WO2004097568A2 - Method and apparatus for machine learning a document relevance function - Google Patents
Method and apparatus for machine learning a document relevance function Download PDFInfo
- Publication number
- WO2004097568A2 WO2004097568A2 PCT/US2004/012813 US2004012813W WO2004097568A2 WO 2004097568 A2 WO2004097568 A2 WO 2004097568A2 US 2004012813 W US2004012813 W US 2004012813W WO 2004097568 A2 WO2004097568 A2 WO 2004097568A2
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- relevance
- document
- documents
- query
- function
- 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
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/335—Filtering based on additional data, e.g. user or group profiles
- G06F16/337—Profile generation, learning or modification
-
- 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
- G06F16/95—Retrieval from the web
- G06F16/951—Indexing; Web crawling techniques
-
- 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
-
- 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
- Y10S707/99937—Sorting
Definitions
- the present invention relates to the field of search engines for locating documents in a database, such as an index of documents stored on servers coupled to the Internet or in an intranet, and in particular the present invention relates to a method and apparatus for determining a document relevance function for estimating a relevance score of a document in a database with respect to a query.
- a search engine To be useful to a user, a search engine must be able to determine, from amongst billions of documents, the two or three documents that a human user would be most interested in, given the query that the user has submitted.
- search engine designers have attempted to construct relevance functions that take a query and a document as their input and return a relevance value.
- the relevance value may be used, for example, to create a list of the documents indexed by the search engine, the list ranking the documents in order of relevance to the query, to serve this need.
- the underlying relevance function For the top two or three documents on this list to be useful to a user, the underlying relevance function must be able to accurately and quickly determine the relevance of a given document to a query.
- a user's perception of true relevance is influenced by a number of factors, many of which are highly subjective. These preferences are generally difficult to capture in an algorithmic set of rules defining a relevance function. Furthermore, these subjective factors may change over time, as for example when current events are associated with a particular query term. As another example, changes over time in the aggregate content of the documents available in the Internet may also alter a user's perception of the relative relevance of a given document to a particular query. A user who receives a return list from a search engine that contains documents that he does not perceive to be highly relevant will quickly become frustrated and abandon the use of the search engine.
- the present invention provides a method for determining a document relevance function for estimating a relevance score of a document in a database with respect to a query. First, for each of a plurality of test queries, a respective set of result documents is collected from the database. Next, for each test query, a subset of the documents in the respective result set is selected and a set of training relevance scores is assigned to documents in the subset. Finally, a relevance function is determined based on the plurality of test queries, the subsets of documents, and the sets of training relevance scores. [0006] Some embodiments further provide a method of selecting a subset of the documents in the respective result set of documents for each query. First, a document is selected from the respective result set.
- a surrogate relevance score relating the selected document to the current query is determined. Then, based on the determined surrogate relevance score, the selected document is assigned to at least one relevance tier of a plurality of relevance tiers. The selecting, determining, and assigning is then repeated until a terrnination condition is reached.
- the termination condition may be that each relevance tier contains at least a respective predefined number of documents, or may be that a highest relevance tier contains at least the predefined number of documents. Other terrnination conditions maybe used as well.
- Some of these embodiments additionally provide a method of assigning a set of training relevance scores to the documents in a subset of the documents selected from the respective result set of documents for each query.
- each of a first plurality of documents in a subset of the documents from a result set is submitted to a respective plurality of human subjects.
- Documents in the first plurality of documents have surrogate relevance scores within a predefined range.
- One or more human subjects determines an individual relevance score for the submitted document with respect to the query.
- a training relevance score is assigned to each submitted document with respect to the query based on the individual relevance scores determined by the human subjects.
- a set of features to be used as predictor variables in the construction of the relevance function is first constructed.
- Each of the features in the set maybe a function of one or more properties of a respective document, a respective query, or both.
- a relevance function is parameterized in terms of a finite set of parameters (e.g., coefficients) and base functions.
- the relevance function takes as its input the set of features and returns a relevance value as its output.
- each base function takes as its input a subset of the features and outputs a value.
- a partial error is defined to relate the training relevance score of a given document with respect to a particular query to a value produced by the document ranking function as applied to the given document with respect to the particular query.
- the defining of the partial error is then repeated with respect to a plurality of given documents and a plurality of particular queries so as to produce a set of partial errors.
- the parameters are then selected so as to minimize a total error which is a function of the set of partial errors.
- a relevance function that, given a document and a query, produces a relevance value is determined.
- a document ranking function is determined, based on the relevance function.
- the document ranking function, given a query and a plurality of documents, produces an ordered list of the documents in which the relevance values of the documents in the list monotonically decreases.
- a relevance function generation module for determining a relevance function based on the plurality of test queries, the subsets of documents, and the sets of training relevance scores.
- FIG. 1 illustrates a client computer submitting a query to a search engine, the search engine employing a document relevance function determined by a relevance function determining system.
- Fig. 2A illustrates an exemplary query containing one or more terms.
- Fig. 2B illustrates a test set of queries and the submission of the test set of queries to a search engine, resulting in a result set of documents.
- Fig. 3 illustrates a result set of documents and a plurality of relevance tiers for assignment of a subset of the documents thereto.
- Fig. 4 is a flow chart of a relevance function determining method.
- Fig. 5 is a block diagram of an exemplary base function, in this case a binary classification tree.
- Fig. 6 is a block diagram of a relevance function determining system.
- a computer network 100 includes one or more client computers 104 connected to a. network 105.
- the network 105 may be the Internet or, in other embodiments, an intranet.
- a collection of documents 103 known as the World Wide Web 102 are available for retrieval by the client computers through the network 105.
- documents are located by a uniform resource locator, such as "http://www.av.com.” By supplying the URL to a document server (not depicted), the document 103 corresponding to the URL can be accessed.
- the computer network 100 includes a search engine.
- search engines available on the Internet include but are not limited to AltaVista (at the URL http://www.av.com), Google (at the URL http://www.google.com), and Yahoo! (at the URL http://www.yahoo.com).
- Search engines typically include a database, the database indexing documents on the World Wide Web.
- a user of a client computer 104-1 who desires to retrieve a document relevant to a particular topic, but who is unsure or unaware of the URL of such a document, submits a query 112 to the search engine, typically through the network 105.
- the search engine 106 after receipt of the query 112, examines the database of documents in an attempt to find those documents that the user will regard as highly relevant to the submitted query 112.
- Some embodiments provide a method of determining a document relevance function, the relevance function used by search engine 106 and determined by a relevance function deterrnining system 108.
- the relevance function deterrnining system 108 may, in some embodiments, be implemented on a different computer system than the computer system that implements the search engine 106. In other embodiments, a single computer system maybe used to implement the functionality of the search engine as well as that of the relevance function determining system 108.
- Embodiments of the relevance function determining system 108 collect a result set of documents for a plurality of test queries.
- the plurality of test queries are determined, at least in part, based on logs of queries submitted by users of client computers 104 to search engine 106.
- the result set of documents is determined by submitting the test queries to one or more search engines 106 and receiving a response, typically a list of URLs, therefrom.
- the relevance function deterrnining system is optionally coupled to network 105, and can thereby retrieve one or more of the documents in the result set.
- Typical embodiments of the relevance function determining system include access to one or more human subjects 110.
- a typical query 112 includes one or more terms 202.
- the query depicted contains three terms. Such a query is referred to as a "three-term” query. Similarly, queries containing only one term are referred to as “one- term” queries, and queries containing two terms are referred to as “two-term” queries.
- the individual terms are delimited by a blank space, or possibly some other means, when submitted by the user.
- a plurality 204 of test queries includes one or more queries 112.
- the plurality 204 includes at least one one-term query 112-1 and at least one two-term query 112-2.
- the plurality 204 may include only one-term queries, only two-term queries, only three-term queries, or any combination of types of queries, possibly including queries having more than three terms.
- the plurality 204 is determined by sampling of queries from one or more logs, stored by search engine 106 of user-submitted queries 112 to the search engine. A preliminary set of query strings is first sampled from the logs. Queries relating to subject matter determined to be outside the scope of knowledge of a user base are then eliminated. Finally, the remaining queries in the preliminary set are assigned to the plurality 204 of test queries.
- the plurality of test queries are selected by sampling words from a lexicon of one-word entries and assigning each of the words so sampled to the plurality of test queries.
- lexicons from which words maybe sampled include dictionaries, such as Merriam- Webster's Collegiate Dictionary, Merriam- Webster, Inc.; 10th edition (1998).
- combinations of two or more words are sampled from the lexicon, and the combinations assigned to the plurality 204 of test queries.
- each test query 112 in the plurality 204 of test queries is submitted to a search engine 106.
- the database is an index of documents retrieved from the World Wide Web.
- the result set of documents is sometimes collected by submitting each test query to a search engine, receiving from the search engine a list of documents on the World Wide Web containing one or more of the terms in the test query, and adding one or more of the documents on the list of documents to the respective result set.
- the search engine responds by returning a list of documents (typically referenced by their URL), each document containing at least one of the terms in the test query 112.
- each respective result set 304 contains one or more documents 210, and each document 210 is associated with one or more test queries 112.
- Each respective result set 304 preferably stores only identifying information for the documents 210 (e.g., the URL of each document, a title or partial title of the document, and a small portion of the document, which may contain one or more of the query terms) rather than the complete contents of the documents.
- the respective result set 304 of each test query 112 is collected by submitting each query 112 from the plurality 204 to two or more search engines indexing documents on the World Wide Web.
- a test query is submitted to both AltaVista's search engine (at the URL http://www.av.com) and a second search engine.
- the second search engine is Google's search engine (at the URL http://www.google.com).
- the URL's of the 200 highest ranked documents (having rankings 1-200), as determined by each search engine, is then received.
- the five documents ranked in positions 1-5 (corresponding to highest relevance to the submitted query) by AltaVista's search engine are added to the respective result set.
- the document at the next highest position is examined, until the end of the list of documents is reached. This is repeated, if possible, until 10 documents from the list of URL's returned by the second search engine have been added to the respective result set.
- the respective result set includes, where possible, 20 documents.
- relatively small means 1% or less of a total.
- the plurality 208 of respective result sets 304 includes more than 4000 documents, but the selected subset of documents to which relevance training scores are assigned includes only 40 documents.
- relevance training scores are assigned to a number of documents ranging between 40 and 200.
- one or more respective result sets 304 each include documents retrieved in reference to a respective one of the plurality of test queries.
- a surrogate relevance score relating each document to the query of the respective result set containing the document is first determined.
- the surrogate relevance score may be determined by submitting the query of the respective result set to a search engine and deteimining the surrogate relevance score of the document as a function of the position of the document in a list returned by the search engine. For example, Query 1 in Fig. 3 is submitted to a search engine, and a surrogate relevance score assigned to document 306-1 as a function of the position of document 306-1 on a list provided by the search engine in response to the submission of Query 1.
- Each set 308 of relevance tiers includes one or more relevance tiers 310.
- each relevance tier 310 has associated with it a minimum surrogate relevance score and, optionally, a maximum surrogate relevance score.
- one or more o " the documents in the respective result set 304 is assigned to each relevance tier 310 in such a way that the surrogate relevance score of the document is greater than or equal to the minimum surrogate relevance score associated with the relevance tier.
- the one or more documents in the respective result set 304 may be selected in a number of ways, for example by random sampling from amongst the documents in the respective result set 304
- An example of assigning a document is illustrated by document 306-1, which may have a surrogate relevance score of 10.
- the minimum relevance surrogate score associated with relevance tier 310-11 may be 8.
- document 306-1 is assigned to tier 310-11.
- Other methods for selecting and assigning documents from the respective result set 304 to relevance tiers 310 are possible.
- the relevance tiers typically include documents representing results with low surrogate relevance scores. [0032]
- each relevance tier 310 additionally has an associated maximum surrogate relevance score.
- tier 310-N2 may have an associated minimum relevance score of 4 and an associated maximum relevance score of 8.
- Relevance tier 310-N1 may have an associated maximum relevance score of 100 and an associated minimum relevance score of 8.
- a document is assigned to a given relevance tier if the surrogate relevance score of the document is less than the maximum surrogate relevance score associated with the tier and greater than or equal to the minimum relevance surrogate score associated the tier.
- document 306-2 may have a surrogate relevance score of 5.
- document 306-2 is assigned to relevance tier 310-N2, but not assigned to relevance tier 310-N1.
- the range of relevance scores associated with the maximum and minimum score for each tier is selected so that the ranges are nonoverlapping. In another embodiment, the ranges overlap, for example with each tier having an assigned nrinimum score but no assigned maximum score.
- the process of assigning documents from the respective result set 304 to relevance tiers 310 is repeated until each relevance tier 310 contains at least a respective predefined number of documents. For example, in some embodiments, the process of assigning is repeated until at least 10 documents are assigned to each relevance tier.
- the minimum number of documents required for each tier maybe different. For example, the minimum number of documents required for tiers 1, 2 and 3 maybe 10, 40 and 100, respectively.
- a first plurality of the documents from the relevance tiers 310 is submitted to one or more human subjects (element 110 in Fig. 1).
- the first plurality of documents may include documents only from the first (i.e., highest) relevance tiers 310-11, ..., 310-N1 associated with each query in the plurality of test queries.
- the human subjects examine each document submitted to them, along with the associated query, and determine an individual relevance score relating the document to the query.
- a large number e.g., fifty
- each human subject is presented with a number of document-query pairings and a survey form including the following statements:
- a human subject assigns an individual relevance score from amongst one of the numbers 1-10, 10 indicating the highest relevance and 1 indicating the lowest relevance.
- the arithmetic average of the individual scores is determined, and the average used as the training relevance score of the document.
- Other methods of deterrnining the training relevance score of the document are also possible, including but not limited to using the median of the individual relevance scores and using the arithmetic average of a sample of the individual relevance scores selected to have a variance below a predetermined threshold.
- training scores are assigned to documents in a second plurality of documents selected from the documents in the subset.
- the second plurality may include all documents in the subset that have not been submitted to human users as part of the first plurality of the documents.
- documents in the second subset are assigned a predetermined low relevance training score. For example, documents from the lowest respective relevance tier 310-M1, 310-M2 (for a second query, not shown), . 310-MN may be assigned a predetermined relevance score of 0. In some embodiments, documents from the next lowest respective relevance tiers for each query may be assigned a predetermined relevance score of 1.5, and so on for the other tiers having documents in the second subset.
- all of the documents in the relevance tiers 310 may be assigned a training relevance score without submitting all of the documents to human subjects.
- Obtaining individual relevance scores from human subjects for all of the documents in the relevance tires 310 may be prohibitively expensive.
- little information is typically gained from low individual relevance scores assigned to documents by human subjects.
- Machine learning techniques in addition to minimizing an error associated with the training relevance scores and the relevance scores produced by the relevance function, determine a relevance function in such a way that the relevance score produces for new queries (not contained in the test set of queries) or new documents (not in the relevance tiers) is close to the fraining relevance score that would have been determined for a document relative to a query, had the document been in the relevance tiers and the query in the test set.
- logistic regression is used as a machine learning technique for determining a relevance function.
- Logistic regression has been demonstrated, via retrospective experiments, to improve relevance ranking in the context of information retrieval. See, for example, Gey, F. C. "Inferring the Probability of Relevance Using the Method of Logistic Regression", SIGIR 1994: 222-231, which is hereby incorporated by reference in its entirety.
- a set of features to be used as predictor variables is first determined.
- the set of features are to be used as predictor variables in the construction of the relevance function.
- a feature is a means of quantifying an aspect of the relationship of a query to a document, or of an aspect of the document itself. Given a document and, possibly, a query, a feature returns a value. Features that return a value based only the contents of a document itself are referred to as query-independent features. Query-independent features may depend on properties of the document itself.
- query-independent features may depend on properties of the server on which the document is located, and possibly properties of the relationship of this server to other servers on the Web.
- Features that require both a document and a query to return one or more values are referred to as query-dependent features.
- z the i th feature selected to relate a document to a query.
- query-independent features that may be included in the set of features include, but are not limited to:
- Eigenrank one or more values associated with elements of an eigenvector of a stochastic matrix derived from the incidence matrix of the web, where links are considered edges of a directed graph, see Page L., Brin S., Motwani R., and Winograd T., "The PageRank citation ranking: Bringing order to the Web," http://citeseer.nj.nec.com/page98pagerank.html, website last accessed April 10, 2003, hereby incorporated by reference in its entirety;
- HUB a value related to the connectivity of the incidence matrix of the web, specifically the so-called Kleinberg application of singular value decomposition to the graph, see Kleinberg L., "Authoritative sources in a hyperlinked environment," in Proceedings of the Nineth Annual ACM-SIAM Symposium on Discrete Algorithms, 1998, hereby incorporated by reference in its entirety;
- URL Depth a value representing the depth of traversal of a web site needed to retrieve a document, determined based on, at least in part, the number of "/"'s in a URL;
- Quality Score a value representing the authority of a document, determined as a function of other features of the document
- Spam Index a value indicating whether a document is likely to be a "spam document,” for example a value of "1" when a document contains excessive repetition of a term, indicating the document has been designed to artificially enhance relevance relative to a query containing the term;
- Document Length the number of individual terms in the document, expressed as an integer.
- query-dependent features examples include, but are not limited to:
- Anchor Text Score a value representing the number of documents containing both a link to the current document and one or more terms in the current query
- Match Location within a document, a value indicating whether a term from a query can be found in the title or the body of the document, and if the term is found in the body but not the title, a value indicative of how deep in the body the term is found;
- Match Frequency a value indicating how many times a term from a query is found in a document, for example, the number of times any term from the query is found in a document;
- Term Weight a standard feature in the art of information retrieval (IR), for example the inverse of the logarithm of the number of times a term from the query is found within the database of documents, expressed as a fraction of the total database size;
- Proximity a value indicating, for multi-term queries, whether or not the terms in the query can be found adjacent to one another in the document.
- the set of features selected includes all of the features described above.
- z ⁇ Eigenrank (ER), Hub, URL Depth, Quality Score, Spam Index, Family Friendliness, Document Length, Anchor Text Score, Match Location, Match Frequency, Term Weight, and Proximity ⁇ .
- a subset of these features maybe selected.
- features in addition to those listed above maybe included in the set of features, for example, derived features that are themselves functions of subsets of the features above.
- the next step in a method 400 of determining a document relevance function is step 404.
- the relevance function is parameterized in terms of a finite set of coefficients and base function.
- the ranking function is parameterized as a linear function of the features
- the parameterization of the relevance function may involve more complex functions of the features, for example
- each base function a r (z) takes as input a subset of the set of features, possibly including all of the set of features, and outputs a value.
- the base functions themselves are parameterized by coefficients.
- the base functions may include wavelet functions, for instance with each wavelet function related to a mother wavelet through a process of dilation, translation, or other possible processes, the precise relationship determined by additional coefficients.
- the base functions may include neural networks.
- the relevance function is parameterized in the following way
- the relevance function is a general, nonlinear function of the coefficients and the base functions
- the base functions include decision trees.
- a base learner may be a binary classification tree 500, as depicted in Fig. 5.
- a binary classification tree in block 502-1 examines a first "splitting variable," x, , and compares its value to a "splitting location" b x .
- the splitting variable is a linear or nonlinear function of one or more of the features in the set of selected features.
- the flow of the tree proceeds to either block 502-2 or block 502-3.
- Each block 502 has associated with it a splitting variable and a splitting location.
- the splitting variables (for example 508-1) and splitting locations (for example 508-2) are coefficients 508 that are needed to determine the base function represented by tree 400.
- each terminal node a terminal value (for example, 508-3), is assigned as the value of the base function.
- the tree 400 is a binary classifier and the terminal values only assume values of "-1" or "1,” corresponding to one or the two classes into which the tree 400 classifies documents.
- the terminal values (for example, 508-3) are also coefficients 508 that are needed to determine the base function represented by tree 400.
- the base functions include classification and regression (CART) trees.
- CART trees may be used to parameterize the relevance function itself or, in some embodiments, the gradient of the relevance function with respect to a subset of the parameters.
- CART trees including methods of selecting parameters of CART trees to minimize an error, see, L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone, Classification and Regression Tress. Belmont, CA: Wadsworth, 1984, which is hereby incorporated by reference in its entirety.
- the relevance function is determined as a function of a linear combination of the terminal values returned from application of each base function to the set of selected features.
- a master classifier is first constructed.
- each base function (tree) is denoted a r
- the parameters (splitting variables, splitting locations, and terminal values) of the base functions are denoted d ⁇ ,...,d j
- the parameters of the combination of the base functions are denoted c r
- the signQ function "votes" by assigning a value of "1" to the master classifier when its argument is positive or zero and a value of "-1" otherwise.
- the relevance function is the inverse logistic transform of the weighted sum of the base functions: exp
- the relevance tiers 310 contain one or more documents, each document assigned a training relevance score relating it to the query associated with the respective result set 304 from which it originated.
- the training relevance score of the h document in the ⁇ relevance tier associated with the n" 1 query is denoted y nmJ .
- the values of the features are also possibly dependent on the query and the document, so the value of the features associated with the 1 document in the m th relevance tier associated with the * query are similarly denoted z nmj .
- Machine learning techniques such as those disclosed herein, determine a document relevance function based on partial errors associated with the training relevance scores and values of the features in the relevance tiers.
- Documents in the relevance tiers represent a small (generally, less than 1%) portion of the documents that will be encountered when the relevance function is used, for example, to characterize documents from the Web relative to queries that are not in the test set of queries.
- machine learning techniques must also attempt to control the "generalization error," i.e. the error associated with the relevance score produced by the relevance function and the relevance score a user would determine for a query not in the test set of queries or for a document not in the relevance tiers.
- step 406 of a method 400 includes defining a partial error e mnj associated with the h document in the m th relevance tier and the « th query as a function of the square of a difference between the training relevance score for the document and query and the value produced by the relevance function,
- the error is defined as a function of the absolute value of a difference between the training relevance score and the value produced by the relevance function
- a training classification function is first defined, the classification function assigning a document to at least one of a number of classes based, at least in part, on the framing relevance score of the document. For example, referring again to Fig. 3, a training classification function may be defined such that all documents from the highest relevance tiers 310-11, ... and 310-N1 associated with each test query is assigned to a first class. Membership in this class may be associated with the value "1.” Documents in the relevance tiers 310 not classified as being members of the first class may be assigned to a second class. Membership in the second class may be associated with the value "-1.” In some embodiments, more than two classes may be defined in a similar manner.
- the framing classification function is a binary classifier, assigning documents to one of two classes, a partial error is defined
- a method 400 of determining a relevance function repeats step 406 of defining a partial error one or more times.
- Step 408 determines whether or not more partial errors are to be determined.
- partial errors are repeatedly defined by step 406 until a partial error has been defined for each document in each relevance tier (310, Fig. 3) for each respective query.
- Step 410 of method 400 selects coefficients of the relevance function so as to minimize a total error.
- the total error is simply the sum of the partial errors for all documents in each relevance tier associated with each query in the plurality of test queries
- the total error may be a more complex function of the partial errors, for example a weighted sum of the partial errors or the sum of the squares of the partial errors.
- the parameters (e.g. coefficients) of the relevance function are determined so as to minimize the total error.
- the selection of the parameters that minimize the total error may also be accomplished via a Boost procedure.
- AdaBoost pseudo-code of which is available from Schapire, R.E. "The Boosting Approach to Machine Learning: An Overview", in MSRI Workshop on Nonlinear Estimation and Classification, 2002, hereby incorporated by reference in its entirety
- AdaBoost pseudo-code of which is available from Schapire, R.E. "The Boosting Approach to Machine Learning: An Overview", in MSRI Workshop on Nonlinear Estimation and Classification, 2002, hereby incorporated by reference in its entirety
- the parameters of the relevance function, for each set of weightings, is determined so as to minimize the total error as determined with that set of weightings. See, for example, Schapire, supra.
- GradientBoost an implementation of the gradient boosting algorithm, GradientBoost, may be used to select the parameters of the relevance function that minimize the total error. See, for example, Friedman, J.H. "Greedy Function Approximation: A Gradient Boosting Machine," The Annals of Statistics 29(5), October 2001, which is hereby incorporated by reference in its entirety.
- GradientBoost is a particularly attractive technique to use for this purpose when the base functions are classification and regression (CART) trees. For a complete description of CART trees, including methods of selecting parameters of CART trees to minimize an error, see, L.
- the relevance function is employed to further determine a ranking function.
- a user submits a query to the search engine.
- a set of documents is retrieved from the database that are to be ranked for relevance with respect to the query.
- only documents including one or more of the terms in the query are including in this set. In other embodiments, other criteria may be used to select this set.
- the values of the selected set of features is evaluated for the document as paired with the query.
- the relevance function is then used to determine a relevance value for the document as paired with the query.
- an ordered list of the set of documents is produced. The ordered list is determined so that the documents on the list are ordered in a sequence of decreasing relevance. Thus, the document that appears first on the ordered list will have the numerically greatest relevance value of all the documents in the set and the document that appears last of the ordered list will have the minimal relevance score of all the documents in the set.
- the present invention can be implemented as a computer program product that comprises a computer program mechanism embedded in a computer readable storage medium.
- the relevance function determining system (108 in Fig.l) includes
- a network interface 606 for communication with other computers on the network (for example the search engine 106 depicted in Fig.l);
- primary and secondary storage 610 comprised of computer-readable media for storing one or more data structures and one or more modules for execution by central processing units(s) 604; and • an internal bus 608 for transmitting and receiving electronic signals amongst the central processing unit(s) 604, network interface 606, and primary and secondary storage medium 608.
- the program modules in storage 610 may be stored on a CD-ROM, magnetic disk storage product, or any other computer readable data or program storage product.
- the software modules in the computer program product may also be distributed electronically, via the Internet or otherwise, by transmission of a computer data signal (in which the software modules are embedded) on a carrier wave.
- Storage 610 may include, at least:
- a data structure 614 for storing data identifying or otherwise representing the result sets of documents; the data in this data structure 614 will typically include URL's referring to documents in the result set of documents.
- Storage 610 may further include:
- a collecting module 616 for collecting a respective result set of documents from the database for each of a plurality of test queries; the collecting module 616 will typically call a search engine module in the same or a different computer system to generate the result set of documents;
- a sampling module 618 for selecting, for each test query of the plurality of test queries, a subset of the documents in the respective result set
- a scoring module 620 for assigning a set of framing relevance scores to the documents in each selected subset
- a relevance function generation module 622 for deterrnining a relevance function based on the plurality of test queries, the subsets of documents, and the sets of training relevance scores.
Abstract
Description
Claims
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP04750656A EP1623298A2 (en) | 2003-04-25 | 2004-04-23 | Method and apparatus for machine learning a document relevance function |
JP2006513331A JP2006524869A (en) | 2003-04-25 | 2004-04-23 | Method and apparatus for machine learning of document relevance functions |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US10/424,170 | 2003-04-25 | ||
US10/424,170 US7197497B2 (en) | 2003-04-25 | 2003-04-25 | Method and apparatus for machine learning a document relevance function |
Publications (2)
Publication Number | Publication Date |
---|---|
WO2004097568A2 true WO2004097568A2 (en) | 2004-11-11 |
WO2004097568A3 WO2004097568A3 (en) | 2006-01-05 |
Family
ID=33299288
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2004/012813 WO2004097568A2 (en) | 2003-04-25 | 2004-04-23 | Method and apparatus for machine learning a document relevance function |
Country Status (6)
Country | Link |
---|---|
US (1) | US7197497B2 (en) |
EP (1) | EP1623298A2 (en) |
JP (1) | JP2006524869A (en) |
KR (1) | KR20060006945A (en) |
CN (1) | CN1826597A (en) |
WO (1) | WO2004097568A2 (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2008538149A (en) | 2005-03-18 | 2008-10-09 | ウィンク テクノロジ−ズ インコ−ポレイテッド | Rating method, search result organizing method, rating system, and search result organizing system |
US9092523B2 (en) | 2005-02-28 | 2015-07-28 | Search Engine Technologies, Llc | Methods of and systems for searching by incorporating user-entered information |
US11157475B1 (en) | 2019-04-26 | 2021-10-26 | Bank Of America Corporation | Generating machine learning models for understanding sentence context |
US11423231B2 (en) | 2019-08-27 | 2022-08-23 | Bank Of America Corporation | Removing outliers from training data for machine learning |
US11449559B2 (en) | 2019-08-27 | 2022-09-20 | Bank Of America Corporation | Identifying similar sentences for machine learning |
US11526804B2 (en) | 2019-08-27 | 2022-12-13 | Bank Of America Corporation | Machine learning model training for reviewing documents |
US11556711B2 (en) | 2019-08-27 | 2023-01-17 | Bank Of America Corporation | Analyzing documents using machine learning |
US11783005B2 (en) | 2019-04-26 | 2023-10-10 | Bank Of America Corporation | Classifying and mapping sentences using machine learning |
Families Citing this family (129)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6883135B1 (en) * | 2000-01-28 | 2005-04-19 | Microsoft Corporation | Proxy server using a statistical model |
AUPR605601A0 (en) * | 2001-07-03 | 2001-07-26 | Blackwood, Miles | Pipeite of sandwich construction |
US7792828B2 (en) | 2003-06-25 | 2010-09-07 | Jericho Systems Corporation | Method and system for selecting content items to be presented to a viewer |
US7610313B2 (en) | 2003-07-25 | 2009-10-27 | Attenex Corporation | System and method for performing efficient document scoring and clustering |
US8548995B1 (en) * | 2003-09-10 | 2013-10-01 | Google Inc. | Ranking of documents based on analysis of related documents |
US7424467B2 (en) * | 2004-01-26 | 2008-09-09 | International Business Machines Corporation | Architecture for an indexer with fixed width sort and variable width sort |
US7293005B2 (en) * | 2004-01-26 | 2007-11-06 | International Business Machines Corporation | Pipelined architecture for global analysis and index building |
US7499913B2 (en) | 2004-01-26 | 2009-03-03 | International Business Machines Corporation | Method for handling anchor text |
US8296304B2 (en) | 2004-01-26 | 2012-10-23 | International Business Machines Corporation | Method, system, and program for handling redirects in a search engine |
US7191175B2 (en) | 2004-02-13 | 2007-03-13 | Attenex Corporation | System and method for arranging concept clusters in thematic neighborhood relationships in a two-dimensional visual display space |
US7584221B2 (en) * | 2004-03-18 | 2009-09-01 | Microsoft Corporation | Field weighting in text searching |
US7260573B1 (en) * | 2004-05-17 | 2007-08-21 | Google Inc. | Personalizing anchor text scores in a search engine |
US7461064B2 (en) | 2004-09-24 | 2008-12-02 | International Buiness Machines Corporation | Method for searching documents for ranges of numeric values |
US7606793B2 (en) | 2004-09-27 | 2009-10-20 | Microsoft Corporation | System and method for scoping searches using index keys |
US7739277B2 (en) * | 2004-09-30 | 2010-06-15 | Microsoft Corporation | System and method for incorporating anchor text into ranking search results |
US7827181B2 (en) | 2004-09-30 | 2010-11-02 | Microsoft Corporation | Click distance determination |
US7761448B2 (en) | 2004-09-30 | 2010-07-20 | Microsoft Corporation | System and method for ranking search results using click distance |
US7779001B2 (en) * | 2004-10-29 | 2010-08-17 | Microsoft Corporation | Web page ranking with hierarchical considerations |
US7716198B2 (en) | 2004-12-21 | 2010-05-11 | Microsoft Corporation | Ranking search results using feature extraction |
JP4977624B2 (en) * | 2005-01-18 | 2012-07-18 | ヤフー! インコーポレイテッド | Matching and ranking of sponsored search listings that incorporate web search technology and web content |
US7356777B2 (en) | 2005-01-26 | 2008-04-08 | Attenex Corporation | System and method for providing a dynamic user interface for a dense three-dimensional scene |
US7404151B2 (en) | 2005-01-26 | 2008-07-22 | Attenex Corporation | System and method for providing a dynamic user interface for a dense three-dimensional scene |
US7921365B2 (en) | 2005-02-15 | 2011-04-05 | Microsoft Corporation | System and method for browsing tabbed-heterogeneous windows |
US7792833B2 (en) * | 2005-03-03 | 2010-09-07 | Microsoft Corporation | Ranking search results using language types |
US20060200460A1 (en) * | 2005-03-03 | 2006-09-07 | Microsoft Corporation | System and method for ranking search results using file types |
US7680772B2 (en) * | 2005-03-09 | 2010-03-16 | Intuit Inc. | Search quality detection |
US7546294B2 (en) * | 2005-03-31 | 2009-06-09 | Microsoft Corporation | Automated relevance tuning |
US20090125498A1 (en) * | 2005-06-08 | 2009-05-14 | The Regents Of The University Of California | Doubly Ranked Information Retrieval and Area Search |
US7627564B2 (en) * | 2005-06-21 | 2009-12-01 | Microsoft Corporation | High scale adaptive search systems and methods |
US8244722B1 (en) | 2005-06-30 | 2012-08-14 | Google Inc. | Ranking documents |
US8249344B2 (en) * | 2005-07-01 | 2012-08-21 | Microsoft Corporation | Grammatical parsing of document visual structures |
US20070005588A1 (en) * | 2005-07-01 | 2007-01-04 | Microsoft Corporation | Determining relevance using queries as surrogate content |
US8195654B1 (en) | 2005-07-13 | 2012-06-05 | Google Inc. | Prediction of human ratings or rankings of information retrieval quality |
US8417693B2 (en) * | 2005-07-14 | 2013-04-09 | International Business Machines Corporation | Enforcing native access control to indexed documents |
US7599917B2 (en) * | 2005-08-15 | 2009-10-06 | Microsoft Corporation | Ranking search results using biased click distance |
CN107103027A (en) * | 2005-10-04 | 2017-08-29 | 汤姆森路透社全球资源公司 | System, method and software for recognizing relevant legal documents |
US7630964B2 (en) * | 2005-11-14 | 2009-12-08 | Microsoft Corporation | Determining relevance of documents to a query based on identifier distance |
US20070150477A1 (en) * | 2005-12-22 | 2007-06-28 | International Business Machines Corporation | Validating a uniform resource locator ('URL') in a document |
US8250061B2 (en) * | 2006-01-30 | 2012-08-21 | Yahoo! Inc. | Learning retrieval functions incorporating query differentiation for information retrieval |
US8509563B2 (en) * | 2006-02-02 | 2013-08-13 | Microsoft Corporation | Generation of documents from images |
US8001121B2 (en) * | 2006-02-27 | 2011-08-16 | Microsoft Corporation | Training a ranking function using propagated document relevance |
US7451120B1 (en) | 2006-03-20 | 2008-11-11 | Google Inc. | Detecting novel document content |
US20070233679A1 (en) * | 2006-04-03 | 2007-10-04 | Microsoft Corporation | Learning a document ranking function using query-level error measurements |
US7647314B2 (en) * | 2006-04-28 | 2010-01-12 | Yahoo! Inc. | System and method for indexing web content using click-through features |
US7593934B2 (en) * | 2006-07-28 | 2009-09-22 | Microsoft Corporation | Learning a document ranking using a loss function with a rank pair or a query parameter |
US7647353B2 (en) * | 2006-11-14 | 2010-01-12 | Google Inc. | Event searching |
US8176055B1 (en) | 2007-03-27 | 2012-05-08 | Google Inc. | Content entity management |
US20080250008A1 (en) * | 2007-04-04 | 2008-10-09 | Microsoft Corporation | Query Specialization |
US8117137B2 (en) * | 2007-04-19 | 2012-02-14 | Microsoft Corporation | Field-programmable gate array based accelerator system |
US7853589B2 (en) * | 2007-04-30 | 2010-12-14 | Microsoft Corporation | Web spam page classification using query-dependent data |
US8073803B2 (en) * | 2007-07-16 | 2011-12-06 | Yahoo! Inc. | Method for matching electronic advertisements to surrounding context based on their advertisement content |
KR100899930B1 (en) * | 2007-07-24 | 2009-05-28 | 엔에이치엔(주) | System and Method for Generating Relating Data Class |
US20090055436A1 (en) * | 2007-08-20 | 2009-02-26 | Olakunle Olaniyi Ayeni | System and Method for Integrating on Demand/Pull and Push Flow of Goods-and-Services Meta-Data, Including Coupon and Advertising, with Mobile and Wireless Applications |
US8645390B1 (en) | 2007-08-31 | 2014-02-04 | Google Inc. | Reordering search query results in accordance with search context specific predicted performance functions |
US7895198B2 (en) * | 2007-09-28 | 2011-02-22 | Yahoo! Inc. | Gradient based optimization of a ranking measure |
US9348912B2 (en) * | 2007-10-18 | 2016-05-24 | Microsoft Technology Licensing, Llc | Document length as a static relevance feature for ranking search results |
US20090106221A1 (en) * | 2007-10-18 | 2009-04-23 | Microsoft Corporation | Ranking and Providing Search Results Based In Part On A Number Of Click-Through Features |
US7840569B2 (en) | 2007-10-18 | 2010-11-23 | Microsoft Corporation | Enterprise relevancy ranking using a neural network |
US8375073B1 (en) | 2007-11-12 | 2013-02-12 | Google Inc. | Identification and ranking of news stories of interest |
US8005774B2 (en) * | 2007-11-28 | 2011-08-23 | Yahoo! Inc. | Determining a relevance function based on a query error derived using a structured output learning technique |
US8099417B2 (en) * | 2007-12-12 | 2012-01-17 | Microsoft Corporation | Semi-supervised part-of-speech tagging |
US8775416B2 (en) * | 2008-01-09 | 2014-07-08 | Yahoo!Inc. | Adapting a context-independent relevance function for identifying relevant search results |
US7984004B2 (en) * | 2008-01-17 | 2011-07-19 | Microsoft Corporation | Query suggestion generation |
US7996379B1 (en) | 2008-02-01 | 2011-08-09 | Google Inc. | Document ranking using word relationships |
US8650144B2 (en) * | 2008-02-14 | 2014-02-11 | Yahoo! Inc. | Apparatus and methods for lossless compression of numerical attributes in rule based systems |
US8812493B2 (en) | 2008-04-11 | 2014-08-19 | Microsoft Corporation | Search results ranking using editing distance and document information |
KR100953488B1 (en) * | 2008-04-16 | 2010-04-19 | 엔에이치엔(주) | Method and system for generating rank learning model using error minimization |
US8126839B2 (en) * | 2008-06-19 | 2012-02-28 | Yahoo! Inc. | Methods and apparatuses for adapting a ranking function of a search engine for use with a specific domain |
US8171021B2 (en) * | 2008-06-23 | 2012-05-01 | Google Inc. | Query identification and association |
US8458170B2 (en) | 2008-06-30 | 2013-06-04 | Yahoo! Inc. | Prefetching data for document ranking |
US8621424B2 (en) * | 2008-06-30 | 2013-12-31 | Yahoo! Inc. | Compiler based code modification for use in document ranking |
US8301638B2 (en) * | 2008-09-25 | 2012-10-30 | Microsoft Corporation | Automated feature selection based on rankboost for ranking |
US8131659B2 (en) * | 2008-09-25 | 2012-03-06 | Microsoft Corporation | Field-programmable gate array based accelerator system |
US8073727B2 (en) * | 2008-10-23 | 2011-12-06 | Sap Ag | System and method for hierarchical weighting of model parameters |
US8671093B2 (en) * | 2008-11-18 | 2014-03-11 | Yahoo! Inc. | Click model for search rankings |
CN101477542B (en) * | 2009-01-22 | 2013-02-13 | 阿里巴巴集团控股有限公司 | Sampling analysis method, system and equipment |
US20100268709A1 (en) * | 2009-04-21 | 2010-10-21 | Yahoo! Inc., A Delaware Corporation | System, method, or apparatus for calibrating a relevance score |
US20100293175A1 (en) * | 2009-05-12 | 2010-11-18 | Srinivas Vadrevu | Feature normalization and adaptation to build a universal ranking function |
US20100332550A1 (en) * | 2009-06-26 | 2010-12-30 | Microsoft Corporation | Platform For Configurable Logging Instrumentation |
US20100332531A1 (en) * | 2009-06-26 | 2010-12-30 | Microsoft Corporation | Batched Transfer of Arbitrarily Distributed Data |
US8635223B2 (en) | 2009-07-28 | 2014-01-21 | Fti Consulting, Inc. | System and method for providing a classification suggestion for electronically stored information |
US8082247B2 (en) * | 2009-07-30 | 2011-12-20 | Microsoft Corporation | Best-bet recommendations |
US20110029516A1 (en) * | 2009-07-30 | 2011-02-03 | Microsoft Corporation | Web-Used Pattern Insight Platform |
EP2471009A1 (en) | 2009-08-24 | 2012-07-04 | FTI Technology LLC | Generating a reference set for use during document review |
US20110264609A1 (en) * | 2010-04-22 | 2011-10-27 | Microsoft Corporation | Probabilistic gradient boosted machines |
US8572496B2 (en) * | 2010-04-27 | 2013-10-29 | Go Daddy Operating Company, LLC | Embedding variable fields in individual email messages sent via a web-based graphical user interface |
US8738635B2 (en) | 2010-06-01 | 2014-05-27 | Microsoft Corporation | Detection of junk in search result ranking |
US8375061B2 (en) * | 2010-06-08 | 2013-02-12 | International Business Machines Corporation | Graphical models for representing text documents for computer analysis |
US20120078062A1 (en) | 2010-09-24 | 2012-03-29 | International Business Machines Corporation | Decision-support application and system for medical differential-diagnosis and treatment using a question-answering system |
US9208231B1 (en) * | 2010-12-01 | 2015-12-08 | Google Inc. | Identifying languages relevant to resources |
CA2829735C (en) * | 2011-03-10 | 2016-05-31 | Textwise Llc | Method and system for information modeling and applications thereof |
US8666914B1 (en) * | 2011-05-23 | 2014-03-04 | A9.Com, Inc. | Ranking non-product documents |
US9477756B1 (en) * | 2012-01-16 | 2016-10-25 | Amazon Technologies, Inc. | Classifying structured documents |
US9495462B2 (en) | 2012-01-27 | 2016-11-15 | Microsoft Technology Licensing, Llc | Re-ranking search results |
US8972328B2 (en) | 2012-06-19 | 2015-03-03 | Microsoft Corporation | Determining document classification probabilistically through classification rule analysis |
US9213770B1 (en) * | 2012-08-14 | 2015-12-15 | Amazon Technologies, Inc. | De-biased estimated duplication rate |
US9122681B2 (en) | 2013-03-15 | 2015-09-01 | Gordon Villy Cormack | Systems and methods for classifying electronic information using advanced active learning techniques |
US9565147B2 (en) | 2014-06-30 | 2017-02-07 | Go Daddy Operating Company, LLC | System and methods for multiple email services having a common domain |
US9996624B2 (en) * | 2014-06-30 | 2018-06-12 | Google Llc | Surfacing in-depth articles in search results |
US10764265B2 (en) * | 2014-09-24 | 2020-09-01 | Ent. Services Development Corporation Lp | Assigning a document to partial membership in communities |
US10509834B2 (en) | 2015-06-05 | 2019-12-17 | Apple Inc. | Federated search results scoring |
US10621189B2 (en) | 2015-06-05 | 2020-04-14 | Apple Inc. | In-application history search |
US10755032B2 (en) * | 2015-06-05 | 2020-08-25 | Apple Inc. | Indexing web pages with deep links |
US10509833B2 (en) * | 2015-06-05 | 2019-12-17 | Apple Inc. | Proximity search scoring |
US10592572B2 (en) | 2015-06-05 | 2020-03-17 | Apple Inc. | Application view index and search |
US10229117B2 (en) | 2015-06-19 | 2019-03-12 | Gordon V. Cormack | Systems and methods for conducting a highly autonomous technology-assisted review classification |
RU2632133C2 (en) * | 2015-09-29 | 2017-10-02 | Общество С Ограниченной Ответственностью "Яндекс" | Method (versions) and system (versions) for creating prediction model and determining prediction model accuracy |
EP3188038B1 (en) * | 2015-12-31 | 2020-11-04 | Dassault Systèmes | Evaluation of a training set |
US10372714B2 (en) | 2016-02-05 | 2019-08-06 | International Business Machines Corporation | Automated determination of document utility for a document corpus |
AU2017274558B2 (en) | 2016-06-02 | 2021-11-11 | Nuix North America Inc. | Analyzing clusters of coded documents |
US10755182B2 (en) * | 2016-08-11 | 2020-08-25 | International Business Machines Corporation | System and method for ground truth evaluation |
US10621492B2 (en) | 2016-10-21 | 2020-04-14 | International Business Machines Corporation | Multiple record linkage algorithm selector |
CN108572900B (en) * | 2017-03-09 | 2021-07-13 | 北京京东尚科信息技术有限公司 | Blank pit position monitoring method, system, electronic equipment and storage medium |
AU2018200643A1 (en) * | 2017-03-09 | 2018-09-27 | Accenture Global Solutions Limited | Smart advisory for distributed and composite testing teams based on production data and analytics |
US10372426B2 (en) * | 2017-11-06 | 2019-08-06 | International Business Machines Corporation | Cognitive redundant coding corpus determination system |
US11409749B2 (en) * | 2017-11-09 | 2022-08-09 | Microsoft Technology Licensing, Llc | Machine reading comprehension system for answering queries related to a document |
RU2693324C2 (en) | 2017-11-24 | 2019-07-02 | Общество С Ограниченной Ответственностью "Яндекс" | Method and a server for converting a categorical factor value into its numerical representation |
US11341138B2 (en) * | 2017-12-06 | 2022-05-24 | International Business Machines Corporation | Method and system for query performance prediction |
US10831770B2 (en) * | 2017-12-12 | 2020-11-10 | International Business Machines Corporation | System and method for estimating query performance in document retrieval |
US10915538B2 (en) | 2018-03-23 | 2021-02-09 | Home Depot Product Authority, Llc | Ranking and presenting search engine results based on category-specific ranking models |
US11093512B2 (en) * | 2018-04-30 | 2021-08-17 | International Business Machines Corporation | Automated selection of search ranker |
RU2721159C1 (en) | 2018-12-13 | 2020-05-18 | Общество С Ограниченной Ответственностью "Яндекс" | Method and server for generating meta-attribute for ranging documents |
RU2744028C2 (en) * | 2018-12-26 | 2021-03-02 | Общество С Ограниченной Ответственностью "Яндекс" | Method and system for storing multiple documents |
US11403300B2 (en) | 2019-02-15 | 2022-08-02 | Wipro Limited | Method and system for improving relevancy and ranking of search result |
US11205506B2 (en) | 2019-05-22 | 2021-12-21 | International Business Machines Corporation | Verifying natural language processing in health care |
CN110598272B (en) * | 2019-08-22 | 2022-11-22 | 合肥工业大学 | Heuristic generation method and device for multi-unmanned platform information interaction topology |
US11436235B2 (en) * | 2019-09-23 | 2022-09-06 | Ntent | Pipeline for document scoring |
CN112231621B (en) * | 2020-10-13 | 2021-09-24 | 电子科技大学 | Method for reducing element detection limit based on BP-adaboost |
US11893981B1 (en) | 2023-07-11 | 2024-02-06 | Seekr Technologies Inc. | Search system and method having civility score |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5696962A (en) * | 1993-06-24 | 1997-12-09 | Xerox Corporation | Method for computerized information retrieval using shallow linguistic analysis |
US6026388A (en) * | 1995-08-16 | 2000-02-15 | Textwise, Llc | User interface and other enhancements for natural language information retrieval system and method |
US20020114394A1 (en) * | 2000-12-06 | 2002-08-22 | Kai-Kuang Ma | System and method for motion vector generation and analysis of digital video clips |
Family Cites Families (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6119114A (en) * | 1996-09-17 | 2000-09-12 | Smadja; Frank | Method and apparatus for dynamic relevance ranking |
US5909510A (en) * | 1997-05-19 | 1999-06-01 | Xerox Corporation | Method and apparatus for document classification from degraded images |
US6651057B1 (en) * | 1999-09-03 | 2003-11-18 | Bbnt Solutions Llc | Method and apparatus for score normalization for information retrieval applications |
US6430559B1 (en) * | 1999-11-02 | 2002-08-06 | Claritech Corporation | Method and apparatus for profile score threshold setting and updating |
US20030074353A1 (en) * | 1999-12-20 | 2003-04-17 | Berkan Riza C. | Answer retrieval technique |
US7062485B1 (en) * | 2000-09-01 | 2006-06-13 | Huaichuan Hubert Jin | Method and apparatus for score normalization for information retrieval applications |
US6701317B1 (en) * | 2000-09-19 | 2004-03-02 | Overture Services, Inc. | Web page connectivity server construction |
US7010527B2 (en) * | 2001-08-13 | 2006-03-07 | Oracle International Corp. | Linguistically aware link analysis method and system |
US7158983B2 (en) * | 2002-09-23 | 2007-01-02 | Battelle Memorial Institute | Text analysis technique |
US7917483B2 (en) * | 2003-04-24 | 2011-03-29 | Affini, Inc. | Search engine and method with improved relevancy, scope, and timeliness |
-
2003
- 2003-04-25 US US10/424,170 patent/US7197497B2/en not_active Expired - Lifetime
-
2004
- 2004-04-23 WO PCT/US2004/012813 patent/WO2004097568A2/en active Application Filing
- 2004-04-23 EP EP04750656A patent/EP1623298A2/en not_active Withdrawn
- 2004-04-23 CN CNA2004800174686A patent/CN1826597A/en active Pending
- 2004-04-23 KR KR1020057020306A patent/KR20060006945A/en not_active Application Discontinuation
- 2004-04-23 JP JP2006513331A patent/JP2006524869A/en not_active Abandoned
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5696962A (en) * | 1993-06-24 | 1997-12-09 | Xerox Corporation | Method for computerized information retrieval using shallow linguistic analysis |
US6026388A (en) * | 1995-08-16 | 2000-02-15 | Textwise, Llc | User interface and other enhancements for natural language information retrieval system and method |
US20020114394A1 (en) * | 2000-12-06 | 2002-08-22 | Kai-Kuang Ma | System and method for motion vector generation and analysis of digital video clips |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11341144B2 (en) | 2005-02-28 | 2022-05-24 | Pinterest, Inc. | Methods of and systems for searching by incorporating user-entered information |
US10311068B2 (en) | 2005-02-28 | 2019-06-04 | Pinterest, Inc. | Methods of and systems for searching by incorporating user-entered information |
US11693864B2 (en) | 2005-02-28 | 2023-07-04 | Pinterest, Inc. | Methods of and systems for searching by incorporating user-entered information |
US9092523B2 (en) | 2005-02-28 | 2015-07-28 | Search Engine Technologies, Llc | Methods of and systems for searching by incorporating user-entered information |
US10157233B2 (en) | 2005-03-18 | 2018-12-18 | Pinterest, Inc. | Search engine that applies feedback from users to improve search results |
JP2012212455A (en) * | 2005-03-18 | 2012-11-01 | Search Engine Technologies Llc | Ranking method and ranking system |
US11036814B2 (en) | 2005-03-18 | 2021-06-15 | Pinterest, Inc. | Search engine that applies feedback from users to improve search results |
JP2008538149A (en) | 2005-03-18 | 2008-10-09 | ウィンク テクノロジ−ズ インコ−ポレイテッド | Rating method, search result organizing method, rating system, and search result organizing system |
US9367606B1 (en) | 2005-03-18 | 2016-06-14 | Search Engine Technologies, Llc | Search engine that applies feedback from users to improve search results |
US11429896B1 (en) | 2019-04-26 | 2022-08-30 | Bank Of America Corporation | Mapping documents using machine learning |
US11328025B1 (en) | 2019-04-26 | 2022-05-10 | Bank Of America Corporation | Validating mappings between documents using machine learning |
US11244112B1 (en) | 2019-04-26 | 2022-02-08 | Bank Of America Corporation | Classifying and grouping sentences using machine learning |
US11423220B1 (en) | 2019-04-26 | 2022-08-23 | Bank Of America Corporation | Parsing documents using markup language tags |
US11429897B1 (en) | 2019-04-26 | 2022-08-30 | Bank Of America Corporation | Identifying relationships between sentences using machine learning |
US11694100B2 (en) | 2019-04-26 | 2023-07-04 | Bank Of America Corporation | Classifying and grouping sentences using machine learning |
US11157475B1 (en) | 2019-04-26 | 2021-10-26 | Bank Of America Corporation | Generating machine learning models for understanding sentence context |
US11783005B2 (en) | 2019-04-26 | 2023-10-10 | Bank Of America Corporation | Classifying and mapping sentences using machine learning |
US11423231B2 (en) | 2019-08-27 | 2022-08-23 | Bank Of America Corporation | Removing outliers from training data for machine learning |
US11449559B2 (en) | 2019-08-27 | 2022-09-20 | Bank Of America Corporation | Identifying similar sentences for machine learning |
US11526804B2 (en) | 2019-08-27 | 2022-12-13 | Bank Of America Corporation | Machine learning model training for reviewing documents |
US11556711B2 (en) | 2019-08-27 | 2023-01-17 | Bank Of America Corporation | Analyzing documents using machine learning |
Also Published As
Publication number | Publication date |
---|---|
CN1826597A (en) | 2006-08-30 |
KR20060006945A (en) | 2006-01-20 |
US20040215606A1 (en) | 2004-10-28 |
JP2006524869A (en) | 2006-11-02 |
WO2004097568A3 (en) | 2006-01-05 |
US7197497B2 (en) | 2007-03-27 |
EP1623298A2 (en) | 2006-02-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7197497B2 (en) | Method and apparatus for machine learning a document relevance function | |
US6560600B1 (en) | Method and apparatus for ranking Web page search results | |
US7356530B2 (en) | Systems and methods of retrieving relevant information | |
EP1596313B1 (en) | Method and system for schema matching of web databases | |
US8655888B2 (en) | Searching documents for ranges of numeric values | |
CA2618854C (en) | Ranking search results using biased click distance | |
US6112203A (en) | Method for ranking documents in a hyperlinked environment using connectivity and selective content analysis | |
CN1702654B (en) | Method and system for calculating importance of a block within a display page | |
JP4908214B2 (en) | Systems and methods for providing search query refinement. | |
US6792419B1 (en) | System and method for ranking hyperlinked documents based on a stochastic backoff processes | |
US20090089274A1 (en) | Gradient based optimization of a ranking measure | |
US20030018617A1 (en) | Information retrieval using enhanced document vectors | |
WO2005024661A2 (en) | Improved search engine optimisation | |
Inkpen | Information retrieval on the internet | |
Sathiya et al. | A Study of Various Result Merging Strategies for a Meta Search Engine | |
Chaudhary | Context driven approach for extracting relevant documents from WWW | |
Dupret et al. | Information retrieval and ranking on the web: Benchmarking studies I | |
King et al. | Information retrieval and ranking on the Web: Benchmarking studies II | |
Chen et al. | Real-time adaptive feature and document learning for Web search | |
Park et al. | User Intention based Personalized Search: HPS (Hierarchical Phrase Serch) | |
Bhushan et al. | Effective Information Retrieval Method Based on Hierarchical Agglomerative Clustering | |
Usha et al. | Analytical survey of Web Page Rank Algorithm | |
Du | A Web Meta-Search Engine | |
Rani et al. | Web Search Result using the Rank Improvement | |
GÜLTEPE et al. | Application of PageRank Algorithm in Linked Data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A2 Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BW BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE EG ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NA NI NO NZ OM PG PH PL PT RO RU SC SD SE SG SK SL SY TJ TM TN TR TT TZ UA UG US UZ VC VN YU ZA ZM ZW |
|
AL | Designated countries for regional patents |
Kind code of ref document: A2 Designated state(s): BW GH GM KE LS MW MZ SD SL SZ TZ UG ZM ZW AM AZ BY KG KZ MD RU TJ TM AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LU MC NL PL PT RO SE SI SK TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
WWE | Wipo information: entry into national phase |
Ref document number: 2006513331 Country of ref document: JP Ref document number: 1020057020306 Country of ref document: KR |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2004750656 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 20048174686 Country of ref document: CN |
|
WWP | Wipo information: published in national office |
Ref document number: 1020057020306 Country of ref document: KR |
|
WWP | Wipo information: published in national office |
Ref document number: 2004750656 Country of ref document: EP |