) - Determine the similarity threshold to be used.
- the string for which it is desired to find similar dictionary words is translated into each of the reduced alphabets of the layers in the similarity search index. For example, if it is desired to find dictionary words which are similar to ADCINS using the similarity index illustrated in part in Fig. 13, ADCINS is translated into (6) in the layer 1 alphabet, (4,2) in the layer 2 alphabet, and (1,3,2) in the layer 3 alphabet.
- b. Scan similarity search index (process 320).
- a similarity search index the structure of which is de ⁇ scribed above with reference to Fig. 13, is scanned recursively, layer by layer, in order to find all dic ⁇ tionary words that are similar (better than or equal to the similarity threshold) to the input string arriving from spellguess 30 or translator 40.
- Process 320 recursively invokes processes 330 - 370 in order to perform a depth-first search for similar nodes.
- the search employs a conventional distance-be- tween-strings computation method to compute the distance between each node and the input string represented in the reduced alphabet of the layer to which the node belongs.
- Parameters may be externally defined to deter ⁇ mine the relative costs of symbol insertion and symbol deletion, for each symbol in the dictionary and reduced alphabets and for symbol substitution, for each pair of symbols in the dictionary and reduced alphabets.
- the output of this step is a set of all dic ⁇ tionary words whose similarity to the "dictionary alpha ⁇ bet" representation of the input string is equal to or exceeds the similarity threshold defined in the initiali-
- Fig. 5 which illus ⁇ trates a preferred method of operation for the database lookup unit 90 and the candidate sorter unit 100 (Fig. 1).
- the database lookup unit 90 accepts as input eacn individual dictionary entry in the list of dictionary entries generated by dictionary lookup unit 50 along with its similarity score, also computed by dictionary lookup unit 50, and its entry context (last name, first name, etc. )
- the database lookup unit 90 operates for each database entry component. For example, in zip code direc ⁇ tory applications, the database lookup unit 90 operates after the state name is elicited from a user, then after the city's name is elicited and then after the street name is elicited. After each operation, the database lookup unit 90 updates the similarity vector 92 (Fig. 1) such that the similarity vector accumulates a similarity score for each database entry.
- the output of the database lookup 90 and the candidate sorter 100 is a list of a predetermined number of database entries which are most similar to the words supplied by the user thus far, sorted according to their similarity scores. This list if stored in the best candi ⁇ dates box 94 of Fig. 1.
- the method of Fig. 5 preferably includes the following processes: a. Initialization (process 380) - Uses concordance 72 to find the logical and physical location of all database entries that include the requested dictionary entry value in the appropriate context. For example, if the user supplies SMITH as a last name, all subscribers whose last name is similar to Smith will be located.
- initialization process 380 also clears the similarity vector 92 and the best candidates box 94.
- b. Database scan (process 390) - Scans all the database entries that were selected by the initializa ⁇ tion process 380. For each entry, the following 2 proc ⁇ esses are performed: i. Update database entry similarity score (process 400) - For each database entry, updates the similar ⁇ ity score by modifying the similarity score to reflect the degree of similarity of that database entry to the current word. For example, the word similarity score may be added to the appropriate similarity vector compo ⁇ nent. ii.
- Update best candidates box Determines whether or not the previously updated database entry has to be included in the best candidates list, based on the following criteria: if the database entry is already in the best candidates box, it is kept there with the new score updated if the best candidates box is not full, the database entry is added unconditionally if the best candidates box is full and the similarity score of the updated database entry is higher than the last (worst score) database entry in the candi ⁇ dates box, the last database entry is replaced with the new one.
- the similarity vector includes updated similarity scores and the contents of the best candidates box is determined.
- a candidate sorter 100 is invoked in order to resort the selected database entries in the best candi ⁇ dates box.
- the current ordering in the best candidates box is based on the sum of similarity scores of individu ⁇ al words where the similarity score of any word, if under a predetermined threshold, is considered to be zero.
- the candidates sorter computes the similarity of each zero- scoring word.
- the candidate sorter 100 prefera- bly performs the following processes: a. Best candidates scan (process 420) - Scans all the database entries in the best candidates box. b. Single candidate words scan (process 430) Scans all the words of a single database entry in the best candidates box. This process decides whether the similarity score of any one of the database entry strings may be improved, according to the following criteria:
- the process invokes 2 additional processes (improve similarity 440, and update best candidates 450), in order to refine the similarity scores, first at the word level and then at the database entry level.
- Improve similarity scores (process 440) - Re ⁇ initiates a minimum edit distance similarity check, similar to process 330 in Fig. 4, between a current database entry word in the best candidates box 94 and each word previously guessed by the spellguess unit 30, using a lower threshold measure than that originally employed by process 330. If the resulting similari ⁇ ty score differs from 0, the word similarity score and the database entry similarity score are both updated in process 450.
- d improve similarity scores (process 440) - Re ⁇ initiates a minimum edit distance similarity check, similar to process 330 in Fig. 4, between a current database entry word in the best candidates box 94 and each word previously guessed by the spellguess unit 30, using a lower threshold measure than that originally employed by process 330. If the resulting similari ⁇ ty score differs from 0, the word
- Update best candidates box (process 450) - Com ⁇ pute the total similarity score of each changed database entry in the best candidates box and the similarity vector.
- Candidate sorter (process 460) - Sorts the entire best candidates box according to the similarity score, using a suitable conventional sort mechanism, such as a heap.
- the output of the entire database lookup 90 and candidate sorter 100 is a list of all best candidates database entries, sorted according to their similarity scores.
- the conversation manager controls the conversation between the caller and the system , from the moment the caller gains access to the computer and until the conversation is finished.
- the actions taken by the conversation manager, including message providing action, depend on the contents of the best candidates box 94.
- the conversation manager 80 is preferably implemented as a state machine whose logic is defined by a decision table.
- the decision table is pref ⁇ erably completely external to the system to enable it to be easily modified.
- the conversation manager 80 includes three building blocks: the state, the conditio and the action.
- the conversation begins at the initial state, and each additional item of information acquired, leads the con ⁇ versation to a different state.
- the con ⁇ versation manager checks for the first condition that is fulfilled, and performs the action that is associated with the state and the condition. After that, the conver ⁇ sation manager proceeds to the next state. For example, one state may be "beginning of conversation", a condition may be "was a word entered by caller?", and an action may be "propose a selection list", or "search the database”.
- Transfer to operator (processes 520 and 550) - If the conversation manager comes to the conclusion that the conversation has to be transferred to a human opera ⁇ tor, it generates the appropriate routing instructions. This action may be useful, if for example, the conver ⁇ sation is too long, or the information provided by the caller does not identify a specific piece of informa ⁇ tion, or if the caller requests to be transferred to a human operator.
- New state (process 530) - The conversation manager proceeds to a new state, once it has finished handling the actions of previous state.
- Fig. 9 is a sim ⁇ plified block diagram of one embodiment of the apparatus of Fig. 1 in which the ambiguous input comprises tele ⁇ phone keypad input.
- the ambiguous input generator 10 of Fig. 1 includes a plural ⁇ ity of conventional telephone instruments 715 communicat ⁇ ing via a telephone network with database accessing unit 20 via a conventional DTMF (dual tone multiple frequency) detection module 720 which is operative to convert the keystroke information into conventional digital data signals such as ASCII data signals.
- DTMF dual tone multiple frequency
- the ambiguous input is keystroke information wherein each key corre ⁇ sponds to more than one letter of the English alphabet and may also correspond to the digit corresponding to the key.
- the data is ambiguous because every letter is en ⁇ tered by a single stroke on one of the telephone keypad buttons, and every button represents up to 3 different letters.
- button 2 represents either A or B or C
- button 3 represents either D or E or F.
- the sequence of digits "2"; "3” could represent any of the following letter combinations: AD, AE, AF, BD, BE, BF, CD, CE, CF, as well as combinations of letters and the numbers "2" and "3".
- n a string of length n can be interpreted in 3 to the power of n different ways.
- a sequence of 6 keystrokes corresponds to 729 alternative words.
- the names ARNOLD and BROOKE are keyed with the same keystroke sequence (276653). The same is true with names like BREEN and ARDEN (27336), or BUSH and BURG (2874).
- Fig. 10 is another embodiment of the apparatus of Fig. 9 in which the ambiguous input is ambiguous oral character input rather than ambiguous keystroke input.
- the ambiguous input generator 10 of Fig. 1 comprises a plurality of telephone instruments 740 or other speech input devices providing ambiguous speech input to the database accessing unit 20 via a probabilis ⁇ tic spoken character recognition device 750.
- the user is prompted to provide oral input by spelling the desired words.
- the conversation manager may prompt the user to "SPELL THE STREET NAME YOU ARE LOOKING FOR.”
- the input provided by device 750 is ambiguous since the probabilistic spoken character recognition box 750 outputs several possible characters, along with their relative probabilities, for every input spoken character. For example, a user's attempt to utter a "D", may be interpreted by the probabilistic spoken character recog ⁇ nition box 750, as "B” with probability 0.5, "D” with probability 0.3, and "T” with probability 0.2. A subse ⁇ quent utterance of the letter "D”, may yield different output (letters and probabilities).
- Fig. 11 is another embodiment of the apparatus of Fig. 1 in which the ambiguous input comprises ambiguous spoken phoneme input rather than ambiguous oral character input.
- the probabilistic spoken phoneme recognition device 770 is similar to conventional speech recognition units, except that the output is phonemes with associated probabilities. In other words, the final decision stage in which probabilistic input, as to the contents of the speech, is reduced to a single possibility, typically by selecting the most probable speech content, is omitted.
- a suitable design for a probabilistic spoken phoneme recog ⁇ nition device is described in Rabiner L., Juang B., Fundamentals of Speech Recognition. Prentice-Hall, 1993, pp. 43 - 52.
- DAVE may be phonetical ⁇ ly broken into the following phonemes: D, Ey and V (when using a certain phonetic alphabet).
- the word "DAVE” When the word "DAVE" is input to a probabilis ⁇ tic spoken phoneme recognition box 770, it may recognize every phoneme in several ways, with a certain probability applied to every phoneme. For instance, the first phoneme may be expressed as “(D, 0.4), or (B, 0.4), or (T, 0.2)”; the second phoneme may be expressed as "(Ey, 0.6) or (Eh, 0.3), or (Ay, 0.1)”; and the third phoneme may be expressed as "(V, 0.8), or (B, 0.1), or (F, 0.1)".
- the software components of the present invention may, if desired, be implemented in ROM (read-only memory) form.
- the software components may, generally, be implemented in hardware, if desired, using conventional techniques.
- the term "probabilistic string” refers to a sequence S over an alphabet A (such as the English alphabet, any non- English alphabet, the phonetic alphabet or any other set of characters) of n elements where each element S is a i finite non-empty set of (character, probability) pairs.
- keyboard refers to an input device including a plurality of input buttons or keys each of which, separately or in combination with other input buttons or keys, represents one or more characters. For example, each key on a conventional telephone keypad represents three, two or only one characters.
- the term "erroneous user input” refers to a database accessing request provided by a user which comprises an inaccurate production of a word in the database.
- the inaccuracy may comprise one, some or all of the following types of inaccuracy: spelling inaccuracy, pronunciation inaccuracy or factual inaccuracy.
- pelling inaccuracy refers to a cognitive or typographical error which results in an incorrect spelling of a database word.
- a user wishing to access a street name BROADWAY may key in or spell out BROAWDAY or DROADWAY.
- a user wishing to access a locality PHILADELPHIA may key in or spell out FILADELFIA.
- pronunciation inaccuracy refers to an incorrect oral rendition of a database word. For example, a user wishing to access the name MARIE may state this request as "MARY".
- JOHN REMINGTON a user wish ⁇ ing to access JOHN REMINGTON may request JACK REMINGTON.
- the user accessing the system shown and described herein can be either a human user or a computerized system.
- the ambiguous input may, alternatively, comprise the output of an OCR (optical character recognition) device such as a handwriting recognition device.
- OCR optical character recognition
Abstract
A database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including a sequence of representations of possibly erroneous user inputs (10), the system including a similar word finder operative (30), for at least one interpretation of each representation, to find at least one database word which is similar to that interpretation, and a database entry evaluator operative (50), for each database word found by the similar word finder, to assign similarity values for relevant entries in the database, the values representing the degree of similarity between each database entry and the request.
Description
1 A DATABASE ACCESSING SYSTEM
FIELD OF THE INVENTION
The present invention relates to methods and apparatus for accessing databases generally and for accessing databases via the telephone, in particular.
BACKGROUND OF THE INVENTION
Publications describing conventional methods and systems for accessing databases via the telephone and publications on subjects related to these technolo¬ gies, such as speech recognition publications, include the following:
U.S. Patent 4,608,460 to Carter, entitled "Comprehensive Automatic Directory Assistance Apparatus and Method Thereof;
WO Patent 88/02965 to Daudelin, entitled "Directory Assistance Call Processing and non-Supervisory Signal Monitoring Arrangements";
US Patent 4,650,927 to James, entitled "Proces¬ sor Assisted Communication System using Tone-Generating Telephones";
US Patent 5,255,310, to Kim, entitled "Method of Approximately Matching an Input Character String with a Key Word and Vocally Outputting Data";
US Patent 4,674,112 to Kondraske, entitled "Character pattern Recognition and Communications Appara¬ tus";
US Patent 5,214,689 to O'Sullivan, entitled "Interactive Transit Information System";
WO Patent 89/00793 to Padden, entitled "Direc¬ tory Assistance Systems";
US Patent 4,979,206 to Padden, entitled "Direc¬ tory Assistance Systems";
US Patent 5,131,045 to Roth, entitled "Audio- Augmented Data Keying";
US Patent 5,052,038 to Shepard, entitled "Apparatus and Method for Obtaining information in a Wide-Area Telephone System with Digital Data Transmission Between a Local Exchange and an Information Storage Site";
US Patent 3,928,724 to Byram, entitled "Voice- Actuated Telephone Directory-Assistance System";
Church K.W., Phonological Parsing in Speech Recognition, Kluwer Academic Publ, 1987;
Colangelo R.D., "Directory Assistance for the Visually Impaired", SouthEastcon 89, Vol 3, pp. 1168- 1171, 1989;
Fast L., Ballantine R. , "Dialing a Name: Alpha¬ betic Entry Through a Telephone Keypad", SIGCHI, Vol 20(2) pp. 34, 1988;
Glowacz D., "RBOCs Vocalize Operator Services", Telephone Engineer & Management, Vol 95(7) pp. 25-30, 1991;
Holmgren J.E., "Toward bell System Applications of Automatic Speech Recognition", The Bell System Techni¬ cal Journal, Vol 62(6) pp. 1865-1881, 1983;
Jack M.A., "Design and Evaluation of Automated Telephone Services", IEE Colloquium on Telecommunica¬ tions, pp. 61-64, 1992;
Klatt D.H., "Review of Text-to-Speech Conver¬ sion for English", Journal of the Acoustical Society of America, Vol 82(3) pp. 737-793, 1987;
Lennig M. , "Putting Speech Recognition to Work in the Telephone Network", Computer, Vol 23(8) pp. 35-41, 1990;
Myers C.S., Rabiner L.R., "An Automated Direc¬ tory Listing Retrieval System based on Recognition of Connected Letter Strings", Journal of the Acoustical Society of America, Vol 71(3), pp. 716-727, 1982;
Ohyama Minoru, "A Knowledge-Based Directory
Assistance System", Future Generation Computer Systems, Vol 5(1) pp. 109-117, 1989;
Pelton G.E., "Designing the Telephone Interface for Voice Processing Applications", Speech Technology, Vol 5(1) pp. 18-21, 1989;
Schalk T.B., "Automating Operator-Assisted Calls Using Speech Recognition", Eurospeech 89, Vol 1 pp. 441-444, 1989;
Veith R.H., "Information Retrieval via Tele¬ phone Keypad and Voice: Experiences with Telephone Direc¬ tories", National Online Meeting Proceedings, Vol 1 pp. 443-451, 1990;
Veith R.H., "Operator Assisted Services with Automated Responses", Audiotex, Vol 1 pp. 19-28, 1986;
Rabiner L., Juang B., Fundamentals of Speech Recognition, Prentice-Hall, 1993, pp. 43 - 52;
Hall, P. A. "Approximate string matching", Computing surveys, Vol. 12(4), pp. 381 -402, 1980;
Salomaa, A., "Computation and Automatation", Cambridge University Press, 1985, pp. 44 - 55;
Tremblay, J. P.; Sorenson, P. G., "An introduc¬ tion to data structures with applications", McGraw-Hill, 1976, pp. 316 - 318;
Vitale, T. "An algorithm for high accuracy name pronunciation by parametric speech synthesizer", Computa¬ tional Linguistics, Vol. 17(3), pp. 257 - 276, 1991;
Lorenz, A., Stochastic Automata Constructive Theory, John Wiley and Sons, 1974; and
Aldefeld, B. "Automated directory listing retrieval system based on isolated word recognition". Proceedings of the IEEE, Vol. 68(11), pp. 1364 - 1379, 1980.
The disclosures of all above references and publications cited therein are hereby incorporated by reference.
SUMMARY OF THE INVENTION
The present invention seeks to provide an improved system for accessing a computerized database from a location which is not in computer communication with the database.
There is thus provided in accordance with a preferred embodiment of the present invention a database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including a sequence of representations of possibly erroneous user inputs, the system including a similar word finder opera¬ tive, for at least one interpretation of each representa¬ tion, to find at least one database word which is at least similar to that interpretation, and a database entry evaluator operative, for each database word found by the similar word finder, to assign similarity values for relevant entries in the database, the values repre¬ senting the degree of similarity between each database entry and the request.
Further in accordance with a preferred embodi¬ ment of the present invention the request includes a sequence of unambiguous representations of possibly erroneous user inputs.
Still further in accordance with a preferred embodiment of the present invention the relevant entries include all database entries having the database word in an individual context.
Additionally in accordance with a preferred embodiment of the present invention the request includes a sequence of ambiguous representations of possibly erroneous user inputs and also including an ambiguous input evaluator operative to identify at least one inter-
pretation of each ambiguous representation and to provide the interpretation to the similar word finder.
Further in accordance with a preferred embodi¬ ment of the present invention the ambiguous input evalua¬ tor includes a database analyzer operative to extract database characteristics from the database.
Still further in accordance with a preferred embodiment of the present invention the system also includes a session manager receiving similarity values from the database entry evaluator and operative to inter¬ actively provide a user sensible message based on the similarity values.
Additionally in accordance with a preferred embodiment of the present invention the session manager is operative to prompt the user to provide the request by orally spelling at least one database word.
Further in accordance with a preferred embodi¬ ment of the present invention the session manager is operative to prompt the user to provide the request by pronouncing at least one database word.
Still further in accordance with a preferred embodiment of the present invention the session manager is operative to prompt the user to provide the request by keying at least one database word on the keypad.
There is additionally provided in accordance with a preferred embodiment of the present invention a database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including at least one ambiguous representation of at least one possibly erroneous user input, the system including an ambiguous input evaluator operative to identify at least one interpretation of each ambiguous representation, a similar word finder operative, for at least one interpretation of each representation, to find at least one database word which is at least similar to that interpretation, and a database entry evaluator
c
operative, for each database word found by the similar word finder, to assign similarity values for relevant entries in the database, the values representing the degree of similarity between each database entry and the request.
Further in accordance with a preferred embodi¬ ment of the present invention the relevant entries in¬ clude all database entries having the database word in an individual context.
Still further in accordance with a preferred embodiment of the present invention the ambiguous input evaluator includes a database analyzer operative to extract database characteristics from the database.
Yet further in accordance with a preferred embodiment of the present invention the system also includes a session manager receiving similarity values from the database entry evaluator and operative to inter¬ actively provide a user sensible message based on the similarity values.
Additionally in accordance with a preferred embodiment of the present invention the session manager is operative to prompt the user to provide the request by orally spelling at least one database word.
Further in accordance with a preferred embodi¬ ment of the present invention the session manager is operative to prompt the user to provide the request by pronouncing at least one database word.
Yet further in accordance with a preferred embodiment of the present invention the session manager is operative to prompt the user to provide the request by keying at least one database word on the keypad.
There is still further provided in accordance with a preferred embodiment of the present invention a database accessing method for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including a sequence of representations of possibly
erroneous user inputs, the method including finding, for at least one interpretation of each representation, at least one database word which is at least similar to that interpretation, and assigning similarity values for relevant entries in the database, for each database word found, the values representing the degree of similarity between each database entry and the request.
There is provided in accordance with a pre¬ ferred embodiment of the present invention a database accessing method for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request including at least one ambiguous representation of at least one possibly erroneous user input, the method including identifying at least one interpretation of each ambiguous representation, finding, for at least one interpretation of each representation, at least one database word which is at least similar to that interpretation, and for each database word found by the similar word finder, assigning similarity values for relevant entries in the database, the values representing the degree of similarity between each database entry and the request.
In the present specification and claims, many examples from zip code database applications are provided in which address information is provided and, in re¬ sponse, the zip code of the address is accessed. However, the term "database" as used herein is intended to include any type of database including but not limited to data¬ bases including financial and commercial data, databases including administrative data such as personnel data, transportation vehicle flight schedule databases, databases including technological data such as weather data, and databases including population data such as zipcode directories, classified and unclassified tele¬ phone directories.
BRIEF DESCRIPTION OF THE DRAWINGS
The present invention will be understood and appreciated from the following detailed description, taken in conjunction with the drawings in which:
Fig. 1 is a high level functional block diagram of a database access system constructed and operative in accordance with a preferred embodiment of the present invention;
Fig. 2 is a flowchart illustrating a preferred method of operation for the spellguess unit of Fig. 1;
Fig. 3 is a flowchart illustrating a preferred method for generating the phonetic translator of Fig. 1;
Fig. 4 is a flowchart illustrating a preferred method of operation for the dictionary lookup unit of Fig. 1;
Fig. 5 is a flowchart illustrating a preferred method of operation for database lookup and candidate sorter units of Fig. 1;
Fig. 6 is a flowchart illustrating a preferred method of operation for the conversation manager of Fig.
1;
Fig. 7 is a flowchart illustrating a preferred method for loading the dictionary, similarity index, numeric index and trigram stochastic finite state automa¬ ton into memory, as well as for preparing the concord¬ ance;
Fig. 8 is a flowchart illustrating a preferred method for loading the database into memory and for updating the concordance prepared by the method of Fig. 7; Fig. 9 is a simplified block diagram of one implementation of the apparatus of Fig. 1 in which the ambiguous input comprises telephone keypad input;
Fig. 10 is a simplified block diagram of one implementation of the apparatus of Fig. 1 in which the ambiguous input comprises output of a probabilistic spoken character recognition device;
Fig. 11 is a simplified block diagram of one implementation of the apparatus of Fig. 1 in which the ambiguous input comprises output of a probabilistic spoken phoneme recognition device;
Fig. 12 is a diagram of the computer memory data structures serving the apparatus of Fig. 1, includ¬ ing a dictionary, concordance and database storing typi¬ cal data and the internal connections between these structures;
Fig. 13 is a diagram of a similarity search index, storing typical data, which is useful in perform¬ ing the dictionary look-up method of Fig. 4.
DETAILED DESCRIPTION OF A PREFERRED EMBODIMENT
Reference is now made to Fig. 1 which is a high level functional block diagram of a database acce&ό system constructed and operative in accordance with a preferred embodiment of the present invention.
The database access system includes an ambigu¬ ous input generator 10 and a database accessing unit 20 accessing a database on the basis of ambiguous input received from generator 10. The ambiguous input generator 10 provides ambiguous input such as ambiguous keystroke input or ambiguous speech input such as ambiguous oral character input or ambiguous oral phoneme input, as illustrated below with reference to Figs. 9 - 11 respec¬ tively.
The output of the ambiguous input generator 10 is received by a spellguess unit 30 which is operative to select at least one string of symbols to which the ambig¬ uous input might correspond, and preferably a plurality of strings of symbols, each associated with a probabili¬ ty. Each string of symbols may, for example, comprise a string of characters such as English language letters or may comprise a string of phonemes suitable for represent¬ ing the dictionary alphabet. For example, the output of the spellguess unit may comprise the N most probable strings and their probabilities, where N is any suitable number such as 20.
A preferred method of operation for spellguess unit 30 is described below with reference to Fig. 2.
A phonetic translator 40 receives the possible strings generated by spellguess unit 30 and translates each of them into phonetic form. A preferred method for generating translator 40 is described below with refer¬ ence to Fig. 3. If translator 40 is generated by the method of Fig. 3, the complexity of phonetic translator 40 is independent of the number of phonetic rules as
described below with reference to Fig. 3. Alternatively, however, conventional phonetic translation methods may be employed such as the methods described in the following publication:
Vitale, T. "An algorithm for high accuracy name pronunciation by parametric speech synthesizer", Computa¬ tional Linguistics, Vol. 17(3), pp. 257 - 276, 1991.
In the illustrated embodiment, the phonetic translator converts strings in the English alphabet to strings in a phonetic alphabet such as the DECvoice phonetic alphabet developed by Digital Equipment Corpora¬ tion. More generally, however, the phonetic translator may be operative to translate strings from any first alphabet into any second alphabet. The DECvoice phonetic alphabet is described in the following publication:
Klatt, D. H. , "Review of text-to-speech conver¬ sation for English", Journal of the Acoustical Society of America, Vol. 82(3), p. 769.
Alternatively, the phonetic translator may be omitted in which case the access to the dictionary may be based directly on the guessed strings rather than on the translation of the guessed strings into phonetic form.
A dictionary look up unit 50 receives each phonetic or alphabetical string and conducts a similarity search within a dictionary 60.
A preferred method of operation for dictionary lookup unit 50 is described below with reference to Fig. 4.
The dictionary 60 comprises an alphabetical or phonetically ordered or otherwise ordered list of some or all words or strings appearing in the database 70 which is to be accessed. A concordance 72 associates the dictionary 60 with the database 70. A similarity search index 74 allows the dictionary 60 to be searched for entries which are similar to a desired user-selected string. A numeric index 76 is provided if a telephone keypad is used which extracts from the dictionary all
entries having a given numerical code. A trigram stochas¬ tic finite state automaton 78 stores trigrams and their probabilities and is accessed by spellguess unit 30, as described in more detail below.
Dictionary 60, database 70, concordance 71, numeric index 76 and trigram stochastic finite state automaton 78 may be stored in any suitable electronic medium such as a hard disc or RAM. A preferred method for loading the dictionary, similarity index, numeric index and trigram automaton into memory, as well as for prepar¬ ing the concordance, is described below with reference to Fig. 7. A preferred method for loading the database into memory and for updating the concordance prepared by the method of Fig. 7 is described below with reference to Fig. 8.
Alternatively, the dictionary 60 may comprise the database itself, particularly if the database is small and/or if the database does not include many dupli¬ cated strings.
The output of the dictionary lookup unit 50 typically comprises, for each string received from unit 30 or 40: a. A possibly empty list of dictionary entries whose similarity to the string is better than or equal to a predetermined similarity threshold; and b. A similarity score representing the degree of similarity between that entry and the corresponding string.
EXAMPLE: The operation of spellguess unit 30, phonetic translator 40 and dictionary lookup 50 for the following ambiguous telephone keypad input: 228437463 is now described.
Spellguess unit 30 guesses various spellings, among which, for the purposes of this example, the word CATHERINE is assumed to appear. This word is input to the phonetic translator 40, which may translate it into (K Ae Th Eh R lh N) . The phonetically translated string is
sent to the dictionary look up unit 50, which may produce the following list of dictionary words and corresponding similarity scores:
(KATHERINE, 100), (CATHERINE, 100), (CATHYRINE, 95),
(KATHERYN, 92), (KATHRINE, 92), (KATHRYN, 85), (CATHRYN,
85).
According to one embodiment of the present invention, the database 70 is now accessed by a database look up unit 90 so as to identify, for each dictionary entry, all database entries which include that dictionary entry.
Preferably, however, the database entries identified by database lookup 90 for each dictionary entry include only those database entries which include the dictionary entry in the correct context. For example, in zip code directory applications, a conversation manag¬ er 80 may direct the user to enter the name of the city of interest and subsequently to enter the name of the street of interest. In this case, the database entries identified for a dictionary entry comprising a putative street name comprise only those database entries having that street name.
Database lookup unit 90 generates and continu¬ ally updates, as the call or database access session progresses, a similarity vector 92. The similarity vector 92 has a number of vector components which correspond to the number of database entries. Each similarity vector component stores a similarity score of the corresponding database entry. This similarity score is provided by the dictionary lookup 50 for each user-supplied word as the database access session proceeds. The database lookup 90 accumulates these similarity scores in the appropriate locations within the similarity vector.
After each word is supplied by the user, and the similarity vector is updated by the database lookup 90, or concurrently with the updating process, a candi¬ date sorter 100 extracts a predetermined number of the
highest scoring similarity vector components and stores these components in a best candidates box 94. The simi¬ larity vector 92 and best candidates box 94 are prefera¬ bly stored in computer memory and are initialized at the beginning of each database access session.
A preferred method of operation for units 90 and 100 is described below with reference to Fig. 5.
The conversation manager 80 receives the output of unit 90 or 100 and determines a course of action. The considerations of the conversation manager may, for example, include the following: a. If one database entry has a similarity score which significantly exceeds the similarity score of all other candidates and which exceeds a threshold similarity score, this entry may be deemed correct and therefore is presented to the user for confirmation. b. Typically, each entry in database 70 comprises a plurality of words or strings. For example, each entry in a zipcode directory typically includes some or all of the following words: state name, street name, street type, house number, locality. If this is the case, and processing of one string (such as "locality") does not provide a conclusive indication of the correct database entry, the conversation manager may decide to prompt the user to enter information identifying another word within the desired database entry (such as "street name"). c. If a small number of entries have similar or identical scores, the conversation manager may decide to present all of these to the user and prompt the user to select one of them about which s/he wishes to receive more information. d. If a large number of entries have similar or identical scores and all strings within the entry have been processed, the conversation manager may wish to refer the user to the services of a human operator. In this case, the conversation manager 80 provides a suit¬ able routing instruction to conventional telephone equip-
ment which routes the user to a human operator.
A preferred method of operation for conversa¬ tion manager 80 is described below with reference to Fig. 6.
The output of the conversation manager is a suitable message to the user which is received by a text- to-speech voice synthesizer 110, such as DECvoice, com¬ mercially available from Digital Equipment Corporation, Maynard, MA, USA or Text Assist, commercially available from Creative Labs, Inc., Milpitas, CA, USA. Alternative¬ ly or in addition, the output of the conversation manager may be received by a pre-recorded voice player 115 which plays a selected message from among a plurality of re¬ corded messages. Examples of a commercially available voice player 115 include the DECvoice or Sound Blaster, commercially available from Creative Labs, Inc., Milpi¬ tas, CA, USA. The text-to-speech voice synthesizer 110 and/or the voice player 115 provide an audible message which is transmitted over the telephone line to the user.
As described briefly above, the apparatus of Fig. 1 utilizes the following data structures in order to perform a similarity dictionary lookup, rather than an exact match lookup, and to support spell guess functions: a. Word dictionary 60 - A list of entries which includes a single entry for each uniquely spelled database word for which direct access is required.
For example, if the database is a zip code directory, the words in the database which are to be accessed directly may include only the localities or alternatively, the state as well as the street name and locality. The dictionary includes a count for every word appearance in the database in the specific context, and a pointer to the concordance. For example, if the word MARYLAND appears in the database as a state name, a city name and a street name, the dictionary will contain only a single entry for the word MARYLAND, with separate
counters for the number of times the word appears as state name, city name and street name, and a single pointer to the concordance, which points to the first occurrence of the word MARYLAND in the concordance, b. Concordance 72 - A list of entries including a single entry for each occurrence of each dictionary word in the database.Concordance entries relating to the same dictionary word are always arranged consecutively and sorted, within the dictionary word, according to the entry context (for example, appearances of the dictionary word in a locality context, followed by ap¬ pearances of the dictionary word in a street name con¬ text). Each database entry is pointed to by all concord¬ ance entries corresponding to that database entry.
For example, the following database entry:
573 MARYLAND, BUFFALO, NY 14201 will be pointed to by a corresponding MARYLAND entry in the concordance, a corresponding BUFFALO entry in the concordance and a corresponding NY entry in the concord¬ ance.
If database entries residing in consecutive addresses correspond to a plurality of concordance en¬ tries residing in consecutive addresses, the plurality of concordance entries may be replaced by a single entry pointing to the range of database entries.
The associations between the database, diction¬ ary and concordance are illustrated in the diagram of Fig. 12. c. Similarity search index 74 - An index into the word dictionary that enables access to the word dictionary by word similarity rather than by exact match of words.
The similarity search index is preferably organized as an N-ary tree in which each node points to its left son and its right brother. An N-ary tree is a well-known data structure which is described in Tremblay, J. P.; Sorenson, P. G., "An introduction to data struc-
tures with applications", McGraw-Hill, 1976, pp. 316 318.
The similarity search index is built in several tree layers, such as 7 or 9 layers. Each layer corre¬ sponds to a different "reduced alphabet" of the diction¬ ary alphabet, which may be either graphic or phonetic. One example of a reduced alphabet for the English lan¬ guage alphabet is an alphabet including only two symbols, c for consonants and v for vowels. The correspondence between the English alphabet and the reduced alphabet is that all consonants are replaced by c and all vowels are replaced by v. More generally, a reduced alphabet is an alphabet that contains less symbols than an original alphabet and complies with the following criteria:
1) each symbol in the reduced alphabet represents a group of symbols of the original alphabet;
2) a symbol of the original alphabet appears in only one group; and
3) the union of all groups comprises the complete origi¬ nal alphabet.
Each layer of the similarity search index contains the same dictionary words but in a different reduced alphabet.
Each word in the similarity search index is represented in vector format with a reduced alphabet. A word in vector format is represented by a vector V of length L, where L is the number of symbols in the alpha¬ bet. Each element Vi of the vector contains the number of occurrences of the ith alphabetic symbol in the word. For example, given an alphabet {a,b,c} and a word "abbac", the vector representation is "2,2,1". To continue the above example of the {c,v} alphabet, the words BLAKE and GRACE will both be expressed by the pair "2,3". Each layer of the similarity search index is built with a different reduced alphabet adhering to the above guide¬ lines.
Fig. 13 illustrates an example of a "grapheme
based" similarity search index for an English language dictionary. It is appreciated that, alternatively, a phonetic-based similarity search index may be employed. The index of Fig. 13 is an N-ary tree which includes the following three layers:
Layer 1 is based on a single symbol reduced alphabet which transforms each English language word into its length.
Layer 2 is based on a two symbol alphabet, which transforms each English word into an ordered pair (a,b) where a is the number of consonants in the word and b is the number of vowels.
Layer 3 is based on three symbols, which trans¬ forms each English word into an ordered triplet (a,b,c) where a is the number of letters from among the set (P, F, B, V, W, T, D) in the word, b is the number of letters from among the remaining consonants in the word and c is the number of vowels in the word.
Each dictionary word is mapped to one node in each index layer. For example, the word ADKINS appearing in Fig. 13 is mapped into node (6) in layer 1 since ADKINS contains 6 letters, into node (4,2) in layer 2 , since ADKINS contains 4 consonants and 2 vowels, and into node (1,3,2) in layer 3 since only the letter D belongs to the set (P, F, B, V, W, T, D), and there are 3 other consonants, and 2 vowels. d. Numeric index 76 - This index is organized as a balanced tree, and is employed when the ambiguous input generator is a keypad in which one number is used to represent a plurality of letters, such as a tele¬ phone keypad. The numeric index uses the ambiguous numer¬ ic input value keyed in by the user of the keypad to point directly to all dictionary entries having that numeric input value. e. Trigram stochastic finite state automaton 78
An automaton that stores all trigrams appearing in the dictionary along with their probabilities which are
computed based on dictionary appearance frequencies. The trigram stochastic finite state automaton is employed by spell guess unit 30 to anticipate the probable spelling of any given ambiguous string.
Reference is now made to Fig. 7 which illus¬ trates a preferred method for loading the dictionary, similarity index, numeric index and trigram automaton into computer memory, as well as for preparing the con¬ cordance in the computer memory.
The method of Fig. 7 preferably includes the following processes: a. Allocate data structures (process 570) - This process creates the initial structures for the word dictionary, the concordance, similarity search index, the numeric index and the trigram stochastic finite state automaton. b. Scan dictionary words (process 580) - This loop reads all dictionary words from a dictionary organized as a sequential disk file. For each word the following proc¬ esses are performed: i. Update word dictionary (process 590) - This process copies the word into the dictionary. ii. Update similarity search index (process 600) - The word is translated into the reduced alphabet of every layer of the index. All index layers are searched, from the highest to the lowest level, and any missing words are added. iii. Update numeric index (process 610) - If the numeric code of the word does not exist in the index, it is added. Once the numeric code of the word does exist, a pointer is added to it which points at the newly entered word. iv. Extract trigrams (process 620) - All trigrams contained in the new word are stored in a temporary list. Each time a trigram on the list appears, an associated frequency value is increment- ed. The trigram list including frequency values are employed to construct the automaton in process 640, once all diction-
20
ary words have been processed by loop 580.
Prepare concordance entries (process 630) - Each dictionary word occupies one or more entries in the con¬ cordance. The concordance entries are created at this time, but are not associated with database entries pointers until the database itself is loaded into memory, c. Create trigram stochastic finite state automa¬ ton (process 640) - Based on the trigram information (population and frequencies) accumulated by process 620, a stochastic finite state automaton is built. This auto¬ maton will be later used by spellguess unit 30 to guess the spelling of the ambiguous input string. A stochastic finite state automaton is described in the above-refer¬ enced publication by Lorenz.
Reference is now made to Fig. 8 which illus¬ trates a preferred method for loading the database into memory and for updating the concordance prepared by the method of Fig. 7.
The method of Fig. 8 preferably includes the following processes: a. Allocate database (process 650) - Prepares the memory required to hold the entire database. b. Database scan loop 660 - Reads a plurality of database entries stored in sequential format and residing on magnetic media, such as hard disk drive. For every database entry, the following processes are performed:
I. Place database entry in memory database (proc¬ ess 670) - Finds the next available place in memory database, places the database entry there, sets a direct pointer from every database word to the dictionary, and keeps the direct memory address in order to allow further direct addressing to the entry (from the concordance).
II. Inner database entry word scan loop (process 680) - Scans all the database entry words that are de¬ fined as requiring direct access through the dictionary and concordance. For every such word, the process per¬ forms the following two processes:
i. Locate word in dictionary (process 690) Since the dictionary is loaded before the database, and since every word requiring direct access has to be present in the dictionary, this process locates the word entry in the dictionary, in order to locate the right entry in the concordance, that corresponds to this data¬ base entry. ii. Update concordance (process 700) - Adds a pointer from the concordance entry to the database entry. This pointer establishes a link between a dictionary word and the corresponding database entries.
Reference is now made to Fig. 2 which illus¬ trates a preferred method of operation for spellguess unit 30. As described above, spellguess unit 30 is preferably operative to select a plurality of strings of characters to which ambiguous input might correspond, each associated with a probability. In the illustrated embodiment, the output of the spellguess unit is the N most probable strings and their probabilities, where N is any suitable number such as 20.
The method of Fig. 2 preferably includes the following processes: a. An initialization process 140 in which an initial string is constructed. The process then invokes a recursive string lengthening process 144, until all possible spelling alternatives have been checked. b. A recursive string lengthening process 144 including blocks 150 - 220 which generates a list of the N most probable strings and their probabilities, based on pre-known frequencies of various "trigrams" or 3-letter strings in the dictionary. Process 144 receives as input a partial string of length L whose probability has al¬ ready been computed and computes the probability of at least one string of length L+l which includes the input partial string. If the length of the partial string is less than 3 (if trigram frequencies are employed), the partial string is taken to include two "dummy" characters
such as two # ' s . c. Output process 230 — the N most probable strings are copied onto a digital storage medium such as computer memory. Preferably, the N most probable strings are retained as a heap sorted by their probabilities.
For example, suppose that the numeric string "2874" is a keypad input to spellguess unit 30, which is intended by the caller to represent the name "BUSH". Spellguess unit 30 uses the method of Fig. 2 in order to produce a list of the most probable strings represented by the numeric string "2874", along with their probabil¬ ities.
An example of such a list is as follows:(BURG, 0.1611), (BURI, 0.1259), (CURG, 0.1118), (BUSH, 0.1086), (CUSH, 0.1018), (CURI, 0.0874).
As described in detail below, the probability of an individual candidate string is based on the proba¬ bilities of each trigram included in the string. For example, the probability assigned to BURG might be rough¬ ly the product of the known frequencies of occurrence of each of the following trigrams: BUR, URG.
The blocks of the recursive spellfinding proc¬ ess are now described in detail:
Block 150: "Minimum probability" is typically 0 if less than N strings have been accumulated. If N strings have already been accumulated, "minimum probabil¬ ity" is equal to the probability of the least probable of the N strings.
If the probability of the input partial string is less than "minimum probability", the process back¬ tracks to the previous recursive level because the proba¬ bility of the input partial string will only decline further as its length increases.
Block 160: If the partial string is, in fact, a "final" string, i.e. a string of the length which the user entered, then it is entered onto a list, typically organized as a heap, of most probable strings (block
190). Otherwise, the process advances directly to block 200.
If the list already contains N strings (block 170), the lowest probability string is discarded (block 180) before the new string is inserted (block 190).
At this point, the string is lengthened by considering the possible characters which might fill the (L+l)'th position in the string. For example, if keypad input is employed and the user punched the #2 key for the L+l'th position, the candidate characters are A, B and C. For each candidate character, blocks 210 and 220 are performed.
Block 210: The candidate character is concate¬ nated with the partial string and the probability of the new, longer partial string is defined by generating the product of the probability of the previous partial string and the probability of the trigram consisting of the last three characters of the new, longer partial string.
Block 220: The next recursive level of the string lengthening process begins, thereby to add yet another character to the current partial string.
Reference is now made to Fig. 3 which illus¬ trates a preferred method for generating phonetic trans¬ lator unit 40. As described above, phonetic translator unit 40 is preferably operative to convert strings from one alphabet (such as English), into another alphabet (such as the DECvoice phonetic alphabet). The complex¬ ity of the phonetic translator is n, where n is the length of the string to be converted, regardless of the number of phonetic translation rules. The method of Fig. 3 preferably includes the following processes:
Process 260: Load translation rules: read an external file that describes phonetic translation rules for translating between the alphabets employed. Prefera¬ bly, the syntax of the external file is checked. The structure of a rule is as follows: grapheme->phoneme/"preceding substring"_"following
substring"
The above rule instructs the phonetic transla¬ tor to translate a certain grapheme to a certain phoneme if the grapheme is preceded by a certain "preceding substring" and is followed by a certain "following sub¬ string". Any of the phoneme, preceding substring and following substring can be empty.
Following are some examples of phonetic trans¬ lation rules: BB->B double Bs sound like a single B
(BUBBLE) PH->F P followed by H always sound like F
(PHONE) 0->Uh/W_L O preceded by W and followed by L, sounds like the phoneme Uh (WOLF) Y->Ay/c_cE Y preceded by any consonant and fol¬ lowed by a consonant that is followed by E, sounds like (BYTE) A non-deterministic automaton is now generated. Processes 270, 272 and 274 create a single non-determin¬ istic finite state automaton (NDFSA) from all the pho¬ netic translation rules loaded in process 260. Each of these processes is now described in detail:
Process 270: Build states SO, SI and S2 where: SO is the initial state for all rules that have 2 charac¬ ters in their "preceding substring" SI is the initial state for all rules that have 1 character in their "preceding substring", and S2 is the initial state for all rules that do not have a "preceding substring".
Additional states may be defined similarly if it is desired to accommodate preceding strings which are more than 2 characters long.
Process 272: Add unconditional transitions from
50 to SI and from SI to S2 such that the entire grapheme alphabet causes the automaton to proceed from state SO to
51 or from SI to S2.
Process 274: Build a subautomaton for each
phonetic rule. The subautomaton for an individual phonet¬ ic rule has n states, where n is the total number of characters in the "preceding substring", the grapheme and the "following substring". The last, n'th, state is also termed herein the Accepting State. Add transitions from each state to the next state which are respectively determined by the characters of the preceding substring, grapheme and following substring. The output generated by the subautomaton is the phoneme which the rule in ques¬ tion associates with the preceding substring, grapheme and following substring.
For example, the rule 0->Uh/W_L, has 3 states, such as for example (S10,S11,S12) and 3 transitions: W as preceding substring causes a transition from the initial state to state S10, 0 as grapheme causes a transition from S10 to Sll, and L as following substring causes a transition from Sll to the accepting state S12. The output of the subautomaton for the above sample rule is the phoneme (Uh) . Since this rule has one character, W, in the "preceding substring", then state SI serves as its initial state. Therefore, the subautomaton for this rule comprises the following transitions: (S1,W,S10), (S10,0,S11), (S11,L,S12).
Process 280: Convert the non-deterministic automaton (NDFSA) constructed in processes 270-274 into a deterministic finite state automaton (DFSA) including all the phonetic translation rules. The conversion process may be based on well known methods, such as the conver¬ sion method described in Salomaa, A. , Computation and automata, Cambridge University Press, 1985, pp. 44 - 55.
Process 290 — Build states matrix: Convert the DFSA created by process 280 into a matrix, where the rows correspond respectively to the states in the DFSA and the columns correspond respectively to the possible inputs i.e. to the grapheme alphabet. The contents of each matrix cell includes: the next state, whether the state is an accepting state or not, and the phoneme
26
generated by an accepting state.
Once the phonetic translator 40 has been gener¬ ated as described above, the phonetic translator is operative to translate a graphic word into a phonetic word as follows:
Given a graphic word, such as a word in Eng¬ lish, apply the states matrix created in process 290 thereon, and output a phonetic word, which is the phonet¬ ic representation of the input grapheme word. The fol¬ lowing examples illustrate the outputs of the phonetic translator unit 40 for a variety of inputs: grapheme input word: BUBBLE phonetic output word: B Ah B L grapheme input word: PHONE phonetic output word: F Ao N grapheme input word: BYTE phonetic output word: B Ay T grapheme input word: WOLF phonetic output word: W Uh L F
Reference is now made to Fig. 4 which illus¬ trates a preferred method of operation for dictionary lookup unit 50 of Fig. 1.
The input to the dictionary look up unit 50 are the strings that were guessed by spellguess unit 30, if the phonetic translator is omitted from the system of Fig. 1, or the strings that were generated by the phonet¬ ic translator unit 40. Dictionary look up unit 50 does not search for an exact match with the input strings, but rather performs a similarity search in order to find all dictionary words that are similar, to a predetermined extent, to the input strings, if not necessarily identi¬ cal. Dictionary look up 50 uses a similarity threshold in order to define the predetermined extent of similarity.
The method of Fig. 4 preferably includes the following processes: a. Initialization (process 310) - Determine the similarity threshold to be used. The string for which it
is desired to find similar dictionary words is translated into each of the reduced alphabets of the layers in the similarity search index. For example, if it is desired to find dictionary words which are similar to ADCINS using the similarity index illustrated in part in Fig. 13, ADCINS is translated into (6) in the layer 1 alphabet, (4,2) in the layer 2 alphabet, and (1,3,2) in the layer 3 alphabet. b. Scan similarity search index (process 320). A similarity search index, the structure of which is de¬ scribed above with reference to Fig. 13, is scanned recursively, layer by layer, in order to find all dic¬ tionary words that are similar (better than or equal to the similarity threshold) to the input string arriving from spellguess 30 or translator 40.
Process 320 recursively invokes processes 330 - 370 in order to perform a depth-first search for similar nodes. The search employs a conventional distance-be- tween-strings computation method to compute the distance between each node and the input string represented in the reduced alphabet of the layer to which the node belongs.
A suitable method for computing distance be¬ tween strings is described in Hall, P. A. "Approximate string matching", Computing surveys. Vol. 12(4), pp. 381 -402, 1980.
Parameters may be externally defined to deter¬ mine the relative costs of symbol insertion and symbol deletion, for each symbol in the dictionary and reduced alphabets and for symbol substitution, for each pair of symbols in the dictionary and reduced alphabets.
It is appreciated that either a spoken language alphabet or a phonetic alphabet may be used as the dic¬ tionary alphabet.
The output of this step is a set of all dic¬ tionary words whose similarity to the "dictionary alpha¬ bet" representation of the input string is equal to or exceeds the similarity threshold defined in the initiali-
SUBSTITUTE SHEET(R(JL£^
zation process 310.
Reference is now made to Fig. 5 which illus¬ trates a preferred method of operation for the database lookup unit 90 and the candidate sorter unit 100 (Fig. 1). The database lookup unit 90 accepts as input eacn individual dictionary entry in the list of dictionary entries generated by dictionary lookup unit 50 along with its similarity score, also computed by dictionary lookup unit 50, and its entry context (last name, first name, etc. )
The database lookup unit 90 operates for each database entry component. For example, in zip code direc¬ tory applications, the database lookup unit 90 operates after the state name is elicited from a user, then after the city's name is elicited and then after the street name is elicited. After each operation, the database lookup unit 90 updates the similarity vector 92 (Fig. 1) such that the similarity vector accumulates a similarity score for each database entry.
The output of the database lookup 90 and the candidate sorter 100 is a list of a predetermined number of database entries which are most similar to the words supplied by the user thus far, sorted according to their similarity scores. This list if stored in the best candi¬ dates box 94 of Fig. 1.
The method of Fig. 5 preferably includes the following processes: a. Initialization (process 380) - Uses concordance 72 to find the logical and physical location of all database entries that include the requested dictionary entry value in the appropriate context. For example, if the user supplies SMITH as a last name, all subscribers whose last name is similar to Smith will be located.
If this is the first database look up in the database access session, initialization process 380 also clears the similarity vector 92 and the best candidates box 94.
b. Database scan (process 390) - Scans all the database entries that were selected by the initializa¬ tion process 380. For each entry, the following 2 proc¬ esses are performed: i. Update database entry similarity score (process 400) - For each database entry, updates the similar¬ ity score by modifying the similarity score to reflect the degree of similarity of that database entry to the current word. For example, the word similarity score may be added to the appropriate similarity vector compo¬ nent. ii. Update best candidates box (process 410) Determines whether or not the previously updated database entry has to be included in the best candidates list, based on the following criteria: if the database entry is already in the best candidates box, it is kept there with the new score updated if the best candidates box is not full, the database entry is added unconditionally if the best candidates box is full and the similarity score of the updated database entry is higher than the last (worst score) database entry in the candi¬ dates box, the last database entry is replaced with the new one.
At the end of processes 380 through 410, the similarity vector includes updated similarity scores and the contents of the best candidates box is determined. At this point, a candidate sorter 100 is invoked in order to resort the selected database entries in the best candi¬ dates box. The current ordering in the best candidates box is based on the sum of similarity scores of individu¬ al words where the similarity score of any word, if under a predetermined threshold, is considered to be zero. The candidates sorter computes the similarity of each zero- scoring word.
Specifically, the candidate sorter 100 prefera-
bly performs the following processes: a. Best candidates scan (process 420) - Scans all the database entries in the best candidates box. b. Single candidate words scan (process 430) Scans all the words of a single database entry in the best candidates box. This process decides whether the similarity score of any one of the database entry strings may be improved, according to the following criteria:
- the word was supplied by the caller; and
- the word's similarity score is 0.
For each of the words fulfilling this criteria, the process invokes 2 additional processes (improve similarity 440, and update best candidates 450), in order to refine the similarity scores, first at the word level and then at the database entry level. c. Improve similarity scores (process 440) - Re¬ initiates a minimum edit distance similarity check, similar to process 330 in Fig. 4, between a current database entry word in the best candidates box 94 and each word previously guessed by the spellguess unit 30, using a lower threshold measure than that originally employed by process 330. If the resulting similari¬ ty score differs from 0, the word similarity score and the database entry similarity score are both updated in process 450. d. Update best candidates box (process 450) - Com¬ pute the total similarity score of each changed database entry in the best candidates box and the similarity vector. e. Candidate sorter (process 460) - Sorts the entire best candidates box according to the similarity score, using a suitable conventional sort mechanism, such as a heap.
The output of the entire database lookup 90 and candidate sorter 100 is a list of all best candidates database entries, sorted according to their similarity scores.
Reference is now made to Fig. 6 which illus¬ trates a preferred method of operation for the conversa¬ tion manager 80. The conversation manager controls the conversation between the caller and the system , from the moment the caller gains access to the computer and until the conversation is finished. The actions taken by the conversation manager, including message providing action, depend on the contents of the best candidates box 94.
The conversation manager 80 is preferably implemented as a state machine whose logic is defined by a decision table. However, the decision table is pref¬ erably completely external to the system to enable it to be easily modified.
The conversation manager 80 includes three building blocks: the state, the conditio and the action. The conversation begins at the initial state, and each additional item of information acquired, leads the con¬ versation to a different state. At each state, the con¬ versation manager checks for the first condition that is fulfilled, and performs the action that is associated with the state and the condition. After that, the conver¬ sation manager proceeds to the next state. For example, one state may be "beginning of conversation", a condition may be "was a word entered by caller?", and an action may be "propose a selection list", or "search the database".
The method of Fig. 6 preferably includes the following processes: a. Initialize conversation + welcome user (process 470) - Set the conversation to the initial state, and, typically, provide a welcome message, which in¬ cludes the instructions for using the database access system. b. Check condition (processes 480, 490 and 540) - Check whether the condition associated with the state is fulfilled or not. At least one condition has to be ful¬ filled per state, so the conversation manager loops until it finds the first one.
c. Action (process 500) - Perform the action that is associated with the state and with the first condi¬ tion that is fulfilled. d. End of conversation (process 510) - If the system enters the "end of the conversation" state, the conversation is ended. e. Transfer to operator (processes 520 and 550) - If the conversation manager comes to the conclusion that the conversation has to be transferred to a human opera¬ tor, it generates the appropriate routing instructions. This action may be useful, if for example, the conver¬ sation is too long, or the information provided by the caller does not identify a specific piece of informa¬ tion, or if the caller requests to be transferred to a human operator. f. New state (process 530) - The conversation manager proceeds to a new state, once it has finished handling the actions of previous state.
Reference is now made to Fig. 9 which is a sim¬ plified block diagram of one embodiment of the apparatus of Fig. 1 in which the ambiguous input comprises tele¬ phone keypad input. In the embodiment of Fig. 9, the ambiguous input generator 10 of Fig. 1 includes a plural¬ ity of conventional telephone instruments 715 communicat¬ ing via a telephone network with database accessing unit 20 via a conventional DTMF (dual tone multiple frequency) detection module 720 which is operative to convert the keystroke information into conventional digital data signals such as ASCII data signals.
In the embodiment of Fig. 9, the ambiguous input is keystroke information wherein each key corre¬ sponds to more than one letter of the English alphabet and may also correspond to the digit corresponding to the key. The data is ambiguous because every letter is en¬ tered by a single stroke on one of the telephone keypad buttons, and every button represents up to 3 different letters. For example, button 2 represents either A or B
or C, while button 3 represents either D or E or F. For example, the sequence of digits "2"; "3" could represent any of the following letter combinations: AD, AE, AF, BD, BE, BF, CD, CE, CF, as well as combinations of letters and the numbers "2" and "3". So, a string of length n can be interpreted in 3 to the power of n different ways. For example, a sequence of 6 keystrokes corresponds to 729 alternative words. For example, the names ARNOLD and BROOKE are keyed with the same keystroke sequence (276653). The same is true with names like BREEN and ARDEN (27336), or BUSH and BURG (2874).
Reference is now made to Fig. 10 which is another embodiment of the apparatus of Fig. 9 in which the ambiguous input is ambiguous oral character input rather than ambiguous keystroke input. In the embodiment of Fig. 10, the ambiguous input generator 10 of Fig. 1 comprises a plurality of telephone instruments 740 or other speech input devices providing ambiguous speech input to the database accessing unit 20 via a probabilis¬ tic spoken character recognition device 750.
The probabilistic spoken character recognition device 750 is similar to conventional speech recognition units, except that the output is probabilistic. In other words, the final decision stage in which probabilistic input, as to the contents of the speech, is reduced to a single possibility, typically by selecting the most probable speech content, is omitted.
A design suitable for a probabilistic spoken character recognition device is described in Rabiner L., Juang B., Fundamentals of Speech Recognition, Prentice- Hall, 1993, pp. 43 - 52.
In this embodiment, the user is prompted to provide oral input by spelling the desired words. For example, the conversation manager may prompt the user to "SPELL THE STREET NAME YOU ARE LOOKING FOR."
The input provided by device 750 is ambiguous since the probabilistic spoken character recognition box
750 outputs several possible characters, along with their relative probabilities, for every input spoken character. For example, a user's attempt to utter a "D", may be interpreted by the probabilistic spoken character recog¬ nition box 750, as "B" with probability 0.5, "D" with probability 0.3, and "T" with probability 0.2. A subse¬ quent utterance of the letter "D", may yield different output (letters and probabilities).
Reference is now made to Fig. 11 which is another embodiment of the apparatus of Fig. 1 in which the ambiguous input comprises ambiguous spoken phoneme input rather than ambiguous oral character input.
In the embodiment of Fig. 11, the ambiguous input generator 10 comprises a plurality of telephone instruments 760 or other speech input devices, and a probabilistic spoken phoneme recognition box 770. In this embodiment, the user is assumed to provide speech input in the form of spoken words. For example, the conversa¬ tion manager may prompt the user to "STATE THE STREET NAME YOU ARE LOOKING FOR."
The probabilistic spoken phoneme recognition device 770 is similar to conventional speech recognition units, except that the output is phonemes with associated probabilities. In other words, the final decision stage in which probabilistic input, as to the contents of the speech, is reduced to a single possibility, typically by selecting the most probable speech content, is omitted. A suitable design for a probabilistic spoken phoneme recog¬ nition device is described in Rabiner L., Juang B., Fundamentals of Speech Recognition. Prentice-Hall, 1993, pp. 43 - 52.
For example, the word "DAVE" may be phonetical¬ ly broken into the following phonemes: D, Ey and V (when using a certain phonetic alphabet).
When the word "DAVE" is input to a probabilis¬ tic spoken phoneme recognition box 770, it may recognize every phoneme in several ways, with a certain probability
applied to every phoneme. For instance, the first phoneme may be expressed as "(D, 0.4), or (B, 0.4), or (T, 0.2)"; the second phoneme may be expressed as "(Ey, 0.6) or (Eh, 0.3), or (Ay, 0.1)"; and the third phoneme may be expressed as "(V, 0.8), or (B, 0.1), or (F, 0.1)".
It is appreciated that the software components of the present invention may, if desired, be implemented in ROM (read-only memory) form. The software components may, generally, be implemented in hardware, if desired, using conventional techniques.
In the present specification and claims, the term "probabilistic string" refers to a sequence S over an alphabet A (such as the English alphabet, any non- English alphabet, the phonetic alphabet or any other set of characters) of n elements where each element S is a i finite non-empty set of (character, probability) pairs.
The term "keypad" refers to an input device including a plurality of input buttons or keys each of which, separately or in combination with other input buttons or keys, represents one or more characters. For example, each key on a conventional telephone keypad represents three, two or only one characters.
The term "erroneous user input" refers to a database accessing request provided by a user which comprises an inaccurate production of a word in the database. The inaccuracy may comprise one, some or all of the following types of inaccuracy: spelling inaccuracy, pronunciation inaccuracy or factual inaccuracy.
The term "spelling inaccuracy" refers to a cognitive or typographical error which results in an incorrect spelling of a database word. For example, a user wishing to access a street name BROADWAY may key in or spell out BROAWDAY or DROADWAY. A user wishing to access a locality PHILADELPHIA may key in or spell out FILADELFIA.
The term "pronunciation inaccuracy" refers to an incorrect oral rendition of a database word. For
example, a user wishing to access the name MARIE may state this request as "MARY".
The term "factual inaccuracy" refers to provi¬ sion of incorrect information. For example, a user wish¬ ing to access JOHN REMINGTON may request JACK REMINGTON.
It is appreciated that the user accessing the system shown and described herein can be either a human user or a computerized system.
It is appreciated that the ambiguous input may, alternatively, comprise the output of an OCR (optical character recognition) device such as a handwriting recognition device.
It is appreciated that various features of the invention which are, for clarity, described in the con¬ texts of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features of the invention which are, for brevity, de¬ scribed in the context of a single embodiment may also be provided separately or in any suitable subcombination.
It will be appreciated by persons skilled in the art that the present invention is not limited to what has been particularly shown and described hereinabove. Rather, the scope of the present invention is defined only by the claims that follow:
Claims
1. A database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request comprising a sequence of representations of possibly erroneous user inputs, the system comprising: a similar word finder operative, for at least one interpretation of each representation, to find at least one database word which is at least similar to that interpretation; and a database entry evaluator operative, for each database word found by the similar word finder, to assign similarity values for relevant entries in the database, said values representing the degree of similarity between each database entry and said request.
2. A system according to claim 1 wherein the request comprises a sequence of unambiguous representa¬ tions of possibly erroneous user inputs.
3. A system according to claim 1 wherein said relevant entries comprise all database entries having said database word in an individual context.
4. A system according to claim 1 wherein the request comprises a sequence of ambiguous representations of possibly erroneous user inputs and also comprising an ambiguous input evaluator operative to identify at least one interpretation of each ambiguous representation and to provide said interpretation to the similar word finder.
5. A system according to claim 4 and wherein said ambiguous input evaluator comprises a database analyzer operative to extract database characteristics from the database.
6. A system according to claim 1 and also compris¬ ing a session manager receiving similarity values from the database entry evaluator and operative to interac¬ tively provide a user sensible message based on said similarity values.
7. A system according to claim 6 and wherein the session manager is operative to prompt the user to pro¬ vide the request by orally spelling at least one database word.
8. A system according to claim 6 and wherein the session manager is operative to prompt the user to pro¬ vide the request by pronouncing at least one database word.
9. A system according to claim 6 and wherein the session manager is operative to prompt the user to pro¬ vide the request by keying at least one database word on the keypad.
10. A database accessing system for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the re¬ quest comprising at least one ambiguous representation of at least one possibly erroneous user input, the system comprising: an ambiguous input evaluator operative to identify at least one interpretation of each ambiguous representation; a similar word finder operative, for at least one interpretation of each representation, to find at least one database word which is at least similar to that interpretation; and a database entry evaluator operative, for each database word found by the similar word finder, to assign similarity values for relevant entries in the database, said values representing the degree of similarity between each database entry and said request.
11. A system according to claim 10 wherein said relevant entries comprise all database entries having said database word in an individual context.
12. A system according to claim 10 and wherein said ambiguous input evaluator comprises a database analyzer operative to extract database characteristics from the database.
13. A system according to claim 10 and also com¬ prising a session manager receiving similarity values from the database entry evaluator and operative to inter¬ actively provide a user sensible message based on said similarity values.
14. A system according to claim 13 and wherein the session manager is operative to prompt the user to pro¬ vide the request by orally spelling at least one database word.
15. A system according to claim 13 and wherein the session manager is operative to prompt the user to pro¬ vide the request by pronouncing at least one database word.
16. A system according to claim 13 and wherein the session manager is operative to prompt the user to pro¬ vide the request by keying at least one database word on the keypad.
17. A database accessing method for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the request comprising a sequence of representations of possibly erroneous user inputs, the method comprising: finding, for at least one interpretation of each representation, at least one database word which is at least similar to that interpretation; and assigning similarity values for relevant entries in the database, for each database word found, said values representing the degree of similarity between each database entry and said request.
18. A database accessing method for processing a request to access a database including a multiplicity of entries, each entry including at least one word, the re¬ quest comprising at least one ambiguous representation of at least one possibly erroneous user input, the method comprising: identifying at least one interpretation of each ambiguous representation; finding, for at least one interpretation of each representation, at least one database word which is at least similar to that interpretation; and for each database word found by the similar word finder, assigning similarity values for relevant entries in the database, said values representing the degree of similarity between each database entry and said request.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
AU37343/95A AU3734395A (en) | 1994-10-03 | 1995-10-02 | A database accessing system |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US31704094A | 1994-10-03 | 1994-10-03 | |
US08/317,040 | 1994-10-03 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO1996010795A1 true WO1996010795A1 (en) | 1996-04-11 |
Family
ID=23231844
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US1995/012600 WO1996010795A1 (en) | 1994-10-03 | 1995-10-02 | A database accessing system |
Country Status (3)
Country | Link |
---|---|
US (2) | US6018736A (en) |
AU (1) | AU3734395A (en) |
WO (1) | WO1996010795A1 (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0848536A2 (en) * | 1996-12-13 | 1998-06-17 | AT&T Corp. | Statistical database correction of alphanumeric account numbers for speach recognition and touch-tone recognition |
US6141661A (en) * | 1997-10-17 | 2000-10-31 | At&T Corp | Method and apparatus for performing a grammar-pruning operation |
US6154579A (en) * | 1997-08-11 | 2000-11-28 | At&T Corp. | Confusion matrix based method and system for correcting misrecognized words appearing in documents generated by an optical character recognition technique |
US6163596A (en) * | 1997-05-23 | 2000-12-19 | Hotas Holdings Ltd. | Phonebook |
WO2001011487A2 (en) * | 1999-08-04 | 2001-02-15 | Board Of Trustees Of The University Of Illinois | Apparatus, method and product for multi-attribute drug comparison |
US6205261B1 (en) | 1998-02-05 | 2001-03-20 | At&T Corp. | Confusion set based method and system for correcting misrecognized words appearing in documents generated by an optical character recognition technique |
US6205428B1 (en) | 1997-11-20 | 2001-03-20 | At&T Corp. | Confusion set-base method and apparatus for pruning a predetermined arrangement of indexed identifiers |
US6219453B1 (en) | 1997-08-11 | 2001-04-17 | At&T Corp. | Method and apparatus for performing an automatic correction of misrecognized words produced by an optical character recognition technique by using a Hidden Markov Model based algorithm |
US6223158B1 (en) | 1998-02-04 | 2001-04-24 | At&T Corporation | Statistical option generator for alpha-numeric pre-database speech recognition correction |
US6400805B1 (en) | 1998-06-15 | 2002-06-04 | At&T Corp. | Statistical database correction of alphanumeric identifiers for speech recognition and touch-tone recognition |
US7031925B1 (en) | 1998-06-15 | 2006-04-18 | At&T Corp. | Method and apparatus for creating customer specific dynamic grammars |
US7630899B1 (en) | 1998-06-15 | 2009-12-08 | At&T Intellectual Property Ii, L.P. | Concise dynamic grammars using N-best selection |
US20210006559A1 (en) * | 2019-07-02 | 2021-01-07 | Verint Systems Ltd. | System and method for identifying pairs of related information items |
Families Citing this family (139)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE69525178T2 (en) * | 1994-10-25 | 2002-08-29 | British Telecomm | ANNOUNCEMENT SERVICES WITH VOICE INPUT |
US6385202B1 (en) * | 1997-10-01 | 2002-05-07 | At&T Corp. | Method and system for allowing access to traditional voice mail and interactive voice response systems over internet-like networks |
JP3275816B2 (en) * | 1998-01-14 | 2002-04-22 | 日本電気株式会社 | Symbol string search method, symbol string search device, and recording medium recording symbol string search program |
US6411932B1 (en) * | 1998-06-12 | 2002-06-25 | Texas Instruments Incorporated | Rule-based learning of word pronunciations from training corpora |
US6144958A (en) * | 1998-07-15 | 2000-11-07 | Amazon.Com, Inc. | System and method for correcting spelling errors in search queries |
US6347295B1 (en) * | 1998-10-26 | 2002-02-12 | Compaq Computer Corporation | Computer method and apparatus for grapheme-to-phoneme rule-set-generation |
US6330564B1 (en) * | 1999-02-10 | 2001-12-11 | International Business Machines Corporation | System and method for automated problem isolation in systems with measurements structured as a multidimensional database |
US6477551B1 (en) * | 1999-02-16 | 2002-11-05 | International Business Machines Corporation | Interactive electronic messaging system |
US7292980B1 (en) * | 1999-04-30 | 2007-11-06 | Lucent Technologies Inc. | Graphical user interface and method for modifying pronunciations in text-to-speech and speech recognition systems |
US7136947B1 (en) * | 1999-06-10 | 2006-11-14 | Cadence Design Systems, Inc. | System and method for automatically synthesizing interfaces between incompatible protocols |
JP3476008B2 (en) * | 1999-09-10 | 2003-12-10 | 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 | A method for registering voice information, a method for specifying a recognition character string, a voice recognition device, a storage medium storing a software product for registering voice information, and a software product for specifying a recognition character string are stored. Storage media |
JP3476007B2 (en) * | 1999-09-10 | 2003-12-10 | 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 | Recognition word registration method, speech recognition method, speech recognition device, storage medium storing software product for registration of recognition word, storage medium storing software product for speech recognition |
KR100318762B1 (en) * | 1999-10-01 | 2002-01-04 | 윤덕용 | Phonetic distance method for similarity comparison of foreign words |
US6587818B2 (en) * | 1999-10-28 | 2003-07-01 | International Business Machines Corporation | System and method for resolving decoding ambiguity via dialog |
US7403888B1 (en) | 1999-11-05 | 2008-07-22 | Microsoft Corporation | Language input user interface |
US7165019B1 (en) | 1999-11-05 | 2007-01-16 | Microsoft Corporation | Language input architecture for converting one text form to another text form with modeless entry |
US6848080B1 (en) | 1999-11-05 | 2005-01-25 | Microsoft Corporation | Language input architecture for converting one text form to another text form with tolerance to spelling, typographical, and conversion errors |
US7725307B2 (en) * | 1999-11-12 | 2010-05-25 | Phoenix Solutions, Inc. | Query engine for processing voice based queries including semantic decoding |
US7050977B1 (en) * | 1999-11-12 | 2006-05-23 | Phoenix Solutions, Inc. | Speech-enabled server for internet website and method |
US9076448B2 (en) | 1999-11-12 | 2015-07-07 | Nuance Communications, Inc. | Distributed real time speech recognition system |
US7392185B2 (en) * | 1999-11-12 | 2008-06-24 | Phoenix Solutions, Inc. | Speech based learning/training system using semantic decoding |
DE19963812A1 (en) * | 1999-12-30 | 2001-07-05 | Nokia Mobile Phones Ltd | Method for recognizing a language and for controlling a speech synthesis unit and communication device |
WO2001059625A1 (en) * | 2000-02-10 | 2001-08-16 | Involve Technology, Llc | System for creating and maintaining a database of information utilizing user opinions |
JP2001249684A (en) * | 2000-03-02 | 2001-09-14 | Sony Corp | Device and method for recognizing speech, and recording medium |
US7047493B1 (en) * | 2000-03-31 | 2006-05-16 | Brill Eric D | Spell checker with arbitrary length string-to-string transformations to improve noisy channel spelling correction |
US7280964B2 (en) * | 2000-04-21 | 2007-10-09 | Lessac Technologies, Inc. | Method of recognizing spoken language with recognition of language color |
US6625595B1 (en) * | 2000-07-05 | 2003-09-23 | Bellsouth Intellectual Property Corporation | Method and system for selectively presenting database results in an information retrieval system |
US7136854B2 (en) * | 2000-07-06 | 2006-11-14 | Google, Inc. | Methods and apparatus for providing search results in response to an ambiguous search query |
US8706747B2 (en) * | 2000-07-06 | 2014-04-22 | Google Inc. | Systems and methods for searching using queries written in a different character-set and/or language from the target pages |
DE10042944C2 (en) * | 2000-08-31 | 2003-03-13 | Siemens Ag | Grapheme-phoneme conversion |
GB2367917A (en) * | 2000-10-12 | 2002-04-17 | Qas Systems Ltd | Retrieving data representing a postal address from a database of postal addresses using a trie structure |
WO2002063493A1 (en) * | 2001-02-08 | 2002-08-15 | 2028, Inc. | Methods and systems for automated semantic knowledge leveraging graph theoretic analysis and the inherent structure of communication |
US7233899B2 (en) * | 2001-03-12 | 2007-06-19 | Fain Vitaliy S | Speech recognition system using normalized voiced segment spectrogram analysis |
US6662190B2 (en) * | 2001-03-20 | 2003-12-09 | Ispheres Corporation | Learning automatic data extraction system |
US7117144B2 (en) * | 2001-03-31 | 2006-10-03 | Microsoft Corporation | Spell checking for text input via reduced keypad keys |
US7103534B2 (en) * | 2001-03-31 | 2006-09-05 | Microsoft Corporation | Machine learning contextual approach to word determination for text input via reduced keypad keys |
US6625600B2 (en) | 2001-04-12 | 2003-09-23 | Telelogue, Inc. | Method and apparatus for automatically processing a user's communication |
US7076731B2 (en) * | 2001-06-02 | 2006-07-11 | Microsoft Corporation | Spelling correction system and method for phrasal strings using dictionary looping |
WO2003001318A2 (en) * | 2001-06-26 | 2003-01-03 | Allot Communications Ltd. | Method for filter selection and array matching |
GB0118184D0 (en) * | 2001-07-26 | 2001-09-19 | Ibm | A method for generating homophonic neologisms |
US6687697B2 (en) * | 2001-07-30 | 2004-02-03 | Microsoft Corporation | System and method for improved string matching under noisy channel conditions |
US6804670B2 (en) * | 2001-08-22 | 2004-10-12 | International Business Machines Corporation | Method for automatically finding frequently asked questions in a helpdesk data set |
US7031910B2 (en) * | 2001-10-16 | 2006-04-18 | Xerox Corporation | Method and system for encoding and accessing linguistic frequency data |
US20030149566A1 (en) * | 2002-01-02 | 2003-08-07 | Esther Levin | System and method for a spoken language interface to a large database of changing records |
US7324942B1 (en) | 2002-01-29 | 2008-01-29 | Microstrategy, Incorporated | System and method for interactive voice services using markup language with N-best filter element |
CA2371731A1 (en) * | 2002-02-12 | 2003-08-12 | Cognos Incorporated | Database join disambiguation by grouping |
US7769592B2 (en) * | 2002-02-22 | 2010-08-03 | Nuance Communications, Inc. | Automatic selection of a disambiguation data field for a speech interface |
JP3799280B2 (en) * | 2002-03-06 | 2006-07-19 | キヤノン株式会社 | Dialog system and control method thereof |
US7673234B2 (en) * | 2002-03-11 | 2010-03-02 | The Boeing Company | Knowledge management using text classification |
JP4255239B2 (en) * | 2002-03-29 | 2009-04-15 | 富士通株式会社 | Document search method |
US8234115B2 (en) * | 2002-03-29 | 2012-07-31 | At&T Intellectual Property Ii, L.P. | Systems and methods for determining the N-best strings |
US6691103B1 (en) * | 2002-04-02 | 2004-02-10 | Keith A. Wozny | Method for searching a database, search engine system for searching a database, and method of providing a key table for use by a search engine for a database |
US6792096B2 (en) | 2002-04-11 | 2004-09-14 | Sbc Technology Resources, Inc. | Directory assistance dialog with configuration switches to switch from automated speech recognition to operator-assisted dialog |
US20040133558A1 (en) * | 2003-01-06 | 2004-07-08 | Masterwriter, Inc. | Information management system plus |
DE10306022B3 (en) * | 2003-02-13 | 2004-02-19 | Siemens Ag | Speech recognition method for telephone, personal digital assistant, notepad computer or automobile navigation system uses 3-stage individual word identification |
US7421418B2 (en) | 2003-02-19 | 2008-09-02 | Nahava Inc. | Method and apparatus for fundamental operations on token sequences: computing similarity, extracting term values, and searching efficiently |
US7236631B1 (en) | 2003-04-18 | 2007-06-26 | Kamran Reihani | System and method for automated symbolic recognition including logical symbols |
US7218782B1 (en) | 2003-04-18 | 2007-05-15 | Kamran Reihani | System and method for automated symbolic recognition including multi-phase symbolic reshaping |
US7227995B1 (en) | 2003-04-18 | 2007-06-05 | Kamran Reihani | System and method for automated symbolic recognition including spatial reasoning |
US7239749B1 (en) | 2003-04-18 | 2007-07-03 | Kamran Reihani | System and method for automated symbolic recognition including evidence-based technique |
US7242805B1 (en) | 2003-04-18 | 2007-07-10 | Kamran Reihani | System and method for automated symbolic recognition including database modeling |
US7363224B2 (en) * | 2003-12-30 | 2008-04-22 | Microsoft Corporation | Method for entering text |
US7299329B2 (en) * | 2004-01-29 | 2007-11-20 | Micron Technology, Inc. | Dual edge command in DRAM |
US20050261788A1 (en) * | 2004-05-21 | 2005-11-24 | Singman-Aste Michael K | System and method for programmatically searching backwards in a string |
GB0413743D0 (en) | 2004-06-19 | 2004-07-21 | Ibm | Method and system for approximate string matching |
US8392453B2 (en) * | 2004-06-25 | 2013-03-05 | Google Inc. | Nonstandard text entry |
US8972444B2 (en) | 2004-06-25 | 2015-03-03 | Google Inc. | Nonstandard locality-based text entry |
US7647294B2 (en) * | 2004-10-27 | 2010-01-12 | Fairbanks Jr William E | Indexing and querying engines and methods of indexing and querying |
US8175889B1 (en) | 2005-04-06 | 2012-05-08 | Experian Information Solutions, Inc. | Systems and methods for tracking changes of address based on service disconnect/connect data |
US7908242B1 (en) | 2005-04-11 | 2011-03-15 | Experian Information Solutions, Inc. | Systems and methods for optimizing database queries |
US9471566B1 (en) * | 2005-04-14 | 2016-10-18 | Oracle America, Inc. | Method and apparatus for converting phonetic language input to written language output |
JP2006309446A (en) * | 2005-04-27 | 2006-11-09 | Toshiba Corp | Classification dictionary updating device, classification dictionary updating program, and classification dictionary updating method |
US7412441B2 (en) * | 2005-05-31 | 2008-08-12 | Microsoft Corporation | Predictive phonetic data search |
US20070150279A1 (en) * | 2005-12-27 | 2007-06-28 | Oracle International Corporation | Word matching with context sensitive character to sound correlating |
US8005759B2 (en) | 2006-08-17 | 2011-08-23 | Experian Information Solutions, Inc. | System and method for providing a score for a used vehicle |
US7912865B2 (en) | 2006-09-26 | 2011-03-22 | Experian Marketing Solutions, Inc. | System and method for linking multiple entities in a business database |
US8036979B1 (en) | 2006-10-05 | 2011-10-11 | Experian Information Solutions, Inc. | System and method for generating a finance attribute from tradeline data |
US7979425B2 (en) * | 2006-10-25 | 2011-07-12 | Google Inc. | Server-side match |
US7831431B2 (en) | 2006-10-31 | 2010-11-09 | Honda Motor Co., Ltd. | Voice recognition updates via remote broadcast signal |
US8606666B1 (en) | 2007-01-31 | 2013-12-10 | Experian Information Solutions, Inc. | System and method for providing an aggregation tool |
US8285656B1 (en) | 2007-03-30 | 2012-10-09 | Consumerinfo.Com, Inc. | Systems and methods for data verification |
US7742982B2 (en) | 2007-04-12 | 2010-06-22 | Experian Marketing Solutions, Inc. | Systems and methods for determining thin-file records and determining thin-file risk levels |
US20080294540A1 (en) | 2007-05-25 | 2008-11-27 | Celka Christopher J | System and method for automated detection of never-pay data sets |
US20080319733A1 (en) * | 2007-06-20 | 2008-12-25 | At&T Corp. | System and method to dynamically manipulate and disambiguate confusable speech input using a table |
CN101689364B (en) * | 2007-07-09 | 2011-11-23 | 富士通株式会社 | Speech recognizer and speech recognition method |
US8301574B2 (en) | 2007-09-17 | 2012-10-30 | Experian Marketing Solutions, Inc. | Multimedia engagement study |
US9690820B1 (en) | 2007-09-27 | 2017-06-27 | Experian Information Solutions, Inc. | Database system for triggering event notifications based on updates to database records |
GB2456369A (en) * | 2008-01-11 | 2009-07-15 | Ibm | String pattern analysis for word or genome analysis |
US8516562B2 (en) | 2008-05-13 | 2013-08-20 | Veritrix, Inc. | Multi-channel multi-factor authentication |
US8006291B2 (en) | 2008-05-13 | 2011-08-23 | Veritrix, Inc. | Multi-channel multi-factor authentication |
US8468358B2 (en) | 2010-11-09 | 2013-06-18 | Veritrix, Inc. | Methods for identifying the guarantor of an application |
US8543393B2 (en) * | 2008-05-20 | 2013-09-24 | Calabrio, Inc. | Systems and methods of improving automated speech recognition accuracy using statistical analysis of search terms |
US8312033B1 (en) | 2008-06-26 | 2012-11-13 | Experian Marketing Solutions, Inc. | Systems and methods for providing an integrated identifier |
US8166297B2 (en) | 2008-07-02 | 2012-04-24 | Veritrix, Inc. | Systems and methods for controlling access to encrypted data stored on a mobile device |
US7991689B1 (en) | 2008-07-23 | 2011-08-02 | Experian Information Solutions, Inc. | Systems and methods for detecting bust out fraud using credit data |
JP2010039721A (en) * | 2008-08-04 | 2010-02-18 | Just Syst Corp | Data input device, method, and program |
US20100114887A1 (en) * | 2008-10-17 | 2010-05-06 | Google Inc. | Textual Disambiguation Using Social Connections |
WO2010051342A1 (en) | 2008-11-03 | 2010-05-06 | Veritrix, Inc. | User authentication for social networks |
US8494857B2 (en) | 2009-01-06 | 2013-07-23 | Regents Of The University Of Minnesota | Automatic measurement of speech fluency |
US20100332292A1 (en) | 2009-06-30 | 2010-12-30 | Experian Information Solutions, Inc. | System and method for evaluating vehicle purchase loyalty |
US8364518B1 (en) | 2009-07-08 | 2013-01-29 | Experian Ltd. | Systems and methods for forecasting household economics |
US8423351B2 (en) * | 2010-02-19 | 2013-04-16 | Google Inc. | Speech correction for typed input |
US8725613B1 (en) | 2010-04-27 | 2014-05-13 | Experian Information Solutions, Inc. | Systems and methods for early account score and notification |
US9152727B1 (en) | 2010-08-23 | 2015-10-06 | Experian Marketing Solutions, Inc. | Systems and methods for processing consumer information for targeted marketing applications |
US8639616B1 (en) | 2010-10-01 | 2014-01-28 | Experian Information Solutions, Inc. | Business to contact linkage system |
US20130238321A1 (en) * | 2010-11-22 | 2013-09-12 | Nec Corporation | Dialog text analysis device, method and program |
US9147042B1 (en) | 2010-11-22 | 2015-09-29 | Experian Information Solutions, Inc. | Systems and methods for data verification |
JP5666937B2 (en) * | 2011-02-16 | 2015-02-12 | 株式会社東芝 | Machine translation apparatus, machine translation method, and machine translation program |
US9263045B2 (en) * | 2011-05-17 | 2016-02-16 | Microsoft Technology Licensing, Llc | Multi-mode text input |
US9483606B1 (en) | 2011-07-08 | 2016-11-01 | Consumerinfo.Com, Inc. | Lifescore |
US8775299B2 (en) | 2011-07-12 | 2014-07-08 | Experian Information Solutions, Inc. | Systems and methods for large-scale credit data processing |
US8474014B2 (en) | 2011-08-16 | 2013-06-25 | Veritrix, Inc. | Methods for the secure use of one-time passwords |
US9317606B1 (en) * | 2012-02-03 | 2016-04-19 | Google Inc. | Spell correcting long queries |
US9576593B2 (en) | 2012-03-15 | 2017-02-21 | Regents Of The University Of Minnesota | Automated verbal fluency assessment |
US9853959B1 (en) | 2012-05-07 | 2017-12-26 | Consumerinfo.Com, Inc. | Storage and maintenance of personal data |
US9098487B2 (en) * | 2012-11-29 | 2015-08-04 | Hewlett-Packard Development Company, L.P. | Categorization based on word distance |
US9697263B1 (en) | 2013-03-04 | 2017-07-04 | Experian Information Solutions, Inc. | Consumer data request fulfillment system |
WO2014136173A1 (en) * | 2013-03-04 | 2014-09-12 | 三菱電機株式会社 | Search device |
US20150058006A1 (en) * | 2013-08-23 | 2015-02-26 | Xerox Corporation | Phonetic alignment for user-agent dialogue recognition |
US10102536B1 (en) | 2013-11-15 | 2018-10-16 | Experian Information Solutions, Inc. | Micro-geographic aggregation system |
US9529851B1 (en) | 2013-12-02 | 2016-12-27 | Experian Information Solutions, Inc. | Server architecture for electronic data quality processing |
US10262362B1 (en) | 2014-02-14 | 2019-04-16 | Experian Information Solutions, Inc. | Automatic generation of code for attributes |
JP5930228B2 (en) | 2014-02-25 | 2016-06-08 | 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 | Information processing apparatus, method, and program |
US9344419B2 (en) | 2014-02-27 | 2016-05-17 | K.Y. Trix Ltd. | Methods of authenticating users to a site |
US9576030B1 (en) | 2014-05-07 | 2017-02-21 | Consumerinfo.Com, Inc. | Keeping up with the joneses |
US20160062979A1 (en) * | 2014-08-27 | 2016-03-03 | Google Inc. | Word classification based on phonetic features |
US10242019B1 (en) | 2014-12-19 | 2019-03-26 | Experian Information Solutions, Inc. | User behavior segmentation using latent topic detection |
CN106293431A (en) * | 2015-05-26 | 2017-01-04 | 富泰华工业(深圳)有限公司 | Handwriting input recognition methods, system and electronic installation |
US10102189B2 (en) * | 2015-12-21 | 2018-10-16 | Verisign, Inc. | Construction of a phonetic representation of a generated string of characters |
US9910836B2 (en) * | 2015-12-21 | 2018-03-06 | Verisign, Inc. | Construction of phonetic representation of a string of characters |
US10102203B2 (en) * | 2015-12-21 | 2018-10-16 | Verisign, Inc. | Method for writing a foreign language in a pseudo language phonetically resembling native language of the speaker |
US9947311B2 (en) | 2015-12-21 | 2018-04-17 | Verisign, Inc. | Systems and methods for automatic phonetization of domain names |
WO2018039377A1 (en) | 2016-08-24 | 2018-03-01 | Experian Information Solutions, Inc. | Disambiguation and authentication of device users |
US9805073B1 (en) * | 2016-12-27 | 2017-10-31 | Palantir Technologies Inc. | Data normalization system |
CN116205724A (en) | 2017-01-31 | 2023-06-02 | 益百利信息解决方案公司 | Large scale heterogeneous data ingestion and user resolution |
US10803059B2 (en) * | 2018-01-30 | 2020-10-13 | Verizon Patent And Licensing Inc. | Searching using query graphs |
US10963434B1 (en) | 2018-09-07 | 2021-03-30 | Experian Information Solutions, Inc. | Data architecture for supporting multiple search models |
US11941065B1 (en) | 2019-09-13 | 2024-03-26 | Experian Information Solutions, Inc. | Single identifier platform for storing entity data |
US11880377B1 (en) | 2021-03-26 | 2024-01-23 | Experian Information Solutions, Inc. | Systems and methods for entity resolution |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4979206A (en) * | 1987-07-10 | 1990-12-18 | At&T Bell Laboratories | Directory assistance systems |
US5131045A (en) * | 1990-05-10 | 1992-07-14 | Roth Richard G | Audio-augmented data keying |
Family Cites Families (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5218536A (en) * | 1988-05-25 | 1993-06-08 | Franklin Electronic Publishers, Incorporated | Electronic spelling machine having ordered candidate words |
AU631276B2 (en) * | 1989-12-22 | 1992-11-19 | Bull Hn Information Systems Inc. | Name resolution in a directory database |
JP2943447B2 (en) * | 1991-01-30 | 1999-08-30 | 三菱電機株式会社 | Text information extraction device, text similarity matching device, text search system, text information extraction method, text similarity matching method, and question analysis device |
US5621857A (en) * | 1991-12-20 | 1997-04-15 | Oregon Graduate Institute Of Science And Technology | Method and system for identifying and recognizing speech |
GB9220404D0 (en) * | 1992-08-20 | 1992-11-11 | Nat Security Agency | Method of identifying,retrieving and sorting documents |
AU5803394A (en) * | 1992-12-17 | 1994-07-04 | Bell Atlantic Network Services, Inc. | Mechanized directory assistance |
US5457770A (en) * | 1993-08-19 | 1995-10-10 | Kabushiki Kaisha Meidensha | Speaker independent speech recognition system and method using neural network and/or DP matching technique |
US5873056A (en) * | 1993-10-12 | 1999-02-16 | The Syracuse University | Natural language processing system for semantic vector representation which accounts for lexical ambiguity |
US5623578A (en) * | 1993-10-28 | 1997-04-22 | Lucent Technologies Inc. | Speech recognition system allows new vocabulary words to be added without requiring spoken samples of the words |
US5479489A (en) * | 1994-11-28 | 1995-12-26 | At&T Corp. | Voice telephone dialing architecture |
-
1995
- 1995-10-02 AU AU37343/95A patent/AU3734395A/en not_active Abandoned
- 1995-10-02 WO PCT/US1995/012600 patent/WO1996010795A1/en active Application Filing
-
1996
- 1996-11-20 US US08/749,374 patent/US6018736A/en not_active Expired - Lifetime
-
1999
- 1999-06-17 US US09/335,456 patent/US6256630B1/en not_active Expired - Lifetime
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4979206A (en) * | 1987-07-10 | 1990-12-18 | At&T Bell Laboratories | Directory assistance systems |
US5131045A (en) * | 1990-05-10 | 1992-07-14 | Roth Richard G | Audio-augmented data keying |
Non-Patent Citations (1)
Title |
---|
NATIONAL ONLINE MEETING PROCEEDINGS, Vol. 1, issued 1990, VEITH R.H., "Information Retrieval Via Telephone Keypad and Voice: Experience with Telephone Directories", pages 443-451. * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0848536A2 (en) * | 1996-12-13 | 1998-06-17 | AT&T Corp. | Statistical database correction of alphanumeric account numbers for speach recognition and touch-tone recognition |
EP0848536A3 (en) * | 1996-12-13 | 1999-03-17 | AT&T Corp. | Statistical database correction of alphanumeric account numbers for speach recognition and touch-tone recognition |
US6137863A (en) * | 1996-12-13 | 2000-10-24 | At&T Corp. | Statistical database correction of alphanumeric account numbers for speech recognition and touch-tone recognition |
US6163596A (en) * | 1997-05-23 | 2000-12-19 | Hotas Holdings Ltd. | Phonebook |
US6154579A (en) * | 1997-08-11 | 2000-11-28 | At&T Corp. | Confusion matrix based method and system for correcting misrecognized words appearing in documents generated by an optical character recognition technique |
US6219453B1 (en) | 1997-08-11 | 2001-04-17 | At&T Corp. | Method and apparatus for performing an automatic correction of misrecognized words produced by an optical character recognition technique by using a Hidden Markov Model based algorithm |
US6141661A (en) * | 1997-10-17 | 2000-10-31 | At&T Corp | Method and apparatus for performing a grammar-pruning operation |
US6205428B1 (en) | 1997-11-20 | 2001-03-20 | At&T Corp. | Confusion set-base method and apparatus for pruning a predetermined arrangement of indexed identifiers |
US6223158B1 (en) | 1998-02-04 | 2001-04-24 | At&T Corporation | Statistical option generator for alpha-numeric pre-database speech recognition correction |
US6205261B1 (en) | 1998-02-05 | 2001-03-20 | At&T Corp. | Confusion set based method and system for correcting misrecognized words appearing in documents generated by an optical character recognition technique |
US6400805B1 (en) | 1998-06-15 | 2002-06-04 | At&T Corp. | Statistical database correction of alphanumeric identifiers for speech recognition and touch-tone recognition |
US7031925B1 (en) | 1998-06-15 | 2006-04-18 | At&T Corp. | Method and apparatus for creating customer specific dynamic grammars |
US7630899B1 (en) | 1998-06-15 | 2009-12-08 | At&T Intellectual Property Ii, L.P. | Concise dynamic grammars using N-best selection |
US7937260B1 (en) | 1998-06-15 | 2011-05-03 | At&T Intellectual Property Ii, L.P. | Concise dynamic grammars using N-best selection |
US8682665B2 (en) | 1998-06-15 | 2014-03-25 | At&T Intellectual Property Ii, L.P. | Concise dynamic grammars using N-best selection |
US9286887B2 (en) | 1998-06-15 | 2016-03-15 | At&T Intellectual Property Ii, L.P. | Concise dynamic grammars using N-best selection |
WO2001011487A2 (en) * | 1999-08-04 | 2001-02-15 | Board Of Trustees Of The University Of Illinois | Apparatus, method and product for multi-attribute drug comparison |
WO2001011487A3 (en) * | 1999-08-04 | 2003-12-24 | Univ Illinois | Apparatus, method and product for multi-attribute drug comparison |
US20210006559A1 (en) * | 2019-07-02 | 2021-01-07 | Verint Systems Ltd. | System and method for identifying pairs of related information items |
Also Published As
Publication number | Publication date |
---|---|
AU3734395A (en) | 1996-04-26 |
US6018736A (en) | 2000-01-25 |
US6256630B1 (en) | 2001-07-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO1996010795A1 (en) | A database accessing system | |
US5960385A (en) | Sentence reconstruction using word ambiguity resolution | |
US5832428A (en) | Search engine for phrase recognition based on prefix/body/suffix architecture | |
US6005495A (en) | Method and system for intelligent text entry on a numeric keypad | |
US6996531B2 (en) | Automated database assistance using a telephone for a speech based or text based multimedia communication mode | |
US8229747B2 (en) | System and method for spelling recognition using speech and non-speech input | |
US7277029B2 (en) | Using language models to expand wildcards | |
US20020126097A1 (en) | Alphanumeric data entry method and apparatus using reduced keyboard and context related dictionaries | |
US20070094008A1 (en) | Conversation control apparatus | |
US20030139921A1 (en) | System and method for hybrid text mining for finding abbreviations and their definitions | |
Vitale | An algorithm for high accuracy name pronunciation by parametric speech synthesizer | |
US6625600B2 (en) | Method and apparatus for automatically processing a user's communication | |
JP2003216181A (en) | Constraint-based speech recognition system and method | |
CN109684928B (en) | Chinese document identification method based on internet retrieval | |
EP0775963A2 (en) | Indexing a database by finite-state transducer | |
US6052443A (en) | Alphanumeric message composing method using telephone keypad | |
CN101218625A (en) | Dictionary lookup for mobile devices using spelling recognition | |
WO2001003123A1 (en) | Multimodal data input device | |
CN111626059B (en) | Information processing method and device | |
CN111414459B (en) | Character relationship acquisition method, device, electronic equipment and storage medium | |
KR930000593B1 (en) | Voice information service system and method utilizing approximate matching | |
JPH10105578A (en) | Similar word retrieving method utilizing point | |
JP3758241B2 (en) | Voice information retrieval device | |
JP3663012B2 (en) | Voice input device | |
EP1187431A1 (en) | Portable terminal with voice dialing minimizing memory usage |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): AM AT AU BB BG BR BY CA CH CN CZ DE DK EE ES FI GB GE HU IS JP KE KG KP KR KZ LK LR LT LU LV MD MG MK MN MW MX NO NZ PL PT RO RU SD SE SG SI SK TJ TM TT UA UG UZ VN |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): KE MW SD SZ UG AT BE CH DE DK ES FR GB GR IE IT LU MC NL PT SE BF BJ CF CG CI CM GA GN ML MR NE SN TD TG |
|
DFPE | Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101) | ||
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
REG | Reference to national code |
Ref country code: DE Ref legal event code: 8642 |
|
122 | Ep: pct application non-entry in european phase | ||
NENP | Non-entry into the national phase |
Ref country code: CA |