JP2011118689A - Retrieval method and system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To express relations between entity pairs from huge text data of WWW or the like, and to achieve relation retrieval using the relations between the entity pairs. <P>SOLUTION: The method for retrieving an entity X having a relation the same as or similar to the relation between entities A, B between the entities C, X of the entity pair (C, X) when the entity pair (A, B) and the entity C are input. By defining the relation between the entity A and the entity B depending on context around the entities A, B in a text including the entity A and the entity B and searching the entities C, X having the context the same as or similar to the context around the entities A, B, the entity X having the relation the same as or similar to the relation of the entity A and the entity B with the entity C is acquired. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

The present invention relates to a search method and system, and more particularly, to a search technique using a relationship between entities in two entities forming a pair. In this specification, an “entity” is one or a plurality of words (ie, phrases), and is typically a single noun or a noun phrase composed of a plurality of words.

In recent years, data on the WWW is increasing enormously. There are many relationships between entities in the vast amount of information. A conventional keyword-based search engine can receive a keyword and find a sentence including the keyword, but cannot search a relationship between entities. A conventional keyword-based search engine cannot actively use a lot of related information potentially existing in a large amount of information on the WWW.

As a system for extracting the relationship between two given entities, there is a TextRunner system shown in Non-Patent Document 1. When TextRunner receives an entity pair (A, B) as an input, it can search for a predicate between A and B. For example, if an entity pair (Bush, United States) is input, an is predicate is (George W. Bush, President of United States) is output. In other words, the relationship between Bush and United States (is president of) is not directly output, but a general predicate that holds between entities including Bush and United States (George W. Bush, President of United States) ( is) only can be output. Thus, TextRunner cannot find a relationship between two entities and cannot perform a relationship search using TextRunner.

The TextRunner system is described in Patent Document 1 as an extraction pattern learning method based on unsupervised learning. In this method, first, a parser is run on a small corpus (a set of texts) to parse sentences in each document. Then, from the analysis result, a noun is taken and the relationship between the noun and the noun is obtained from the syntax tree. Use heuristics to determine if the relationship is to be extracted and create a sample for learning. For example, using the syntax tree of the sentence “Oppenheimer taught at Berkeley and Caltech”, the patterns (Oppenheimer, taught at, Berkeley) and (Oppenheimer, taught at, Caltech) are positive samples, (Berkeley, and, Caltech) is judged as a negative sample. Na &iuml; ve Bayes classifier is made using these samples as training data. In other words, when a relationship is given, the classifier determines whether the relationship is good (correct) or bad (incorrect). The characteristics used in the classification process are whether the two nouns are proper nouns, the distance in the sentence, and the part of speech on the left and right. The classifier learns from a small corpus and has the advantage that it takes less time to use a parser. Then, after learning the classifier, it can be determined using the classifier whether the relationship extracted from the enormous corpus is a good (correct) relationship. This approach can extract good (correct) relationships and may be a good basis for the present invention. However, the method described in Patent Document 1 only provides a relationship extraction and search method. That is, there are a method for extracting a relationship between words and a method for searching for a related or unknown word when a word or relationship is given. In this method, the similarity of the relationship between word pairs is not handled, and the latent relationship search engine of the present invention cannot be realized. Further, since the search of Patent Document 1 has a different purpose from the present invention, the indexing method and search algorithm are also completely different from the present invention. Furthermore, by using a syntax analyzer to create learning data, it becomes language dependent, and there is a high possibility that patterns cannot be extracted well for languages with few language resources (languages without a syntax analyzer).

Several methods for measuring the similarity between two word pairs have been proposed. For example, in the LRA (latent relational analysis) method proposed in Non-Patent Document 2, when two entity pairs ((A, B), (C, D)) are given, the interval between (A, B) Automatically generate lexical patterns frequency vectors that characterize the relationship using a keyword-based search engine. Also, (C, D) is processed in the same manner, and the relationship similarity between (A, B) and (C, D) is measured by measuring the cosine similarity of the feature vectors of (A, B) and (C, D). You can measure the degree. Further, Non-Patent Document 3 proposes a method for creating a word pair feature vector at high speed from the WWW and a feature vector dimension reduction method for improving the accuracy of the relational similarity.

Although the methods disclosed in Non-Patent Document 2 and Non-Patent Document 3 are the basis of the present invention, the relationship similarity measurement system requires four complete entities of two entity pairs, and one entity is unknown. In this case, measurement cannot be performed, and the feature vector of an entity pair having an unknown entity cannot be output. Therefore, the functions of the search engine proposed in the present invention cannot be realized only by the systems proposed in those documents.

ここで、τ(w_i，w_j)は単語w_iとw_jの類似度で、δ(w)は単語wのWordNet階層における深さである。P_ijは単語w_i，w_jの共通上位語(hypernym)である。例えば、図１３の概念“abstract entity”の深さは1で、Indo-European，Greek，Englishの深さはそれぞれ5，6，8である。そこで、(English，Greek)の類似度は、

となる。
そして単語間の類似度を用いて、単語ペア間の類似度を次のように定義する。

αとβは重みで、より似ている単語ペアに優先するようにα=2，β=1とする。
この式を使うと、例えば、π((English，A)，(Greek，Zeus))＝0.52で、
π((English，A)，(Greek，Alpha))＝0.82である。つまり、(English，A)と(Greek，Alpha)との類似度は(English，A)，(Greek，Zeus)との類似度よりも大となる。 Non-Patent Document 4 proposes a method of measuring the similarity between word pairs using WordNet. WordNet is a conceptual dictionary as shown in FIG. Using this dictionary, the similarity between two words is defined as follows.

Here, τ (w _i , w _j ) is the similarity between the words w _i and w _j , and δ (w) is the depth of the word w in the WordNet hierarchy. P _ij is a common broad word (hypernym) of the words w _i and w _j . For example, the depth of the concept “abstract entity” in FIG. 13 is 1, and the depths of Indo-European, Greek, and English are 5, 6, and 8, respectively. Therefore, the similarity of (English, Greek) is

It becomes.
And the similarity between word pairs is defined as follows using the similarity between words.

α and β are weights, and α = 2 and β = 1 so that priority is given to more similar word pairs.
Using this formula, for example, π ((English, A), (Greek, Zeus)) = 0.52,
π ((English, A), (Greek, Alpha)) = 0.82. That is, the similarity between (English, A) and (Greek, Alpha) is greater than the similarity between (English, A) and (Greek, Zeus).

It is also conceivable to implement a relational search using the method described in Non-Patent Document 4. That is, when there is a query (English, A) and (Greek,?), The path from English to A is searched. Then, a path (Greek,?) That has the longest common path with the path is searched. As shown in FIG. 13, when this method is used, the answers to the queries (English, A) and (Greek,?) Are not (Greek, Zeus), but (Greek, Alpha), (Greek, Beta), You can give an answer like ... However, that alone cannot give the correct answer "Alpha".

Furthermore, there is an “Analogical Thesaurus” proposed in Non-Patent Document 5 as a proposal close to the relationship search. An analogy thesaurus is a hierarchical concept dictionary for finding word pairs in a similar relationship. In WordNet, as shown in Figure 11, both English and Greek characters are classified into the {LETER, ALPHABETIC_CHARACTER} concept, and there is no distinction between them, and queries {(English, A), (Greek,?) } Cannot answer.

On the other hand, the thesaurus proposed in Non-Patent Document 5 introduces a concept that divides WordNet in more detail. As shown in FIG. 12, new concepts {GREEK_LETTER}, {ENGLISH_LETTER} and {1 ^st LETTER} are introduced to divide the concept {LETTER, ALPHABETIC_CHARACTER} originally in WordNet in detail. Therefore, when a query {(English, A), (Greek,?)} Is input, the system looks up this dictionary and finds the position of the character {A}. And there are {ENGLISH_LETTER} and {1 ^st LETTER} as ancestor nodes at that position. Since the query wants to find the correspondence between English and Greek, {GREEK_LETTER} is a candidate for {ENGLISH_LETTER}. This is because {GREEK_LETTER} and {ENGLISH_LETTER} have an ancestor node {LETTER, ALPHABETIC_CHARACTER} in common. Thus, it can be seen that the answer candidates are all child nodes of {GREEK_LETTER}. Finally, in order to get the correct answer from the candidate set, among the child nodes of the node “A” and the node {GREEK_LETTER} of the query, the node {ALPHA} is the node {A} and the previous ancestor {1 ^st Since it shares LETTER}, the answer will output node {ALPHA}. Therefore, the output of the query {(English, A), (Greek,?)} Is “ALPHA” (character α). As you can see in the output derivation process, the reason is that A is the first letter of English (1 ^st letter) and α is the first letter of Greek. Therefore, the thesaurus proposed in Non-Patent Document 5 can directly handle the relationship search.

However, using this method, the capabilities of the search system are very limited. Because it is necessary to introduce new concepts (concepts for detailed division) into WordNet in order to be able to search accurately, there is a method to automatically derive those concepts, but the accuracy is not good, and this is the final result This is a major cause of reducing the recall rate. What's worse, you can't search for primitive concepts that don't exist in WordNet (that is, concepts that don't have child nodes). For example, the WWW frequently introduces new concepts (person names, organization names, etc.), but the analogy thesaurus cannot handle this.

M. Banko and O. Etzioni. TheTradeoffs Between Open and Traditional Relation Extraction, Proceedings of ACL, pp. 28-36 (2008). Peter D. Turney. Measuring SemanticSimilarity by Latent Relational Analysis.Proceedings of International JointConf.on Artificial Intelligence (IJCAI), pp.1136-1141 (2005). Danushka Bollegala, Yutaka Matsuo and Mitsuru Ishizuka: Measuring the Similarity between Implicit Semantic Relations from the Web, Proceedings of the 18th International World Wide Web Conference (WWW 2009), pp. 651-660 (2009). Tony Veale. WordNet Sits theS.A.T.- A Knowledge-Based Approach to Lexical Analogy.Proceedings of the 16th European Conf. On Artificial Intelligence (ECAI), pp.606-612 (2004). Tony Veale.The AnalogicalThesaurus.In Proceedings of the Fifteenth Conference on InnovativeApplications of Artificial Intelligence (IAAI), pp.137-141 (2003).

US2008 / 0243479A1.

As described above, existing systems such as TextRunner cannot express the relationship between entity pairs even if there is a text corpus. In addition, the analogy thesaurus can only perform a limited relationship search and cannot use a huge amount of WWW data. Several methods for measuring the relational similarity have been proposed, but the relational search cannot be realized directly.

An object of the present invention is to express a relationship between entity pairs from a large amount of text data such as WWW, and to realize a relationship search using the relationship between entity pairs.

The first technical means adopted by the present invention is that when an entity pair (A, B) and an entity C are input, the entity pair between the entity C (X) and the entity A (B) A method for searching for an entity X having the same or similar relationship as the relationship using a search database or / and an existing search engine,
Entity pair (A, B) and entity C are received as a query,
Defining the relationship between entity A and entity B depending on the context around entities A and B in the text containing entity A and entity B;
By searching for the entities C and X having the same or similar context as the surroundings of the entities A and B, the relationship between the entities A and B is the same as or similar to the relationship between the entities A and B. It is a relationship search method for acquiring an entity X provided.

An “entity” is one or more words or phrases.
The present invention is a technique for ranking (C, X) that has a large relational similarity (A, B) and (C, X) based on the relational similarity.
In the present invention, the position of the entity X is not limited. For example, the entity pair (X, D) can be handled in the same manner as the entity pair (C, X). That is, when an entity pair (A, B) and an entity D are input as a query, the entity pair (C, X) is replaced with an entity pair (X, D), and the entity C is replaced with an entity D. When the entity pair (A, B) and the entity D are input, the entity having the same or similar relationship as the relationship between the entities A and B between the entities X and D of the entity pair (X, D) It functions as a method for searching for X.

  In one aspect, the context is one or any combination of vocabulary patterns, bag-of-words, part of speech patterns, dependency patterns.

  In one aspect, the contexts are clustered and each context cluster includes substantially the same or similar context.

  In one aspect, the relationship between two entities in a pair of entities is represented by a feature vector, which is obtained from a context or context cluster around the two entities in the entity pair.

In one aspect, the relationship between entity A and entity B is represented by a first feature vector, which is obtained from a context or context cluster around entities A and B;
The step of acquiring the entity X is to acquire the entity X having a second feature vector that is the same as or similar to the first feature vector with the entity C, and the second feature vector is the entity C And the context or context cluster around the entities C and X in the text including the entity X.

In one aspect, obtaining the entity X comprises
The similarity between the pair (A, B) and each pair (C, X) is calculated based on the distance between both feature vectors, and the calculated similarity between the pairs (relationship similarity) is used as an index. Sorting a plurality of pairs (C, X) in descending order of similarity between pairs and ranking X candidates in the order of sorting;
Outputting a part or all of the ranked X as a search result;
It has.
Typically, since the relationship similarity is calculated by the distance between feature vectors, it is necessary to create a feature vector first. In one aspect, the feature vector can be defined by “weight based on pattern frequency”.
In a typical aspect, the relationship similarity is calculated based on the cosine similarity between both feature vectors, but other distance measures may be used.

In one aspect, the context is a vocabulary pattern, and the feature vector has the frequency of the vocabulary pattern as an element. That is, the relationship between the entity pairs is expressed by a feature vector having a feature value of “weight calculated based on the vocabulary pattern frequency”.
More specifically, the elements of the feature vector are “feature (dimension)” and its “value (feature amount)”. When the context is a vocabulary pattern, “feature (dimension)” = “vocabulary pattern” ”,“ Value (feature amount) ”=“ vocabulary pattern frequency ”.
When clustering vocabulary patterns, the dimension of the feature vector is “vocabulary pattern cluster”, and the value of each dimension is “sum of frequency of vocabulary patterns included in the vocabulary pattern cluster”.

In one aspect, obtaining the entity X comprises
Acquire a plurality of pairs (C, X) having a second feature vector whose elements include the vocabulary pattern of the first feature vector of the acquired pair (A, B) (frequency is greater than a preset threshold). Steps,
The similarity between the pair (A, B) and each pair (C, X) is calculated using both feature vectors, and the calculated similarity between the pairs is used as an index. Sorting the pairs (C, X) and ranking the candidates for X in the sort order;
Outputting a part or all of the ranked X as a search result;
It has.

In one aspect, the search database stores correspondences between a large number of entity pairs and surrounding contexts expressing relationships between the entities of each entity pair.
In one aspect, the context is a vocabulary pattern, and the search database stores correspondence between a large number of entity pairs, each entity pair, the vocabulary pattern and the frequency of the vocabulary pattern.

In one aspect, the search database is
A first index defining the ID of each entity in the entity pair;
A second index defining entity pair IDs corresponding to two entity ID pairs;
A third index defining a vocabulary pattern ID corresponding to each vocabulary pattern;
A fourth index that defines the correspondence between the entity pair ID and the vocabulary pattern ID / frequency of the vocabulary pattern;
With
An entity pair ID corresponding to the input entity pair (A, B) is acquired, and using the acquired entity pair ID, a vocabulary pattern ID corresponding to the entity pair (A, B) (the frequency of the vocabulary pattern is determined in advance). Obtaining a frequency of the vocabulary pattern to form a first feature vector;
Obtaining a candidate entity pair ID including the vocabulary pattern ID of the first feature vector (the frequency of the vocabulary pattern is greater than a preset threshold) and the ID of entity C;
Obtaining a second feature vector corresponding to each candidate entity pair ID and obtaining an entity pair (C, X) such that the second feature vector is similar to the first feature vector;
It has.
In one embodiment, the frequency threshold of the vocabulary pattern is frequency 1. Actually, the low frequency often includes noise such as spelling mistakes, and it is desirable to appropriately select and set a threshold value in advance.
In one aspect, the fourth index is
An index for searching the frequency of the vocabulary pattern ID / vocabulary pattern from the entity pair ID;
An index for searching the frequency of entity pair ID / vocabulary pattern from the vocabulary pattern ID;
It has.

In one mode, in the search database, the vocabulary patterns form vocabulary pattern clusters, and each vocabulary pattern cluster includes similar or substantially identical vocabulary patterns.
In one aspect, the step of acquiring the entity pair (C, X) includes the same vocabulary pattern (frequency greater than a preset threshold) as the vocabulary pattern of the feature vector of the entity pair (A, B) as an element. Vocabulary patterns belonging to the vocabulary pattern cluster to which the vocabulary pattern of the feature vector of the entity pair (A, B) belongs as well as the entity pair (C, X) with the feature vector A step of searching for an entity pair (C, X) having a feature vector as an element.

In one aspect, an entity cluster composed of similar entities is formed in the search database.
In one aspect, the score of the entity cluster to which the entity X belongs is used as a score that serves as an index for ranking the candidate of the entity X.
In one aspect, in the output step, a plurality of entities included in the entity cluster are output. The score of the entity cluster is a cluster average score obtained by normalizing the sum of the scores of the entities included in the entity cluster divided by the entity cluster size.

In one aspect, a feature vector representing a relationship between entities of an entity pair is
The sentence acquired from the text corpus is divided into words,
Create candidate entity pairs from the segmented words,
Obtained from the context around the candidate entity pair.

In one aspect, the element of the feature vector is a vocabulary pattern frequency;
Extracting the frequency of vocabulary patterns
In the original sentence including the candidate entity pair, the first entity of the entity pair is replaced with X, the second entity is replaced with Y, and a partial sequence S including X and Y is generated, and an n-gram ( generating n) from 1 to K;
Using the obtained n-gram as one dimension of a feature vector, counting the frequency in all sentences including entity pairs;
With
The n-gram frequency is stored as the vocabulary pattern frequency in association with the entity pair.
Here, the partial sequence S is a word sequence composed of m ₁ words immediately before X, a word sequence between X and XY (excluding X and Y), and m ₂ words immediately after Y and Y. is there. m ₁ , m ₂ , and K are parameters (integers).
When the length of a word string between XY (excluding X and Y) is D, in one aspect, if D is greater than K-2, then the sequence between [X, (X, Y), Y] n-gram is also generated.
In one aspect, the context around the candidate entity pair is acquired only when the distance D between the two entities in the candidate entity pair is equal to or less than a threshold value _Dth . This distance D is the number of words.
In one aspect, in the step of creating the candidate entity pair, entity pairs including proper nouns are preferentially extracted as candidate entity pairs.
In one aspect, in the step of creating the candidate entity pair, a frequent entity pair is preferentially extracted as a candidate entity pair.
One aspect includes forming clusters of vocabulary patterns from similar or substantially identical vocabulary patterns.

In one aspect, the search method of the present invention uses an existing search engine to obtain a context around entities A and B in text including entity A and entity B;
Using existing search engines to obtain entities C, X having the same or similar context as the surrounding context of entities A, B;
It has.
The processing inside the search system changes depending on whether or not an entity given as a query exists in the search database. If one or more of the entities A, B, and C do not exist in the search database, the existing keyword-based Web search engine is used to dynamically extract a vocabulary pattern that represents the relationship between the entities A and B, and In C, the entity X associated with the vocabulary pattern is extracted and ranked.

In one aspect, the context around the entities A and B is obtained from a snippet as a search result of the search engine using the entities A and B as a query with the distance between the entities A and B limited. To do.
Specifically, the search here is a query created by inserting a predetermined number of * (*: wildcard corresponding to 0 or 1 word) between entities A and B (eg, “A *”). * * Search using B ”). In a normal keyword search engine, when entities A and B are entered as a query, a snippet including entities A and B can be obtained. In this wildcard search, only entities A and B are searched. It can be said that the input between the queries is to limit the distance between the entities A and B. Thus, a search with a query (eg “A *** B”) yields a snippet that also includes the contexts before and after entities A and B, and the vocabulary that includes the contexts before and after (outside) entities A and B. A pattern can be generated. The number of “*” is appropriately set in advance by those skilled in the art.

In one aspect, the context around the entities A and B is the snippet (“A
* * * This is a vocabulary pattern extracted from a snippet obtained using B ”as a query.
Entity X candidates are obtained from snippets as query search results obtained by substituting entity C for the extracted vocabulary pattern.
When the initial search query is (A, B), (C,?), The query here is entity C, vocabulary pattern (pattern),? The query is “C pattern *” that retains the order. If the initial search query is (A, B), (?, D), what is the query here? , Entity D and query “* pattern D” holding the order of vocabulary patterns.
If you enter the query “C pattern *”, many existing search engines will not return only “C pattern W1” (W1 is a single word (letters in Japanese)), but will match that query. For a text snippet, the context is also returned. Therefore, “… C pattern W1 W2 W3…” (where * corresponds to W1) is actually returned. Therefore, the candidate for D can be W1 or W1 W2 or W1 W2 W3.

The second technical means adopted by the present invention is that when an entity pair (A, B) and an entity C are input, an entity pair (C, X) between entities C and X is between entities A and B. A system for searching for an entity X having the same or similar relationship as
The search database stores correspondences between a large number of entity pairs and surrounding contexts that express relationships between entities of each entity pair,
Means for receiving entity pair (A, B) and entity C as a query;
Means for obtaining, using the search database, a first peripheral context expressing a relationship between entity A and entity B;
Means for obtaining entity C by searching for an entity C, X having a second peripheral context identical or similar to the first peripheral context using the search database;
A search system.

In one aspect, the context is a vocabulary pattern, and the search database stores correspondence between a large number of entity pairs, each entity pair, the vocabulary pattern and the frequency of the vocabulary pattern.
In one aspect, the search system comprises:
A plurality of entity pairs (C, X) having a second feature vector whose elements include the vocabulary pattern of the first feature vector of the acquired entity pair (A, B) (frequency is greater than a preset threshold) Means to obtain,
Means for calculating the similarity between the pair of the entity pair (A, B) and each entity pair (C, X) using both feature vectors;
A means for sorting a plurality of entity pairs (C, X) in descending order of the similarity between pairs using the calculated similarity between pairs as an index, and ranking X candidates in the order of sorting;
Means for outputting a part or all of the ranked X as a search result;
It has.

In one aspect, the search database is
A first index defining the ID of each entity in the entity pair;
A second index defining entity pair IDs corresponding to two entity ID pairs;
A third index defining a vocabulary pattern ID corresponding to each vocabulary pattern;
A fourth index that defines the correspondence between the entity pair ID and the vocabulary pattern ID / frequency of the vocabulary pattern;
With
An entity pair ID corresponding to the input entity pair (A, B) is acquired, and using the acquired entity pair ID, a vocabulary pattern ID corresponding to the entity pair (A, B) (the frequency of the vocabulary pattern is determined in advance). (Greater than the set threshold value)-The frequency of the vocabulary pattern is acquired to form the first feature vector,
Obtaining a candidate pair ID including the vocabulary pattern ID of the first feature vector (the frequency of the vocabulary pattern is greater than a preset threshold) and the ID of the entity C;
A second feature vector corresponding to each candidate pair ID is acquired, and an entity pair (C, X) such that the second feature vector is similar to the first feature vector is acquired.
In one aspect, the fourth index is
An index for searching the frequency of the vocabulary pattern ID / vocabulary pattern from the entity pair ID;
An index for searching the frequency of entity pair ID / vocabulary pattern from the vocabulary pattern ID;
It has.

In one aspect, the search database stores vocabulary pattern clusters, and each vocabulary pattern cluster includes similar or substantially identical vocabulary patterns.
In one aspect, the search database stores entity clusters, and each entity cluster includes similar or substantially identical entities.

  According to the present invention, a relationship between useful entities is expressed by a feature vector from a huge text corpus, and based on this, a relationship search can be practically realized. In addition, the method of the present invention can deal with a huge amount of data sources. In other words, a relational search can be performed in an information space such as a huge WWW, and so-called large-scale Web-scale can be handled.

  The indexing technique by creating a search database according to the present invention is important for reducing the search time. By creating a search database using indexing technology, this system can execute relational searches with high latency and low latency.

  Various problems can be solved using the relationship search engine of the present invention based on latent relationships. For example, in a question answering system, the system teaches an answer to an input question, but it is difficult to give an answer to knowledge that does not exist in the system knowledge base. If the search system of the present invention is used, a wide range of questions can be answered even if there is little knowledge. For example, suppose that there is knowledge that “Tokyo is the capital of Japan” in advance (for example, there is a first-order predicate logical expression of capital (Japan, Tokyo) in the knowledge base). At that time, the answer to the question "What is the French capital?" Is the result of the query {(Japan, Tokyo), (France,?)} And the knowledge of the French capital (logical formula: capital ( Paris, France)) can give answers just by using the WWW or text corpus even if the system does not have it in advance.

  In addition, the search engine can easily search for similar words, and can also be used to support the creation of a thesaurus. For example, if {(A, B), (C,?)} Is entered as a query, and the relationship of (A, B) is a similar word, it is possible that D obtained as a search result is a similar word of C High nature. Similarly, not only similar terms but also broader terms (hypernyms, hyponyms), opposite terms and partially related terms (meronyms) can be searched.

  Further, as an application that can be used immediately, there is a product manufacturer search and the like. For example, a user knows that “Shiseido is famous for cosmetics” on an e-commerce website, but wants to find out which manufacturer is better for a game machine. At that time, if you enter the query (cosmetics, Shiseido) and (game machine,?) In this system, you can get a list of famous game machine manufacturers.

By performing relational searches in an enormous information space such as the World Wide Web, the latest data (knowledge) can be processed, and new concepts and information can be handled. In addition, a plurality of candidates can be presented to the user for the query.
Therefore, the application range of the search engine developed in the present invention is wide.

It is a figure which shows the concept of the search method and system of this invention. It is a figure which shows the search screen of the search method and system of this invention. It is a display screen of simple results (results that do not include evidence) of queries (steve jobs, apple), (?, Microsoft). In the display screen of FIG. 2A, only a part of the evidence is shown. Some of the vocabulary patterns are shown in Debug Info. Query (steve jobs, apple), (steveballmer,?) This is a display screen for simple results (results not including evidence). It is a figure which shows the whole structure of the search method and system of this invention. It is a figure which shows the relationship between an entity pair and a feature vector. It is a figure which shows part-of-speech tagging. It is a figure which shows a response | compatibility of (a) entity and entity ID, (b) correspondence of entity pair and entity pair ID. Fig. 5 shows a correspondence table between vocabulary patterns and vocabulary pattern IDs. A table storing vocabulary patterns including the entity pairs and frequency information from the entity pairs is shown. 9 is a table expressing FIG. 8 by ID. A table of mapping from a vocabulary pattern ID to an entity pair ID and frequency list including the vocabulary pattern as one component of a feature vector is shown. It is a figure which shows WordNet (concept dictionary) (nonpatent literature 5). It is a figure which shows after the introduction of the concept which divides WordNet (concept dictionary) in detail (nonpatent literature 5). It is a figure which shows the path | pass in WordNet (concept dictionary).

[A] Overview of Search Method and System In the search system according to the present invention, as shown in FIG. 1, two entity pairs {(A, B), (C,?)} Are input as queries, and “?” As to find the appropriate entity. The output “?” Of the search system is an entity X that becomes (C, X) having a relationship similar to the relationship between the entities A and B of (A, B). Typically, entity X is obtained as a ranked list. For example, when three entities (words) {(Japan, Tokyo), (France,?)} Are given as input, this search system outputs “Paris” as “?” As the top result. To do. In other words, the relationship between “Japan” and “Tokyo” is very similar to the relationship between “France” and “Paris”. In this specification, the input format is described as {(A, B), (C,?)} For convenience, but in the present invention, the symbol “?” Can be set at any position of each entity pair. . The query of this search system may be in an input format such as {(?, B), (C, D)}, for example.

In order to answer the query as described above, it is necessary to extract the relationship between the entity A and the entity B and search the database for another entity pair (C, X) having the relationship. However, the WWW and text corpus to be searched by this search engine are only large text data, and the relationship between entities is not prepared in advance. Therefore, the relationship between the entity A and the entity B in the entity pair (A, B) is not specified in advance, but is potential. In the present invention, such a latent relationship is expressed by a feature vector, and a relationship search is realized by measuring a relationship similarity between two entity pairs, that is, a latent relational similarity.

In the present invention, if there is a potential relationship between entities A and B of a given entity pair (A, B) (referred to as “stem word pair”), the given third entity (C) Then, an entity X having a similar latent relationship between the entities C and X of the entity pair (C, X) or (X, C) is searched from the WWW and the text corpus. Since the problem of searching for (X, C) and the problem of searching for (C, X) are essentially the same, the search for (C, X) is assumed below.

First, a feature that expresses a potential relationship between entities A and B of a given stem word pair (A, B) is found from a text corpus. The relationship between the entities A and B can be extracted based on the context around the entities A and B in the text including the entities A and B. There are several ways to understand the surrounding context. One or any combination of lexical patterns, bag-of-words, part-of-speech patterns, dependency patterns around entities A and B Can be illustrated. In one aspect, these peripheral contexts (vocabulary patterns, bag-of-words, part-of-speech patterns, dependency patterns, etc.) are acquired in advance for a large number of entity pairs, and the entity pairs and peripheral contexts are associated with each other as a database. Save it. Alternatively, the surrounding context can be obtained on-the-fly when querying.

A lexical pattern is a sequence of vocabulary that maintains order in a sentence. For example, an vocabulary n-gram that appears in a sentence is a vocabulary pattern. Bag-of-words is a technique that treats a set of words without being caught in word order.

Briefly explain the part of speech pattern. For example, in the sentence Obama is the president of the US,
Obama / PERSON
is / VBZ the / DT president / NN of / IN the / DT US / LOCATION
Here, the relationship between Obama and US (entity pair) when interpreted as VBZ: verb-3rd person singular, DT: determiner, NN: noun (single or non-countable noun), IN: predicate or subordinator. (Obama, US) features)
One vocabulary pattern is "is the president of the"
One part-of-speech pattern is “VBZ DT NN IN DT”.
If the part-of-speech pattern is the same, there is a possibility that the relationship between the entities of the entity pair is similar. For example, the part of speech pattern of the above sentence and the following sentence is the same.
Obama is the leader of the US

The dependency pattern is a pattern resulting from dependency analysis in the sentence. Since the result of the dependency analysis represents the role and relationship in the sentence of each entity (word), if the dependency pattern is the same, the entity pair may also be related. For example,
Obama is the president of the US,
Junichiro Koizumi was the prime minister of Japan,
The sentence part pattern is not exactly the same, but the dependency pattern:
PERSON <-subject--VERB (be)-object->NOUN--of-> LOCATION,
Are common, (Obama, US), (Junichiro Koizumi, Japan) is considered to have some relationship. In the above expression, the dependency is represented by an arrow, and the arrow label is the name of the relationship. That is, PERSON (Obama, Junichiro Koizumi) has a relationship that depends on the verb (be), and the noun (president, prime minister) also depends on the verb. Furthermore, LOCATION (US, Japan) depends on the middle noun (president, prime minister).

In the present invention, typically, the vocabulary pattern around the entity pair is used as the context around the entity pair, but a part of speech pattern or a dependency pattern may be used. However, considering the problem of processing time and accuracy when performing advanced language processing (dependency analysis) on the vast and noisy data on the Web, it is advantageous to use only vocabulary patterns. For example, when searching in the dynamic mode in the present invention, the snippet of the keyword-based search engine output is often not a complete sentence, so the morpheme and syntax analysis may not be successful.

In one aspect, lexical patterns in a sentence in which entities A and B appear simultaneously are used as features that express the relationship between entities A and B. For example, when there is a phrase “American President Obama” in the corpus, extract the vocabulary pattern “President of” and use this pattern between “America” and “Obama” in the entity pair (America, Obama). One of the patterns characterizing the potential relationship of Since there may be a plurality of patterns that characterize the relationship between the entities of the entity pair, the characteristic pattern is expressed as a frequency vector. In this method, a potential relationship between entities in an entity pair is expressed as a feature vector (in this embodiment, a vocabulary pattern frequency vector) from the Web or a text corpus.

When a stem word pair is given, an entity pair that has a feature vector that is very similar to the feature vector that expresses the relationship between the entities of the stem word pair and that has the third entity C as one element Find out. Another element of the found entity pair (an entity that is not C) is a candidate answer. The similarity between the candidate entity pair and the stem word pair is measured, sorted by similarity, and the result is output from the sorted list.

In one aspect of the present invention, the dimension of the feature vector (eg, vocabulary pattern) is further reduced by a clustering technique to solve the noise and data pattern problem (data sparseness problem).

If the entity you want to search for is a proper noun, there may be multiple ways to express it (for example, “United States”, “US”, “USA” are all proper nouns that refer to the United States. (The expression format is different.) In one aspect of the present invention, a clustering technique is used for entities in a plurality of expression forms (surface forms), and different expression forms are handled as one entity, thereby improving search accuracy and recall. .

The search system according to the present invention (a specific configuration example is shown in FIG. 3) is composed of one or a plurality of computers. The computer includes processing means (CPU and the like) as hardware, storage means ( Hard disk, RAM, ROM, etc.), input means, output means or display means, and a control program for operating the computer as software. The user terminal is also composed of one or a plurality of computers, and the computer includes processing means, storage means, input means, output means or display means, a control program for operating the computer, and the like. The search system and the user terminal are provided with transmission / reception means that can exchange information with each other via a computer network represented by the Internet, and are connected to each other via a computer network represented by the Internet. Yes. The search system is connected to an existing keyword-based search engine via a computer network represented by the Internet. For example, a search screen as shown in FIG. 2 is displayed on the screen of the user terminal, and an entity is input from the input means of the user terminal, and is transmitted to the search system as a search query. In the search system, the search result is calculated based on the search query and can be viewed from the user terminal.

[B] Specific content of search method and system: lexical pattern generation, vocabulary pattern clustering, similarity calculation between entity pairs, indexing of entity pairs, search by relationship similarity, query using remote corpus The dynamic processing will be described.

[B-1] A feature vector for expressing a relationship between entities of an entity pair is defined from a lexical pattern generation text corpus, and the feature vector is specifically created. The element of the feature vector is the frequency of lexical patterns that express the relationship between entities. For example, when a sentence containing “American President Obama” appears 10 times and a sentence containing “American Congressman Obama” appears 8 times, the feature vector of the entity pair (America, Obama) is the column vector V1 in FIG. It is expressed as The frequency of vocabulary patterns not included in the pair is 0.

There are several ways to create feature vectors. First, there is a text corpus in advance, and by analyzing it, a feature vector can be created in advance (before a query). Second, even without a text corpus, using a normal keyword-based search engine, like “America *** Obama” (where the “*” operator is a wildcard that matches one or zero words) A query can be sent to a keyword-based search engine, lexical patterns can be extracted from the resulting snippet results, and feature vectors can be created. That is, when a query is known, relationships between entities related to the query may be extracted online (in real time) using an existing keyword-based search engine.

In generating feature vectors, vocabulary patterns between entities are particularly important. In order to perform a relationship search with high accuracy, it is important to more accurately extract a vocabulary pattern expressing the relationship in a context including the relationship between entities. The lexical pattern query preprocessing will be described below.

The corpus for query preprocessing is a local corpus (such as a set of texts on a local disk or a WWW page) in the local hard disk of the search system. Such a text corpus contains a large number of entities and potential relationships between them.

These texts are first cut into sentences. Then, divide those sentences into words and add parts of speech. This process is known as part-of-speech (POS) tagging. For example, as shown in FIG. 5, when there is a sentence “Tokyo is the capital of Japan”, after tagging part-of-speech tagging, <Tokyo / noun, ha / particle, Japan / noun, no / particle, capital / A sequence of nouns, de / auxiliary verbs, certain / auxiliary verbs> is obtained. From the obtained word and part-of-speech string, nouns, verbs, adjectives, and adjective verbs are taken and those entities are set as candidates for entity pair elements. In the above sentence, “Tokyo”, “Japan” and “Capital” are extraction candidate entities.

Create candidate entity pairs from candidate entities. A candidate entity pair is two candidate entities that maintain the order in the sentence. For example, in the case of the above sentence, three entity pairs of (Tokyo, Japan), (Japan, capital), and (Tokyo, capital) are set as candidate entity pairs. If a candidate entity pair appears in the sentence, record that entity pair and increase the frequency. Also, the position of the sentence in which the pair appears (document ID and the position of the sentence in the document) is recorded.

In order to preferentially extract proper nouns as candidate entities, NER (Named
Entity Recognizer) is used to extract proper nouns in the sentence. If there is a proper noun in the entity pair, record it.

In order to handle multiple forms of proper nouns (United States, US, USA,...) As a single entity, the expression methods are clustered. That is, after clustering, United States, US, and USA are included in one cluster. To that end, we examine which entity an expression form pairs with. Further, the lexical pattern of the entity pair including the expression form is also examined. It is determined that an expression form having a vocabulary pattern similar to that of the partner entity falls within one cluster. For example, the United States, US, and USA expressions often form (United States, Barack Obama), (US, Barack Obama), (USA, Barack Obama) entity pairs together with "Barack Obama". Furthermore, “President” and “leader” appear in common in the vocabulary pattern of the entity pair. Therefore, “United States”, “US”, and “USA” are likely to be one entity. If a candidate entity is found as “US” during a search, the recall rate can be increased by including entities of “United States” and “USA” as candidates. Similarly,
About Steve Ballmer
(Steve Ballmer, Microsoft): 50 appearances
(Steve Ballmer, Bill Gates): 10
(Steve Ballmer, Microsoft Corp): 8
...
About Ballmer
(Ballmer, Microsoft): 20
(Ballmer, Bill Gates): 15
(Ballmer, Gates): 10
...
If there is a search result, Steve Ballmer and Ballmer are clustered so as to be in one cluster because they are similar. In entity clustering, a dictionary of synonyms may be stored in a database and used depending on the type of entity. In addition, when presenting the search result to the user, it is also easy for the user to understand if a cluster of entities (proprietary nouns) (a plurality of different expression formats belonging to the cluster) is displayed.

In order to count the frequency of the candidate entity pairs in the entire corpus, the above processing is executed on the entire corpus. After extracting candidate entity pairs, the candidate entity pairs and the frequency of those entity pairs appearing in the corpus are known, as well as which entity pairs contain proper nouns. After this processing is completed, when new text enters the corpus, the frequency can be updated by sequentially performing the same processing on the text.

Extract vocabulary patterns using the obtained candidate entity pairs. At that time, lexical patterns of entity pairs including proper nouns and frequent entity pairs are preferentially extracted. Use a variety of filtering techniques (eg, using entity pair appearance frequency or distance between entities) to extract candidate entity pairs that can be found in noise or irrelevant entity pairs. Remove. For example, entity pairs that are infrequent (for example, 4 times or less) are recognized as noise and lexical pattern extraction is not executed. When new text is added to the corpus, if the frequency of the entity pair is more than four times, the entity pair is no longer noise and the lexical pattern is extracted as usual.

To determine the vocabulary pattern, first measure the distance D in the text between the two entities of the candidate entity pair. The distance D is simply the number of words between the two entities. For entity pairs whose distance D is greater than the threshold value D _th , the vocabulary pattern is not extracted because the relationship between the entities is low.

For an entity pair having a threshold value _Dth or less, after replacing the first entity of the entity pair with X and the second entity of the entity pair with Y in the original sentence, the following subsequence (word string) S including X and Y And all n-grams (n is from 1 to K) of the word string S are generated. Using the obtained n-gram as one dimension of the feature vector, the frequency is counted over the sentence including the entity pair. The feature vector has n-gram as a dimension, and its value is the frequency of n-gram.
Subsequence S = PreX X InXY Y PostY
Here, PreX is m ₁ word strings immediately before X (in the sentence), and PostY is m ₂ word strings immediately after Y. InXY is a word string between (X, Y). In one embodiment, the parameters m ₁ and m ₂ are 5. InXY is extracted as a candidate pair if 10 or less (i.e., the length D of InXY is not extracted to be greater than D _th.). The n-gram n is from 1 to K. In one aspect, K = 7. In one embodiment, when D is larger than K-2 (D> K-2, ie, D + 2> K), an n-gram of (X, InXY, Y) is also generated.

More specific description will be given below.
M words before and after an entity, all words between entities
(However, the distance D between entities is within D _th ) n-gram (n = 1, 2, 3,..., K).
For example,
They discussed with Barack Obama, president of the United States in December to find a solution to the problem.
If m ₁ = m ₂ = 2 and D _th = 10,
The subsequence S to be extracted is
S = discussed with Barack Obama , president of the United States in December
It becomes.
The extracted n-gram is
[When K = 7 (D = 4, D <K-2]
n = 4: discussed with X ,; with X, president; X, president of; ...; the Y in December
n = 6: discussed with X, president of ； with X, president of the ； X, president of the Y ； ...
n = 7: discussed with X, president of the; with X, president of the Y; X, president of the Y in ;, president of the Y in December
n = 2: discussed with; with X; X,;, president; president of; of the; the Y; Y in; in December
And as another example, the statement is:
“They discussed with Barack Obama , the first African American president of the United States in December to find a solution to the problem”
If m ₁ = m ₂ = 2 and D _th = 10, the extraction target column S = “discussed with X, the first African American president of the Y in December”
In this case, D = 8 (that is, D = 8 because of the length of “, the first African American president of the”). Further, assuming that K = 7, D = 8> K−2.
[When K = 7, D = 8> K−2]
The extracted n-gram is
All n-grams with n = 1, 2, .., 7 (generate all n-grams with n = 1 to 7 as in the above example) and the following pattern:
X, the first African American president of the Y (n = 8)
That is, when D is larger than K-2, in addition to all n-grams from n = 1 to K, a subsequence (X, InXY, Y) is also extracted as an n-gram (this subsequence is n-gram, where n is greater than K).
All of the generated n-grams are used as features representing relationships.
In one aspect, lexical pattern extraction is performed after a sentence is cut out. That is, a partial sequence is extracted from one sentence, and an n-gram is generated from the partial sequence. In one aspect, even if one or both of the entity pairs appear in the sentence, only one occurrence (X to Y) is replaced, but they may be replaced at the same time.

The parameters D _th , m ₁ , m ₂ , and K can be appropriately set as threshold values by those skilled in the art. In the above embodiment, D _th = 10, m ₁ = 5, m ₂ = 5, and K = 7. However, it is not limited to these. For example, when D _th = 10, vocabulary patterns / feature vectors are not extracted when ten or more entity pairs are separated from each other. Further, when the feature vector is calculated, the threshold value _Dth may be a dynamic threshold value (for example, dynamically changeable from 1 to 10). Further, a minimum D _th value may be recorded in the search database so that a desired vocabulary pattern can appear. For example, for the sentence “Obama is the president of the USA”, the minimum D _th for “X is the president of the Y” to appear is 5. If D _th = 4, the lexical pattern does not appear and the search algorithm is executed.

[B-2] Vocabulary pattern clustering There are many vocabulary patterns that express relationships between entities in an entity pair, and there may be few vocabulary patterns that completely match. For example, the relationship “President of” can be expressed in the form of “Country leader” or “Government leader”. Therefore, in the present embodiment, a pattern clustering method similar to the method disclosed in Non-Patent Document 3 is used to solve the problem of pattern matching low frequency. By clustering vocabulary patterns, low frequency problems can be solved, and even if the vocabulary patterns do not completely match, the search recall can be increased.

Vocabulary pattern clustering will be described. First, a vocabulary pattern is represented by a frequency vector of entity pairs that appear together with the vocabulary pattern. For example, assume that the lexical pattern X, the president of Y appears 10 times in (Barack Obama, USA) and 20 times in (Vladmir Putin, Russia). In this case, a vector such as [10, 20] is created. Similarly, all vocabulary patterns are represented by frequency vectors of entity pairs that appear together.

ここで、ｈはエンティティペア（S,T）と語彙パターンpの出現頻度を表す関数である。 In the vector p representing the vocabulary pattern p, the i-th element is given by the following equation.

Here, h is a function representing the appearance frequency of the entity pair (S, T) and the vocabulary pattern p.

Next, the similarity between vocabulary patterns is calculated by the cosine similarity between the corresponding vectors, and the operation of “clustering similar vocabulary patterns into clusters” is performed. As a clustering algorithm, Algorithm 1 of page 655 of Non-Patent Document 3 is used. Here, the clustering method disclosed in Non-Patent Document 3 is incorporated herein by reference.

Briefly, first, a set of vocabulary patterns is sorted in descending order of the total appearance frequency (the sum of the appearance frequencies in all entity pairs). Thereby, the ranking of the vocabulary pattern in order of frequency is obtained. Next, vocabulary patterns are selected one by one from the beginning, the similarity with existing vocabulary pattern clusters is calculated, and the nearest vocabulary pattern cluster is searched. If the similarity with the nearest vocabulary pattern cluster is larger than a predetermined threshold, the vocabulary pattern of interest is added to the vocabulary pattern cluster. Otherwise, a new vocabulary pattern cluster consisting of one vocabulary pattern is created. This process is performed for all vocabulary patterns.

で定義する。一つの典型的な態様では、関係類似度として、２つの特徴ベクトルのcosine類似度を用いることができる。また、非特許文献3で提案されている手法を使うことも出来る。また、特徴ベクトル間の距離尺度については、ユーグリット距離、マハラノビス距離等の公知の距離尺度を用いることもできる。 [B-3] When a feature vector between entity pairs is obtained, in order to search for an answer to a query, the similarity between two entity pairs (this similarity is referred to as “relationship similarity”) is used. , It must be defined by a feature vector expressing the relationship between entities. The relationship similarity can be calculated by the distance between feature vectors. Here, a function expressing the relationship similarity between two entity pairs is

Define in. In one exemplary aspect, the cosine similarity of two feature vectors can be used as the relationship similarity. The method proposed in Non-Patent Document 3 can also be used. As a distance scale between feature vectors, a known distance scale such as Eugrid distance or Mahalanobis distance can be used.

For example, as shown in FIG. 4, since the feature vector V1 of the entity (word) pair of (USA, Obama) and the feature vector V2 of the entity (word) pair of (France, Sarkozy) are high, RelSim The value of ((USA, Obama), (France, Sarkozy)) will also be high.

If the similarity calculation method is used, when the lexical pattern of the entity pair (A, B) and the vocabulary pattern of the entity pair (C, D) are accurately matched, the relationship similarity between these entity pairs increases. However, there are many vocabulary patterns that express relationships between entities, and there may be few vocabulary patterns that perfectly match. Therefore, it is desirable to perform similarity calculation considering the above vocabulary pattern clustering.

The calculation method of the relationship similarity will be described more specifically. The two entity pairs for calculating the relationship similarity are (A, B) and (C, D), respectively. First, a vocabulary pattern that associates entity A and entity B is extracted. The number of times a certain vocabulary pattern p is extracted for the entity pair (A, B) is represented by h (A, B, p).

Next, the feature vector F (A, B) of the entity pair (A, B) is constructed as follows. The i-th element of F (A, B) is the sum of the appearance frequencies of all patterns p and entity pairs (A, B) that appear in the i-th vocabulary pattern cluster C _i . If you write the i-th element of F (A, B) as an expression,

The meaning of the above formula is as follows. First of all, it should be noted that the vocabulary pattern cluster is intended to organize “different expressions used to express the same relationship”. For example, the relationship of being president of a country can be expressed in various vocabulary patterns such as “X the president of Y”, “Y president X”, “Y's head of state X” (where X is Person name, Y is country name). By treating these “different surface layer expressions representing the same meaning” as if they were the same, even if seemingly different vocabulary patterns are used, the similarity calculation between entity pairs is not affected. By using the above formula, all different vocabulary patterns representing the same meaning appear in one dimension of the vector. Therefore, there is an advantage that the zero frequency problem can be avoided. If the number of clusters in the vocabulary pattern is n, the feature vector has n dimensions. That is, each vocabulary pattern cluster contributes to one feature.

ここでVは全エンティティペアを含む集合である。 Furthermore, another device needs to be added to the above formula. Some vocabulary pattern clusters contain a huge number of patterns, and some lexical pattern clusters do not contain many patterns. This is because the types of vocabulary patterns used to express a certain relationship vary depending on the relationship. For example, “X is a large Y”, “X the largest Y”, “X is a big Y”, “enormous YX”, etc. can be written to express the relationship of “largest” (eg ostrich and bird) There are so many expressions. On the other hand, there is not so much to express a fairly clear relationship such as “a president of a country”. The above formula does not consider how many vocabulary patterns exist in each vocabulary pattern cluster, that is, the size of the vocabulary pattern cluster. Therefore, normalization is performed by dividing the right side of the above expression by “the sum of all patterns in the cluster”.

Here, V is a set including all entity pairs.

[B-4] Entity pair indexing
In order to execute a search, it is not sufficient to measure the similarity between pairs. Because the fourth entity is not known. Since one entity of the second pair is unknown, it needs to be searched. Therefore, in this embodiment, as many entity pairs as possible are stored in the database, and their feature vectors are calculated in advance and stored in the database. This is called entity pair indexing. The result of Indexing is “mapping of feature vectors expressing the relationship between entities of the entity pair from the entity pair” and “mapping of the entity pair list including the vocabulary pattern from the vocabulary pattern”. More specifically, they are “entity pair-feature vector table” and “vocabulary pattern-entity pair table including vocabulary pattern”. A search using a search database obtained by indexing entity pairs will be described later with reference to FIG.

[B-5] When a query {(A, B), (C,?)} Of a query processing related search engine is inputted, a feature vector of (A, B) is searched and included in the feature vector. Find all existing vocabulary patterns (greater than frequency 1).

Next, a vocabulary pattern cluster (the above clustering result) to which those vocabulary patterns belong is found. Then, other word pairs (C, D) including vocabulary patterns belonging to those vocabulary pattern clusters are found, and D candidates are ranked using RelSim ((A, B), (C, D)) as an index. The list of ranked D is the search result.

In addition, for entity pairs that do not appear in the search database obtained by indexing, search the entire Web text using a keyword-based search engine as described below. Since this operation takes some time, it is desirable to ask the user's permission before execution.

[B-6] Dynamic processing at the time of query using a remote corpus Even if the local corpus is enormous, not all queries can be covered. Therefore, use the remote corpus so that you can answer for entities that are not in the local corpus.

The remote corpus is a set of snippets that are transmitted as a result of sending a query “A ***** B” to a keyword-based search engine (Google, Yahoo, etc.) when the query is known. Here, the “*” operator is a wildcard that matches one or zero words and is supported by many existing search engines (eg, Google, YahooBOSS API). Since the text of these snippets is not pre-existing on the local hard disk, it is called a remote corpus. In the above query, it is possible to download a snippet when A and B co-occur within a maximum of 5 words. In a normal existing keyword search engine, when A and B are input as a query, a snippet including A and B can be obtained. In this wildcard search, a simple query of A and B is performed. It can be said that the input between is limited to the distance between A and B. Therefore, when searching with a query (“A ***** B”), A and B co-occur within a maximum of 5 words and a snippet including the context before and after A and B is obtained. A vocabulary pattern including contexts before and after A and B can be generated. By using the snippet, the time taken to actually download the search result from the Web can be reduced.

Dynamic processing at query time using a remote corpus, in one aspect, is performed when at least one of the three entities in the input query is not in the local corpus. If the three entities in the query exist in the local corpus, the remote corpus is not used.

Next, replace all occurrences of A and B with variables X and Y, respectively, in the downloaded context. Convert the context to lowercase and perform lemmatization (base shape extraction). Next, all n-grams that are less than 5-gram including both X and Y are selected as vocabulary patterns. If the extracted vocabulary pattern is already included in the search database, the high-frequency vocabulary pattern may be preferentially selected.

In the vocabulary pattern extracted by the above vocabulary pattern extraction, C is substituted for the variable X and “*” is substituted for Y. In this way, the Web search is performed again using the obtained vocabulary pattern. A portion of the snippet that matches “*” is extracted as a candidate for D. In addition, not only one word but also bi-gram, tri-gram, etc. can be extracted as candidates from the portion corresponding to “*”. By doing so, it is possible to deal with an entity D consisting of one word or more. As already mentioned, the query “C pattern *” does not return only “C pattern W1”, but “… C pattern W1 W2 W3…” is returned as the context, so W1, W2, W3, etc. Can be used as candidates.

At this stage, various entities can appear as candidates for the entity D. Therefore, it is necessary to rank a large number of entity D candidates. The following methods can be exemplified as a ranking method for a huge number of candidates.
a. Rank in descending order of appearance frequency.
b. Rank by number of different vocabulary patterns selected as candidates.
c. Using the candidate D, ranking is performed based on whether or not C is obtained when searching with (A, B), (?, D) queries (inverse search method).
d. Learn any combination of the above techniques using machine learning.

[C] An embodiment of a search system using a relational search database using a search database will be described in detail with reference to FIGS. For simplicity, the entity pair and the entity ID are described as a word pair and a word pair ID.

FIG. 3 is a diagram showing the configuration of the search system. Data source is Local Corpus
corpus) 1 and / or remote corpus 2. The local corpus is a corpus stored in the storage unit of the search system, and includes, for example, a Wikipedia data dump, a crawled WWW page, and the like. On the other hand, a remote corpus is a corpus that does not exist in the storage unit of the present search engine, and includes, for example, a query result of an ordinary keyword-based search engine. If you use a remote corpus, you can handle a huge amount of data sources such as WWW without crawling with this system. The remote corpus may be temporarily stored in the storage unit. In this specification, these local and remote corpora are generally referred to as text corpora.

Before processing a search query, a database for searching is created using a text corpus. The indexing engine 3 and the pattern clustering engine 4 of the search system shown in FIG. 3 are modules for creating a search database. The indexing engine 3 includes a tokenizer 30, an entity pair selector (Word-Pair Selector) 31, and a lexical pattern generator 32. By the indexing engine 3, entity pair index (Word-Pair Index) 50, vocabulary pattern index (Pattern Index) 51, feature vector index (Feature Vector Index) 52, inverted entity-pair index (Inverted Word-Pair Index) 53 One database is created.

The indexing engine 3 receives text from the text corpora 1 and 2 and divides the text into words by the tokenizer 30. For example, as shown in FIG. 5, the tokenizer 30 receives an input sentence (Tokyo is the capital of Japan) and a part-of-speech sequence of words included in the sentence (Tokyo / noun, ha / particle, Japan / noun). , No / particle, capital / noun, de / auxiliary verb, certain / auxiliary verb).

Next, an entity pair selector (Word-pair Selector) 31 receives the word string generated by the tokenizer 30 and creates an entity pair. Entity pair candidates are all nouns, verbs, adjectives, and entity pairs that have adjective verbs as part of speech in a sentence, but ultimately only frequent entity pairs are selected by the entity pair selector, Proceed to the next step. A table for storing the frequency of entity pairs is a feature vector index 52 described later.

Associate a unique integer for each entity. This integer is called an entity ID. In order to find the entity ID from the entity at high speed, a hash table of the entity and the entity ID is created as shown in FIG. 6A and stored in one table of the database. For example, entity IDs 1, 2, 3, 4, and 5 are assigned to Tokyo, France, Japan, Paris, and the capital, respectively. Subsequent processing does not use the entity itself, and only processing for the entity ID is required, so that there is an advantage that the search can be performed at a high speed and the amount of data to be stored can be compressed. The search database also stores the frequency and part of speech information of each entity in association with the entity ID.

Then, for each entity pair candidate, the hash table of FIG. 6B is created using the two entity IDs of the pair as keys. FIG. 6B is a table that maps two entity IDs of an entity pair to an entity pair ID. FIG. 6B shows the contents of the entity pair index 50. Specifically, an entity pair is formed from entity ID1 and entity ID3, and entity pair ID1 is assigned to the entity pair. Similarly, an entity pair is formed from entity ID2 and entity ID4, and entity pair ID2 is assigned to the entity pair.

In the final step of the indexing engine 3, the lexical pattern generator 32 generates lexical patterns for the selected entity pair. As the lexical pattern generation algorithm, the above-described method can be used. The vocabulary pattern generator creates a table (Pattern vs. Pattern ID) 51 as shown in FIG. 7 that associates lexical patterns with lexical pattern IDs. As shown in FIG. 7, the vocabulary pattern “X is the capital of Y” is assigned a vocabulary pattern ID1, and the vocabulary pattern “X is the largest city of Y” is assigned a vocabulary pattern ID2. Vocabulary pattern 3 is assigned to “the capital of Y is X”.

Furthermore, not only the vocabulary pattern ID information but also the total number of vocabulary pattern frequencies is stored from the vocabulary pattern (see FIG. 8). The vocabulary pattern generator 32 also counts the frequency of the vocabulary pattern of the entity pair (the frequency of the vocabulary pattern for the entity pair) and stores it in the feature vector index database 52. The feature vector index 52 is a mapping from “entity pair” to “feature vector of an entity pair”. Conceptually, the feature vector index is a table shown in FIG. In other words, the table stores information on lexical patterns including the entity pairs and their frequencies from the entity pairs. As shown in FIG. 8, the entity pair (Tokyo, Japan) has the vocabulary pattern “X is the capital of Y” / frequency “10”, and the vocabulary pattern “X is the largest city of Y” / frequency “4”. , The vocabulary pattern “the capital of Y is X”, the frequency “6”, and the total frequency “20”. The entity pair (Paris, France) corresponds to the vocabulary pattern “X is the capital of Y” / frequency “12”, the vocabulary pattern “the capital of Y is X” / frequency “3”, and the total frequency “15”. It is attached. Here, for simplicity, FIG. 8 does not show all n-grams of vocabulary patterns.

In addition to the vocabulary pattern, a part of speech pattern including an entity pair may be extracted, and the entity pair and the part of speech pattern may be associated with each other and stored in the database. Extraction of part-of-speech patterns is known to those skilled in the art. In one aspect, information on the part of speech of an entity (ie, whether this entity is a common noun, person name, organization name, place name) can be used for entity clustering. This is because if two entities have different parts of speech (for example, person names and place names), they are less likely to enter the same cluster than the same part of speech, so when measuring similarity, Make the similarity half the normal similarity (penalize different part-of-speech pairs). As an example, if the similarity between Barack Obama (person name) and Obama (person name) is 0.9, the similarity is counted as 0.9 and put in the same cluster. However, even if the similarity between Barack Obama (person name) and Whitehouse (place name) is 0.9, the part of speech is different, so the similarity is halved and reduced to 0.45. So Barack Obama and Whitehouse can't be in the same cluster.

In order to easily search the feature vector index or reduce the size, the entity pair ID and vocabulary pattern ID are used instead of storing the text of the entity pair directly as a key. Therefore, what is actually saved is the table of FIG. In FIG. 9, FIG. 8 is expressed using the entity pair ID and the vocabulary pattern ID. By using the tables shown in FIGS. 6, 7, and 9, the table shown in FIG. 8 can be reproduced. Further, in order to limit candidate entity pairs from stem word pairs, candidate entity pairs are limited to entity pairs that share one or more vocabulary patterns (or vocabulary pattern clusters) with stem word pairs.

In order to search the set of candidate entity pairs at high speed, an inverted entity pair index (Inverted Word-Pair Index) 53 is prepared. As shown in FIG. 10, the inverted entity pair index is a mapping from a vocabulary pattern ID to a list of entity pair IDs and frequencies including the vocabulary pattern as one component of a feature vector. As shown in FIG. 10, the vocabulary pattern ID1 corresponds to the entity pair ID1 · frequency “10”, the entity pair ID2 · frequency “12”, and the total number of frequencies “22”. The vocabulary pattern ID2 corresponds to the entity pair ID1, the frequency “4”, and the total frequency “4”. The vocabulary pattern ID3 corresponds to the entity pair ID1 / frequency “6”, the entity pair ID3 / frequency “3”, and the total frequency “9”. Since this table is an index in the opposite direction to the feature vector index table (FIG. 9), it is called an inverted index.

As described above, lexical pattern clustering and entity clustering (one method of named entity disambiguation) are used to solve the problem of low frequency and perfect matching. The method used in this system is the sequential clustering algorithm proposed in Non-Patent Document 3. A module for realizing clustering by this algorithm is the pattern clustering engine 4. The pattern clustering engine 4 uses the feature vector index 52 and the inverted entity pair index 53 to combine similar vocabulary patterns into one vocabulary pattern cluster. A similar vocabulary pattern is one in which the distribution of word pairs in which the vocabulary pattern appears is similar (based on Distributional hypothesis).

Using feature vector index 52 and inverted entity pair index 53, assuming that there are only three word pair frequency vectors (for example, in FIG. 10, there are only three word pairs (word pair IDs 1, 2, 3) in FIG. Cosine similarity of (10, 12, 0) for word pair frequency vector, (4, 0, 0) for pattern ID 2 and (6, 0, 3) for pattern ID 3) And the sequential clustering algorithm described in Non-Patent Document 3 is applied to cluster the vocabulary patterns. In the search database, a table of vocabulary pattern clusters in which similar vocabulary patterns are collected is prepared, and each vocabulary pattern is distributed to any vocabulary pattern cluster. A vocabulary pattern cluster is associated with a vocabulary pattern ID, which is assigned a vocabulary pattern cluster ID.

Proper nouns (named entities, etc.) are clustered using the co-occurrence frequency and common pattern with the other entity as described above. Similar to the vocabulary pattern clustering, the similarity between entities is calculated by the cosine similarity between the corresponding vectors, and the work of “clustering similar entities together” is performed.
For example, United States, US, and USA are included in one cluster. Which cluster each entity enters is recorded in an entity cluster table in which similar entities are grouped. An entity cluster ID is assigned to the entity cluster, and the entity ID and the entity cluster ID are associated with each other. For entities other than proper nouns, an entity cluster in which synonyms are collected may be prepared. This completes the preparation for searching.

The last module is a query processing engine 6. When {(A, B), (C,?)} Is input as the query 7, the query processing engine 6 first searches the entity pair index 50 for the ID of the input stem word pair (A, B) (stem_id). And).

Using the stem_id, the feature vector f (a, b) of the stem word pair (A, B) is found based on the result of the feature vector index 52 (see FIGS. 8 and 9) and the pattern clustering engine 4. Then, a vocabulary pattern ID greater than frequency 1 in f (a, b) is searched, and a set of candidate entity pair IDs is created with reference to the inverted entity pair index 53 (FIG. 10). To be a candidate entity pair, one entity in the pair needs to be the entity C of the query.

Finally, the similarity between the feature vector of the stem word pair (A, B) and the feature vector of the candidate entity pair (C, D) is calculated using the RelSim ((A, B), (C, D)) function. Calculate and sort the set of candidate entity pairs in descending order of similarity.

In one aspect, when calculating a score (similarity) that serves as an index for ranking, scores of a plurality of entities that are included in one entity cluster (that is, one entity is pointed out but the expression format is different) A cluster average score obtained by dividing the sum of (similarity) by the size of the entity cluster is taken, and ranking is performed using the score (similarity) of the entity (entity). In other words, instead of ranking individual entities by individual entity scores (similarity), the score as a cluster of entities (using the cluster average score obtained by dividing the sum of the scores of the members of the entity cluster by the size of the cluster) Make a ranking.

The cluster score (similarity) will be specifically described below.
Input query: (A, B), (C,?)
A vocabulary pattern containing (A, B) is S ₁ = {p ₁ , p ₂ , .., p _n }
A vocabulary pattern containing (C, X) is S ₂ = {q ₁ , q ₂ , .., q _m }
If q ₁ ∈ S _1, then the score of X is h (q ₁ ,
Add (C, X)) * h (q ₁ , (A, B)).
Here, h (p, wp) is “appearance frequency of word pair wp of vocabulary pattern p”.
Even for q1 not∈S _1, q ₁ and p _i is given the same pattern cluster, also calculates a score as follows.

h (q ₁ , (C, X)) * h (p _i , (A, B))
The cluster score is the sum of all X scores.
For example,
(Obama, US) :( is president of [4], is leader of [3], is a senate of [2],)
(Koizumi, Japan) :( is prime minister of [8], is leader of [5])
Query: (Obama, US), (Koizumi, X)
If “is president of” and “is prime minister of” are in the same cluster, the score of X is: (3 * 5 + 4 * 8) = 15 + 32 = 47.
Next, the cluster average score will be described.
Cluster total score = Σ (entity score),
Cluster average score = Cluster total score / size.
When an entity cluster (United States, US, US) is formed,
For example, if United States score: 15, US score 10, US score 8,
Cluster total score: 15 + 10 + 8 = 33,
Cluster average score: 33/3 = 11.

Using the cluster score of an entity has an important impact on the ranking of search results. For example, if the query is (Apple, Steve Jobs), (Microsoft,?), The result score of D = Steve Ballmer alone is D = Windows (both X introduces Y, X announces Y, ... Can often be lower). D = Steve
By combining Ballmer, D = Ballmer, D = Steve Anthony Ballmer into one cluster, the score of the entity cluster (the total result of the scores of each entity) will be higher than the score of the entity cluster of D = Windows only. It is done. This score reversal has the advantage that the desired result is ranked higher.

A ranked list 8 of search results D (or entity clusters belonging to D) is output from the sorted set of candidate entity pairs. In addition, a sentence or a document including a vocabulary pattern that the entity pair (A, B) and the entity pair (C, D) have in common may be output as a reason for explaining the result. Furthermore, a representative relationship of sharing (A, B) and (C, D) can also be output.

When an entity is obtained as a search result, clustering the entities can output an “entity cluster” as the search result. In other words, by outputting search results using entity clusters, similar items (eg, Obama and Barak Obama) can be displayed together when X candidates are presented to the user.

[D] Search Using Dynamic Processing Above, search using a search database has been described. When one or more of the entities A, B, and C do not exist in the search database, a dynamic processing mode using a remote corpus is executed by an existing search engine. This will be specifically described below.

When the entities A and B of the entity pair (A, B) are not in the search database, the vocabulary pattern representing the relationship between the entities A and B is extracted using an existing search engine. First, an existing search engine is searched for the query “A **** B”. This receives a snippet in which entity A and entity B appear within a maximum of 4 words (* is a wild card matched with 1 to 0 words) as a search result. In order to generate a vocabulary pattern later, a word sequence (n-gram) is extracted. For this purpose, entity A and entity B need to appear nearby. In other words, this query attempts to approximate the “context” of entity A and entity B. The number of wildcards * can be appropriately set by those skilled in the art. For example, a plurality of queries from 1 to a maximum of 7 can be used. For example, “A * B”, “A ** B”, “A *** B”,... “A ******* B” are used. In some cases, the results received with a query with many * and the results with a query with few * are similar, so in one aspect, snippets from the same URL are collected only once. In many existing search engines (for example, Google, YahooBOSS API), the query “A **** B” provides a snippet that includes the context before and after A and B.

From the snippet acquired using the wild card search, what is the context of the entity pair (A, B) is obtained. Next, it is necessary to extract a vocabulary pattern that expresses the relationship between entity A and entity B from this context. The algorithm for extracting the vocabulary pattern is the same as the lexical pattern acquisition method described above. That is, n-grams including entity A and entity B are extracted, and A in the extracted n-grams is replaced with X and B is replaced with Y. An example will be described below.

When entity A = Barack Obama and entity B = USA, a Google search with “A **** B” will give the following results.

Make the above snippet into a lowercase and convert all the used forms to base forms (lemmatization). Lemmatization converts the plural form of the noun into its singular form, and the usage forms of verbs (past and progressive forms) become base forms, and variations due to use are absorbed. By using lowercase, variations due to uppercase and lowercase letters are absorbed.

次のこの中でXもYも両方を含むn-gramを生成する。例えば、n=6で

が生成される。 Next, if X = Barack Obama, Y = USA, the result is as follows.

In the following, generate an n-gram that includes both X and Y. For example, n = 6

Is generated.

The above vocabulary pattern starts with X and ends with Y, but it does not necessarily extract only the vocabulary pattern that connects between X and Y. The condition for extracting the vocabulary pattern is that the length is n and includes both X and Y, and all those satisfying the two conditions are extracted as vocabulary patterns. For example, when n = 7, the following two patterns are extracted from the above snippet.

If the obtained vocabulary pattern also exists in the search database, the one with higher frequency can be prioritized, but if it does not exist in the search database, it is extracted, for example, twice or more for the query (A, B). Search for all patterns by substituting C for X.

For example, in the case of the above vocabulary pattern of X, president of the Y and C = Vladimir Putin, the following query is obtained.
Vladimir Putin, president of the *
You can now search Google for the following results:

In this snippet, the FEDERATION OF RUSSIA part is correct. In order to extract the correct part, for example, 3-gram from the part appearing after the vocabulary pattern is taken, and the answer candidates are ranked by appearance frequency.

Thus, when the entity pair (A, B) is not in the database, it is necessary to extract the vocabulary pattern representing the relationship of the entity pair (A, B) from the web search result snippet. If entity C is in the search database, the candidate (C, X) can be selected by searching the database with the extracted vocabulary pattern. Of course, the dynamic mode can be executed by ignoring the database even when the entities (A, B) and the entity C are in the search database. Even in this case, the entity C is not in the database. All can be processed in the dynamic mode. However, since the dynamic mode actually needs to access the Web search engine dynamically on the spot, compared with the search using the search database (entities A, B, and C are all in the search database) Note that it takes minutes. In one embodiment, a search using the search database outputs a search result within 1 second, whereas in the dynamic mode, a search result is output within 30 seconds.

2A to 2C show examples of search result display screens. FIG. 2A shows a query (steve
It is a display screen (displayed on the display of the client terminal) of simple results (results not including evidence) of jobs, apple), (?, microsoft). In FIG. 2A, the ranking of search results is displayed, specifically, in the order of ballmer, steve ballmer (score 295), bill gates (score 52), and danny thrope (score 27). From the search result displayed in FIG. 2A, it can be seen that the ballmer and steve ballmer are clustered. In the display screen of FIG. 2A, when Show evidence is selected from the client terminal by the input means, the evidence text is displayed on the display screen. FIG. 2B shows only a part of the evidence (bill gates) on the display screen of FIG. A part of the vocabulary pattern is displayed in Debug Info. FIG. 2C is a display screen of a query (steve jobs, apple), (steve ballmer,?) Simple result (result not including evidence).

Claims (34)

When the entity pair (A, B) and the entity C are input, the relationship between the entities A and B between the entities C and X of the entity pair (C, X) is the same or similar. A method of searching for an entity X using a search database or / and an existing search engine,
Entity pair (A, B) and entity C are received as a query,
Defining the relationship between entity A and entity B depending on the context around entities A and B in the text containing entity A and entity B;
By searching for the entities C and X having the same or similar context as the surroundings of the entities A and B, the relationship between the entities A and B is the same as or similar to the relationship between the entities A and B. A search method for acquiring an entity X provided.   The search method according to claim 1, wherein the context is one or any combination of a vocabulary pattern, a bag-of-words, a part-of-speech pattern, and a dependency pattern.   The search method according to claim 1, wherein the contexts are clustered, and each context cluster includes substantially the same or similar context.   The search method according to claim 3, wherein a relationship between two entities of a pair of entities is represented by a feature vector, and the feature vector is obtained from a context or a context cluster around the two entities of the entity pair. . The relationship between entity A and entity B is represented by a first feature vector, which is obtained from the context or context cluster around entities A and B;
The step of acquiring the entity X is to acquire the entity X having a second feature vector that is the same as or similar to the first feature vector with the entity C, and the second feature vector is the entity C 5. The search method according to claim 4, wherein the search is acquired from a context or a context cluster around the entities C and X in the text including the entity X. 5. The step of obtaining the entity X includes:
The similarity between the pair of the entity pair (A, B) and each entity pair (C, X) is calculated based on the distance between the feature vectors of the two, and the calculated pair similarity is used as an index to determine the entity pair. Sorting a plurality of entity pairs (C, X) in descending order of similarity, and ranking X candidates in order of sorting;
Outputting a part or all of the ranked X as a search result;
The search method according to claim 5, further comprising:   The search method according to claim 4, wherein the context is a vocabulary pattern, and the feature vector has a frequency of the vocabulary pattern as an element. The step of obtaining the entity X includes:
A plurality of entity pairs (C, X) having a second feature vector whose elements include the vocabulary pattern of the first feature vector of the acquired entity pair (A, B) (frequency is greater than a preset threshold) A step to obtain,
The similarity between the pair of the entity pair (A, B) and each entity pair (C, X) is calculated using both feature vectors, and the similarity between the entity pairs is calculated using the calculated similarity between the pairs as an index. Sorting a plurality of entity pairs (C, X) in descending order and ranking candidates for X in sort order;
Outputting a part or all of the ranked X as a search result;
The search method according to claim 7, comprising:   The search method according to claim 1, wherein the search database stores correspondences between a large number of entity pairs and surrounding contexts expressing relationships between entities of each entity pair.   The search according to claim 9, wherein the context is a vocabulary pattern, and the search database stores correspondences between a large number of entity pairs, each entity pair, a vocabulary pattern and a frequency of the vocabulary pattern. Method. The search database is
A first index defining the ID of each entity in the entity pair;
A second index defining entity pair IDs corresponding to two entity ID pairs;
A third index defining a vocabulary pattern ID corresponding to each vocabulary pattern;
A fourth index that defines the correspondence between the entity pair ID and the vocabulary pattern ID / frequency of the vocabulary pattern;
With
A pair ID corresponding to the input entity pair (A, B) is acquired, and using the acquired entity pair ID, a vocabulary pattern ID corresponding to the entity pair (A, B) is set in advance. Obtaining a frequency of vocabulary patterns to form a first feature vector; and
Obtaining a candidate entity pair ID including the vocabulary pattern ID of the first feature vector (the frequency of the vocabulary pattern is greater than a preset threshold) and the ID of entity C;
Obtaining a second feature vector corresponding to each candidate entity pair ID and obtaining an entity pair (C, X) such that the second feature vector is similar to the first feature vector;
The search method according to claim 10, comprising: The fourth index is
An index for searching the frequency of the vocabulary pattern ID / vocabulary pattern from the entity pair ID;
An index for searching the frequency of entity pair ID / vocabulary pattern from the vocabulary pattern ID;
The search method according to claim 11, comprising: In the search database, vocabulary patterns form vocabulary pattern clusters, and each vocabulary pattern cluster includes similar or substantially identical vocabulary patterns.
The search method according to claim 10.   The search method according to claim 13, wherein the dimension of the feature vector when calculating the similarity between pairs is a vocabulary pattern cluster, and the value of each dimension is the sum of the frequencies of vocabulary patterns included in the vocabulary pattern cluster.   The search method according to claim 9, wherein an entity cluster composed of similar entities is formed in the search database.   The search method according to claim 15, wherein a score of an entity cluster to which the entity X belongs is used as a score serving as an index for ranking the candidate of the entity X.   The search method according to claim 15, wherein in the output step, a plurality of entities included in the entity cluster are output. A feature vector that represents the relationship between entities in an entity pair is
The sentence acquired from the text corpus is divided into words,
Create candidate entity pairs from the segmented words,
The search method according to claim 4, wherein the search method is acquired from a context around a candidate entity pair. The element of the feature vector is the frequency of the vocabulary pattern,
Extracting the frequency of vocabulary patterns
In the original sentence including the candidate entity pair, the first entity of the entity pair is replaced with X, the second entity is replaced with Y, and a partial sequence S including X and Y is generated, and an n-gram ( n is from 1 to K),
Using the obtained n-gram as one dimension of a feature vector, counting the frequency in all sentences including entity pairs;
With
The search method according to claim 18, wherein the n-gram frequency is stored in association with an entity pair as a vocabulary pattern frequency.
Here, the partial sequence S is m ₁ words immediately before X, a word sequence between X and XY, and m ₂ words immediately after Y and Y. m ₁ , m ₂ , and K are integer parameters. The search method according to any one of claims 18 and 19, wherein a context around a candidate entity pair is acquired only when a distance D between two entities of the candidate entity pair is equal to or less than a threshold value _Dth .   21. The search method according to claim 18, wherein in the step of creating the candidate entity pair, an entity pair including a proper noun is preferentially extracted as a candidate entity pair.   The search method according to any one of claims 18 to 21, wherein, in the step of creating the candidate entity pair, a high-frequency entity pair is preferentially extracted as a candidate entity pair.   The search method according to any one of claims 18 to 22, further comprising the step of forming a cluster of vocabulary patterns from similar or substantially identical vocabulary patterns. Using an existing search engine to obtain the context around entities A and B in the text containing entities A and B;
Using existing search engines to obtain entities C, X having the same or similar context as the surrounding context of entities A, B;
The search method according to claim 1, further comprising:   25. The context around the entities A and B is obtained from a snippet as a search result of the search engine with the entity A and the entity B as a query, the distance between the entity A and the entity B being limited. Search method described in. The context around the entities A and B is acquired as a vocabulary pattern,
Entity X candidates are obtained by substituting C for the position of A in the acquired vocabulary pattern “C
26. The search method according to claim 25, wherein the search method is acquired from a snippet as a search result of the search engine using “pattern *” as a query. When the entity pair (A, B) and the entity C are input, the relationship between the entities A and B between the entities C and X of the entity pair (C, X) is the same or similar. A system for searching for an entity X using a search database,
The search database stores correspondences between a large number of entity pairs and surrounding contexts that express relationships between entities of each entity pair,
Means for receiving entity pair (A, B) and entity C as a query;
Means for obtaining, using the search database, a first peripheral context expressing a relationship between entity A and entity B;
Means for obtaining entity C by searching for an entity C, X having a second peripheral context identical or similar to the first peripheral context using the search database;
Search system equipped with.   The search according to claim 27, wherein the context is a vocabulary pattern, and the search database stores correspondences between a large number of entity pairs, each entity pair, a vocabulary pattern and a frequency of the vocabulary pattern. system. A plurality of entity pairs (C, X) having a second feature vector whose elements include the vocabulary pattern of the first feature vector of the acquired entity pair (A, B) (frequency is greater than a preset threshold) Means to obtain,
Means for calculating the similarity between the pair of the entity pair (A, B) and each entity pair (C, X) using both feature vectors;
Means for sorting a plurality of entity pairs (C, X) in descending order of the similarity between entity pairs using the calculated similarity between pairs as an index, and ranking X candidates in the order of sorting;
Means for outputting a part or all of the ranked X as a search result;
The search system according to claim 28, comprising: The search database is
A first index defining the ID of each entity in the entity pair;
A second index defining entity pair IDs corresponding to two entity ID pairs;
A third index defining a vocabulary pattern ID corresponding to each vocabulary pattern;
A fourth index that defines the correspondence between the entity pair ID and the vocabulary pattern ID / frequency of the vocabulary pattern;
With
A pair ID corresponding to the input entity pair (A, B) is acquired, and using the acquired entity pair ID, a vocabulary pattern ID corresponding to the entity pair (A, B) is set in advance. The frequency of the vocabulary pattern is acquired to form the first feature vector,
Obtaining a candidate pair ID including the vocabulary pattern ID of the first feature vector (the frequency of the vocabulary pattern is greater than a preset threshold) and the ID of the entity C;
Obtaining a second feature vector corresponding to each candidate entity pair ID and obtaining an entity pair (C, X) such that the second feature vector is similar to the first feature vector;
The search system according to claim 29. The fourth index is
An index for searching the frequency of the vocabulary pattern ID / vocabulary pattern from the entity pair ID;
An index for searching the frequency of entity pair ID / vocabulary pattern from the vocabulary pattern ID;
The search system according to claim 30, comprising: The search database stores vocabulary pattern clusters, and each vocabulary pattern cluster includes similar or substantially identical vocabulary patterns.
The search system according to any one of claims 27 to 31.   33. The search system according to claim 27, wherein entity clusters are stored in the search database, and each entity cluster includes similar or substantially identical entities.   When the entity pair (A, B) and the entity D are input by replacing the entity pair (C, X) with the entity pair (X, D) and the entity C with the entity D, the entity pair (X 34) The search system according to any one of claims 27 to 33, which functions as a system for searching for an entity X having the same or similar relationship between the entities A and B between the entities X and D of D).

Images