JP2012043294A - Binomial relationship categorization program, method, and device for categorizing semantically similar word pair by binomial relationship - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a binomial relationship categorization program and the like that can categorize a semantically similar word pair according to a binomial relationship by collectively handling an inter-noun relationship and an inter-verb/adjective relationship as an inter-word relationship, without predefining an inter-word relationship to be acquired.SOLUTION: A plurality of word pairs of a prescribed threshold amount or more that tend to co-occur are extracted from a sentence set storage section, and a set of dependency words that co-occur for each word in the word pairs is extracted from the sentence set storage section. Then, a first set of characteristic dependency words consisting of dependency words appearing in a first set of dependency words and not appearing in a second set of dependency words, and a second set of characteristic dependency words consisting of dependency words appearing in the second set of dependency words and not appearing in the first set of dependency words, are extracted. Further, an appearance frequency of co-occurrences with a word within a set of sentences is counted, for each dependency word, and a vector is derived in order to generate a word pair cluster by division optimization clustering based on the degree of similarity between the vectors.

Description

  The present invention relates to a technique for extracting an inter-word relationship from a sentence.

  2. Description of the Related Art Conventionally, there is a technique for extracting a word relationship such as a relationship between nouns or a verb / adjective relationship including a high-order relationship and a partial whole relationship, which is composed of two terms, from a sentence. There are a first conventional technique using a lexical syntactic pattern and a second conventional technique using co-occurrence information of nouns in a dependency relationship in order to automatically extract a word relationship from a sentence.

According to the first prior art, inter-word relations are extracted by applying a lexical syntactic pattern including a part-of-speech utilization form or a conjunction to a sentence (for example, non-patent documents 1, 2, 3, 8, 9 10). A lexical syntactic pattern is a pattern that uses a dependency relationship with a word, such as “* such as *” (see Non-Patent Document 8, for example). For example, as shown below, the upper and lower relations can be extracted by applying a lexical syntactic pattern to an example sentence.
Example sentence: "Cherry blossoms such as Yoshino cherry tree"
Vocabulary pattern: “*”
Higher-order subordinate relationship: (Sakura, Yoshino cherry tree)

According to the second prior art, for a given verb / adjective pair, the higher the similarity between sets of words related to the elements constituting each pair, the more presumed that there is a semantic relationship between the pairs (for example, (See Patent Documents 1 and 2, Non-Patent Documents 4, 5, and 11). As described below, for example, two predicates are presumed to have a semantic relationship because there is a common noun between “blur” and “go”.
Dependent nouns of the predicate "blur": "Kawara""town""park"
Dependent nouns for the predicate "go": "town""park""sandbeach"
Common nouns for two predicates: "town""park"

JP 2010-129025 A JP 2009-265889 A

T. Inui and M. Okumura, “Investigating the characteristics of causal relations in Japanese text,” in Proceedings of theWorkshop on Frontiers in Corpus Annotations II, 2005, 37-44. K. Torisawa, “Automaticacquisition of expressions representing preparation and utilization of anobject,” in Proceedings of the Recent Advances in Natural Language Processing, 2005, 556-560. S. Abe, K. Inui, and Y. Matsumoto, “Acquiring event relation knowledge by learning cooccurrencepatterns and fertilizing cooccurrence samples with verbal nouns,” inProceedings of the 3rd International Joint Conference on Natural LanguageProcessing, 2008, 497-504. D. Lin and P. Pantel, “DIRT-discovery of inference rules from text,” in Proceedings of ACMSIGKDD Conference on Knowledge Discovery and Data Mining, 2001, 323-328. C. Hashimoto etal., “Large-scale verb entailment acquisition from the web,” in Proceedings ofthe 2009 Conference on Empirical Methods in Natural Language Processing: Volume3, 2009, 1172-1181. Hideyuki Takahashi and Koichi Takeuchi, “Syntax Extraction of Verbs by Simultaneous Co-occurrence Clustering Considering Ambiguity”, IEICE Technical Report. NLC, Language Understanding and Communication, vol. 108, 2009, 37-42 C. Fellbaum, WordNet: An Electronic Lexical Database, The MIT Press, 1998. M. A. Hearst, “Automatic acquisition ofhyponyms from large text corpora,” in Proceedings of the 14th conference on Computational linguistics-Volume 2 (Association for Computational Linguistics Morristown, NJ, USA, 1992), 539-545. T. Chklovskiand P. Pantel, “Verbocean: Mining the web for fine-grained semantic verbrelations,” in Proceedings of EMNLP, vol. 4, 2004, 33-40. O. Etzioni et al., “Unsupervisednamed-entity extraction from the web: An experimental study,” ArtificialIntelligence 165, no. 1 (2005), 91-134. J. Kazama and K. Torisawa, “InducingGazetteers for Named Entity Recognition by Large-scale Clustering of DependencyRelations,” in Proceedings ACL-08: HLT (2008), 407-415.

  According to the first prior art described above, it is possible to acquire the inter-word relation with high accuracy if it matches the lexical syntactic pattern, but it is not possible to extract the inter-word relation that does not match the lexical syntactic pattern. Further, according to the second prior art, it is possible to extract even an inter-word relationship that does not match the lexical syntactic pattern, while an inter-noun relationship cannot be classified into a semantic relationship, and a verb / adjective relationship. For, only relationships that belong to implication relationships can be extracted.

  Moreover, according to the technique described in Patent Document 2, machine learning is performed so that the relationship between nouns can be classified into semantic relationships. However, when performing machine learning, learning data is required.

  Further, according to the first and second prior arts, the semantic relationship to be acquired is determined in advance, and the determined semantic relationship is acquired. However, because it is difficult to predetermine all the existing semantic relationships, it is possible to acquire specific semantic relationships, but acquire various semantic relationships and unexpected semantic relationships. I can't.

  Further, according to the first and second prior arts, either the noun relationship or the verb / adjective relationship is set as the classification target. However, it is not a general-purpose technique that can handle these relationships collectively as inter-word relationships.

  Therefore, the present invention treats the relationship between nouns and the relationship between verbs / adjectives collectively as the relationship between words and binomial relationship between semantically similar word pairs without predefining the relationship between words to be acquired. It is an object of the present invention to provide a binary relation classification program, method and apparatus that can be classified into two categories.

According to the present invention, in a binary relation classification program for executing a computer mounted on a device to classify word pairs into semantic binary relations,
It has a sentence set storage unit that stores a large number of document information,
A first step of extracting a plurality of word pairs consisting of a first word and a second word from the sentence set storage unit;
A second step of extracting a word pair whose similarity representing the ease of co-occurrence among word pairs is equal to or greater than a predetermined threshold;
For the word pairs extracted in the second step, the first dependency word set co-occurs in the first word and the second dependency word set co-occurs in the second word from the sentence set accumulation unit. A third step of extracting
A first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set; and a first feature dependency word set that appears in the second dependency word set and the first A fourth step of extracting a second characteristic dependency word set consisting of dependency words that do not appear in the dependency word set;
For each dependency word that belongs to the first feature dependency word set, the frequency of occurrence in the document set that co-occurs with the first word, and for each dependency word that belongs to the second feature dependency word set, the second A fifth step of counting the frequency of occurrence in the document set co-occurring with
A first vector representing the appearance frequency of each dependency word belonging to the first feature dependency word set based on the first word, and a dependency word belonging to the second feature dependency word set based on the second word A sixth step of deriving a vector obtained by combining the second vector representing the appearance frequency of each;
The computer further functions as a seventh step of generating a word pair cluster by division optimization clustering based on the similarity between vectors.

According to another embodiment of the binary relation classification program of the present invention,
A word is a noun,
A word pair is a noun pair,
A dependency set is a predicate set,
A predicate is a verb or adjective,
It is also preferred that the computer further functions so that the feature dependency set is a feature predicate set.

According to another embodiment of the binary relation classification program of the present invention,
A word is a predicate that is a verb or an adjective,
A word pair is a predicate pair,
The dependency language set is a noun set,
It is also preferable that the computer further functions so that the feature dependency set is a feature noun set.

According to another embodiment of the binary relation classification program of the present invention,
Regarding the seventh step, it is also preferable to further cause the computer to function so as to cluster only pairs that are equal to or greater than a predetermined threshold using mutual information as the similarity in the second step.

According to the present invention, in a binary relation classification method in an apparatus for classifying word pairs into semantic binary relations,
It has a sentence set storage unit that stores a large number of document information,
A first step of extracting a plurality of word pairs consisting of a first word and a second word from the sentence set storage unit;
A second step of extracting a word pair whose similarity representing the ease of co-occurrence among word pairs is equal to or greater than a predetermined threshold;
For the word pairs extracted in the second step, the first dependency word set co-occurs in the first word and the second dependency word set co-occurs in the second word from the sentence set accumulation unit. A third step of extracting
A first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set; and a first feature dependency word set that appears in the second dependency word set and the first A fourth step of extracting a second characteristic dependency word set consisting of dependency words that do not appear in the dependency word set;
For each dependency word that belongs to the first feature dependency word set, the frequency of occurrence in the document set that co-occurs with the first word, and for each dependency word that belongs to the second feature dependency word set, the second A fifth step of counting the frequency of occurrence in the document set co-occurring with
A first vector representing the appearance frequency of each dependency word belonging to the first feature dependency word set based on the first word, and a dependency word belonging to the second feature dependency word set based on the second word A sixth step of deriving a vector obtained by combining the second vector representing the appearance frequency of each;
And a seventh step of generating a word pair cluster by division optimization clustering based on the similarity between vectors.

According to the present invention, in a binary relation classification apparatus in an apparatus for classifying word pairs into semantic binary relations,
A sentence set storage means for storing a large number of document information;
Word pair extraction means for extracting a plurality of word pairs consisting of the first word and the second word from the sentence set storage unit;
Similar word pair extraction means for extracting word pairs whose similarity indicating the ease of co-occurrence among word pairs is equal to or greater than a predetermined threshold;
With respect to the word pairs extracted by the similar word pair extraction unit, the first dependency word set co-occurs with the first word and the second dependency word co-occurs with the second word from the sentence set accumulation unit. A dependency word set extraction means for extracting the set;
A first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set; and a first feature dependency word set that appears in the second dependency word set and the first Feature dependency word set extraction means for extracting a second feature dependency word set consisting of dependency words that do not appear in the dependency word set;
For each dependency word that belongs to the first feature dependency word set, the frequency of occurrence in the document set that co-occurs with the first word, and for each dependency word that belongs to the second feature dependency word set, the second A dependency word appearance frequency counting means for counting the appearance frequency in the document set co-occurring with
A first vector representing an appearance frequency for each dependency word belonging to the first feature dependency word set based on the first word, and a word for each word belonging to the second feature dependency word set based on the second word Word pair similarity calculating means for deriving a vector obtained by combining the second vector representing the appearance frequency;
It has a word pair cluster generation means for generating a word pair cluster by division optimization clustering based on similarity between vectors.

  According to the binary relation classification program, method, and apparatus of the present invention, word pairs that are semantically similar can be classified into a binary relation without pre-defining the inter-word relation to be acquired.

It is a flowchart showing the process in the binary relation classification program of this invention. It is a flowchart of the production | generation of a noun pair cluster. It is a flowchart of the production | generation of a predicate pair cluster. It is a functional block diagram of the binary relation classification apparatus in this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  The present invention can extract word pairs that are semantically similar from a set of sentences and classify the word pairs into binary relations having a semantic relationship.

Here, “word pairs” refer to pairs of “words”. A “word” is one of the structural units of a language and consists of one or more morphemes. A morpheme is the smallest unit that has a meaning for a language, and refers to each of a set of phonemes that have been decomposed into units that no longer make sense after further decomposition. There are simple words composed of one morpheme and compound words composed of a plurality of morphemes. In the following, when referring to “words”, both simple words and compound words are targeted.
Simple word: "Mountain"
Compound word: “Mountain climbing”

Phrases, clauses, sentences and sentences are composed by gathering words. For example, the sentence “go to Yoshinoyama” is composed of three words “Yoshinoyama”, “ni” and “go”. The words can be broadly classified into function words having a grammatical role and content words having other general meanings. Hereinafter, the term “word” refers to a content word.
Examples of content words: noun (Yoshinoyama), verb (go), adjective (beautiful)
Examples of function words: particles (ga, o, ni, no), auxiliary verbs (being, being able to be)

“Word pair” refers to a pair of such words, for example, the following.
Noun pairs: (Sakura, Yoshino cherry), (Beer party, green soybeans)
Verb / adjective pairs: (sleep, wake up), (fast, wake up), (fast, fast)

In general, the “binary relationship” refers to a relationship such as “entailment, synonym, synonym, causality, time” defined by FellBaum (see, for example, Non-Patent Document 7). On the other hand, according to the present invention, not only such binary relations but also surprising semantic relations that cannot be defined manually can be classified.
(A) Relationship between nouns (Sakura, Yoshino Yoshino): Upper and lower relationship (b) Relationship between verbs (sleeping, getting up): Causal relationship (c) Relationship between nouns (beer party, edamame): “Essential tool for event-event For example, the above-mentioned (a) and (b) are relationships belonging to the relationship classification by FellBaum et al. On the other hand, the above-mentioned (c) is not a relationship belonging to the relationship classification by FellBaum et al. According to the present invention, it is possible to classify in consideration of such a semantic relationship, and no huge manual cost is required for defining the relationship classification.

  FIG. 1 is a flowchart showing processing in the binary relation classification program of the present invention.

  The binary relation classification program is executed by a processor (computer) installed in the apparatus to classify “word pairs” made up of word pairs into semantic binary relations. The apparatus includes a sentence set storage unit that stores a large number of pieces of sentence information.

  According to the present invention, word pairs that are likely to co-occur above a predetermined threshold are extracted from the sentence set storage unit. Next, a dependency word set that co-occurs for each word included in the word pair is extracted from the sentence set storage unit. Then, a word pair cluster is generated for each word based on a vector representing the appearance frequency of the dependency word set.

The binary relation classification program executes the following seven steps.
(S1) A plurality of word pairs are extracted from the sentences stored in the sentence set storage unit. A “word pair” consists of a first word and a second word.
(S2) Only word pairs in which the first word and the second word are likely to co-occur are extracted from the extracted word pairs.
(S3) A first dependency word set co-occurring with the first word and a second dependency word set co-occurring with the second word from the sentence set accumulation unit for the word pair extracted in S2. And are extracted.
(S4) a first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set, and appear in the second dependency word set; A second feature dependency word set consisting of dependency words that does not appear in the first dependency word set combination is extracted.
(S5) For each dependency word belonging to the first characteristic dependency word set, the frequency of appearance of the dependency word co-occurring with the first word in the sentence stored in the sentence set storage unit is counted. . Similarly, for each dependency word belonging to the second feature dependency word set, the appearance frequency of the dependency word co-occurring with the second word in the sentences stored in the sentence set storage unit is counted. .
(S6) A first vector representing the appearance frequency of each dependency word belonging to the first feature dependency word set based on the first word is generated. Similarly, a second vector representing the appearance frequency for each dependency word belonging to the second feature dependency word set based on the second word is generated. A vector obtained by combining the first vector and the second vector is derived.
(S7) A word pair cluster is generated by division optimization clustering based on the similarity between vectors generated in S6.

  FIG. 2 is a flowchart for generating a noun pair cluster.

  In FIG. 2, the case of generating a noun pair cluster will be described as an example for generating a word pair cluster. In FIG. 2, the “word” in FIG. 1 is specifically described as “noun” and the “dependency word” is described as “predicate”.

(S21) A plurality of noun pairs are extracted from the sentences stored in the sentence set storage unit. The “noun pair” includes a first noun n1 and a second noun n2. For example, the following lexical syntactic pattern is used for extracting noun pairs.
Pattern: “<n1><n2>”
“<N1> to <n2>”

For example, the following noun pairs are extracted from the text stored in the text set storage using the pattern “<n2><n2>”.
[Sentence] [noun pair]
"I want to see Yoshinoyama's cherry blossoms"==> (Yoshinoyama, Sakura)
“The Yoshino cherry tree in Shinjuku Gyoen is beautiful” ==> (Shinjuku Gyoen, Yoshino cherry tree)
"Cherry blossoms in the garden"==> (Garden, cherry blossom)

(S22) From the extracted noun pairs, only noun pairs in which the first noun and the second noun are likely to co-occur are extracted. Mutual information can be used as a kind of similarity to measure “ease of co-occurrence” of two words. Only noun pairs whose degree of similarity measured for co-occurrence is equal to or greater than a predetermined threshold are extracted.

  By using the mutual information, noun pairs that are likely to co-occur in various sentences, such as (Yoshinoyama, Sakura), have high similarity. On the other hand, noun pairs that co-occur only in specific sentences, such as (Mr. Sumida, socks), have a low similarity. As a result, noun pairs having some semantic relationship can be extracted.

(S23) For the noun pair extracted in S22, a first predicate set that co-occurs with the first noun and a second predicate set that co-occurs with the second noun are extracted from the sentence set storage unit. The

  For example, classification is made for each semantic relationship based on the difference in the concept indicated by the two words “Somei Yoshino” and “Sakura”. However, it is difficult to express explicitly the range of concepts represented by words. To that end, assume that a set of predicates that co-occur with a noun represents a concept. According to the second prior art described above, the relationship is estimated from the common part of the set. On the other hand, according to the present invention, the difference between the predicate sets co-occurring with each noun is explicitly used as representing some semantic relationship. Also, in the present invention, in order to use the set of noun pairs that are likely to co-occur (highly similar) extracted in S22, classification is performed based on the semantic relationship in consideration of the range of concepts represented by words. Can do.

For example, for the noun pair <Yoshinoyama, Sakura>, a first predicate set that co-occurs with the first noun "Yoshinoyama" and a second predicate set that co-occurs with the second noun "Sakura" Extracted as follows.
Noun pair <Yoshinoyama, Sakura>
Noun “Yoshinoyama” ==> Predicate set {go, plant, stop, bloom}
Noun "Sakura"==> Predicate set {blooming, planting, protecting, seeing}
Noun pair <Shinjuku Gyoen, Yoshino Yoshino>
Noun "Shinjuku Gyoen"==> Predicate set {go, improve, stop by}
Noun "Yoshino Yoshino"==> Predicate set {blooming, planting, protecting, seeing}
Noun pair <garden, cherry blossom>
Noun “garden” ==> Predicate set {care, planting, cleaning}
Noun "Sakura"==> Predicate set {blooming, planting, seeing}

(S24) Next, a first feature predicate set consisting of predicates that appear in the first predicate set and do not appear in the second predicate set, and appear in the second predicate set and appear in the first predicate set A second feature predicate set consisting of predicates not to be extracted is extracted.

For example, for the noun pair <Yoshinoyama, Sakura>, all predicate sets include {blooming, planting}, so these predicates are deleted. Similarly, the above-described noun pair has the following feature predicate set.
Noun pair <Yoshinoyama, Sakura>
Noun “Yoshinoyama” ==> Feature Predicate Set {Go, Stop}
Noun “Sakura” ==> Feature Predicate Set {Protect, See}
Noun pair <Shinjuku Gyoen, Yoshino Yoshino>
Noun “Shinjuku Gyoen” ＝＝ ＞ Characteristic Predicate Set {Go, Maintain, Stop}
Noun "Yoshino Yoshino"==> Feature predicate set {blooming, planting, protecting, seeing}
Noun pair <garden, cherry blossom>
Noun “garden” ==> Feature predicate set {care, clean}
Noun "Sakura"==> Feature predicate set {blooming, seeing}

(1) Similarity between the noun pair <Yoshinoyama, Sakura> and the noun pair <Shinjuku Gyoen, Yoshino Yoshino> Any noun pair is common to the first feature predicate set that co-occurs with the first noun. The second feature predicate set that co-occurs with the second noun includes “see” and “protect” in common. Therefore, it is determined that these noun pairs have high similarity.

(2) Similarity between the noun pair <Yoshinoyama, Sakura> and the noun pair <Niwa, Sakura> Both noun pairs are common to the second noun "Sakura". However, the first feature predicate set that co-occurs with the first noun “Yoshinoyama” of the noun pair <Yoshinoyama, Sakura> is a predicate related to a place like “go” or “stop”. On the other hand, the first feature predicate set that co-occurs with the first noun “garden” of the noun pair <garden, cherry blossoms> includes landscaping predicates such as “care” and “clean”. That is, the predicates common to the two noun pairs are different. Therefore, it is determined that these noun pairs have low similarity.

  In this way, for the two noun pairs <Yoshinoyama, Sakura> and <Niwa, Sakura>, the predicate that does not co-occur with the paired nouns is extracted as a feature predicate set, even if they are the same noun “Sakura”. Is done. That is, even if the nouns are exactly the same, the feature predicate sets differ depending on the pair of nouns.

  If the noun pair is highly similar, as in (1) the noun pair <Yoshinoyama, Sakura> and the noun pair <Shinjuku Gyoen, Yoshino Yoshino>, it is determined that these noun pairs have a common semantic relationship. Is done.

(S25) For each predicate belonging to the first feature predicate set, the appearance frequency of the predicate co-occurring with the first noun in the sentences stored in the sentence set storage unit is counted. Similarly, for each predicate belonging to the second feature predicate set, the appearance frequency of the predicate co-occurring with the second noun in the text stored in the text set storage unit is counted.

For example, the frequency of appearance of expressions in which the noun “Yoshinoyama” and the predicate “go” are directly related is counted in the sentences stored in the sentence set storage unit 10. For example, it is expressed as follows.
freq (Yoshinoyama, go) = 132 times
freq (Yoshinoyama, stop by) = 76 times
freq (cherry tree, protect) = 63 times
freq (cherry tree, see) = 142 times

(S26) A predicate belonging to the first feature predicate set (np1 ′) based on the first noun and a first vector (freg_np1 ′) representing the appearance frequency for each predicate belonging to the first feature predicate set are derived. The Similarly, a predicate belonging to the second feature predicate set (np2 ′) based on the second noun and a second vector (freg_np2 ′) representing the appearance frequency for each predicate belonging to the second feature predicate set are derived. The

Each term in the vector corresponds to a predicate and is expressed as follows:
freq (n, p): frequency of occurrence of predicate p co-occurring with noun n
freq_np = [freq (n, p1), freq (n, p2) .....] ^T

The vector for each noun is expressed as follows:
Vector for the noun “Yoshinoyama”: freq_np1 '= [132,76] ^T
Vector for the noun "sakura": freq_np2 '= [63,142] ^T

Then, the generated vectors freq_np1 ′ and freq_np2 ′ are combined so that their dimensions are different.
f (Yoshinoyama, cherry blossoms) = [Go, Stop, Protect, See] ^T
f (Yoshinoyama, Sakura) = [132, 76, 63, 142] ^T

(S27) A noun pair cluster is generated by split optimization clustering based on the similarity between vectors derived in S26.

In S27, for the noun pair (n1, n2), the cluster ID (identifier) and the contribution degree in the cluster to which the noun pair belongs are acquired. Vector generated in S26, the noun pair (Yoshino, cherry) for a [132,76,63,142] ^T, noun pair (Shinjuku Gyoen, Yoshino) for, in [130,78,63,140] ^T Suppose there is. Here, it is assumed that the noun pair (Yoshinoyama, Sakura) and the noun pair (Shinjuku Gyoen, Yoshino Yoshino) belong to the same cluster by clustering based on the similarity between vectors. In this case, when the same ID is obtained as the cluster ID to which the cluster belongs, the cluster contribution is obtained for each noun pair.
Noun pair Vector Cluster ID Cluster contribution
f (Mount Yoshino, Sakura) = [132,76,63,142] ^T r1 0.8
f (Shinjuku Gyoen, Yoshino Yoshino) = [130,78,63,140] ^T r1 0.85

  According to the above example, the noun pair (Yoshinoyama, Sakura) and the noun pair (Shinjuku Gyoen, Yoshino Yoshino) have high similarity between vectors, so they belong to the same cluster as having some common semantic relationship. Be classified. On the other hand, noun pairs (garden, cherry blossoms) are classified as belonging to different clusters because their similarity between vectors is lower than noun pairs (Yoshinoyama, cherry blossoms) and noun pairs (Shinjuku Gyoen, Yoshino Yoshino). The In this way, noun pairs can be clustered assuming that the difference between the concepts represented by the first and second nouns in the noun pair means some semantic relationship.

  As a clustering technique, for example, K-means or EM algorithm (probabilistic clustering) can be used. According to K-means, clustering can be performed without using learning data. On the other hand, according to the EM algorithm, teacher learning can be performed by giving a cluster to which a noun pair belongs in advance as learning data. In the case of K-means, clustering can be performed with unknown relationships that are close in semantic relationship but do not belong to existing classifications. On the other hand, in the case of the EM algorithm, it is possible to perform clustering based on a relationship designed in advance by learning data, for example, an upper-lower relationship or a partial overall relationship.

  Each cluster obtained by clustering is regarded as a set of noun pairs representing some same semantic relationship. Each cluster is assigned a different cluster ID.

  Also, each noun pair is given a contribution to each cluster. The degree of contribution varies depending on the clustering method. In the case of K-means, each noun pair belonging to a cluster and the distance from the center of gravity of the cluster correspond to the contribution to the cluster. On the other hand, in the case of the EM algorithm, the affiliation probability of each noun pair to each cluster corresponds to the contribution to the cluster.

  FIG. 3 is a flowchart of predicate pair cluster generation.

  In FIG. 3, generation of word pair clusters will be described by taking an example of generating predicate pair clusters. In FIG. 3, the “word” in FIG. 1 is specifically described as “predicate”, and the “dependency word” is described as “noun”.

  In FIG. 2, the example which clustered based on the similarity of noun pairs using the feature predicate set which co-occurs with each noun was demonstrated. On the other hand, FIG. 3 illustrates an example of clustering based on the similarity between predicate pairs using feature noun sets that co-occur with each predicate. FIG. 3 is exactly the same as the contents of the process in FIG. 2 except that the noun pair differs from the predicate pair.

(S31) A plurality of predicate pairs are extracted from the sentences stored in the sentence set storage unit. The “predicate pair” includes a first predicate p1 and a second predicate p2. For example, the following lexical syntactic pattern is used for extracting predicate pairs.
Pattern: "<p1> while <p2>"
“<P1> to <p2>”

For example, the following predicate pairs are extracted from the sentences stored in the sentence set storage unit using the pattern “<n2><n2>”.
[Sentence] [predicate pair]
“Walking Mt. Yoshino, see the cherry blossoms” ==> (walk, see)
“Go to Shinjuku Gyoen and see the cherry blossoms” ==> (Go, see)

(S32) Only predicate pairs in which the first predicate and the second predicate are likely to co-occur are extracted from the extracted predicate pairs. As a kind of similarity for measuring “ease of co-occurrence” of two words, the mutual information amount can be used as in S22 of FIG. Only predicate pairs in which the similarity measured for co-occurrence is equal to or greater than a predetermined threshold are extracted.

(S33) For the predicate pair extracted in S32, a first noun set co-occurring with the first predicate and a second noun set co-occurring with the second predicate are extracted from the sentence set storage unit. The

  For example, it is assumed that a set of nouns co-occurring with predicates represents a concept in order to classify each semantic relationship based on the difference in concept indicated by two words “walking” and “going”. According to the present invention, the difference between noun sets co-occurring with each predicate is explicitly used to represent some semantic relationship. In the present invention, in order to use the set of predicate pairs that are likely to co-occur (highly similar) extracted in S32, classification is performed based on the semantic relationship in consideration of the range of concepts represented by the words. Can do.

For example, for the predicate pair <walk, see>, the first noun set co-occurring with the first predicate “walk” and the second noun set co-occurring with the second predicate “view” are as follows: Extracted as follows.
Predicate pair <walk, see>
Predicate "walking"==> noun set {park, road, mountain, town}
Predicate "see"==> noun set {flower, mountain, forest, town}
Predicate pair <go, see>
Predicate "go"==> noun set {company, school, mountain, town}
Predicate "see"==> noun set {flower, mountain, forest, town}

(S34) Next, a first feature noun set consisting of nouns appearing in the first noun set and not appearing in the second noun set, and appearing in the second noun set and appearing in the first noun set A second feature noun set consisting of nouns to be extracted is extracted.

For example, for the predicate pair <walking, seeing>, {nouns, towns} are included in all noun sets, so these nouns are deleted. Similarly, the predicate pair described above has the following characteristic noun set.
Predicate pair <walk, see>
Predicate "walking"==> characteristic noun set {park, road}
Predicate "see"==> feature noun set {flower, forest}
Predicate pair <go, see>
Predicate "go"==> noun set {company, school}
Predicate "see"==> noun set {flower, forest}

(S35) For each noun belonging to the first characteristic noun set, the frequency of appearance of the noun co-occurring with the first predicate in the sentence accumulated in the sentence set accumulating unit co-occurring with the first predicate is counted. Is done. Similarly, for each noun belonging to the second characteristic noun set, the frequency of appearance of the noun co-occurring with the second predicate in the sentence accumulated in the sentence set accumulating unit co-occurring with the second predicate is counted. Is done.

For example, the appearance frequency of expressions in which the predicate “walking” and the noun “park” are directly related is counted in the sentences accumulated in the sentence set accumulation unit. For example, it is expressed as follows.
freq (walking, park) = 128 times
freq (walking, road) = 60 times
freq (see, flower) = 48 times
freq (see, forest) = 122 times

(S36) A vector representing the appearance frequency for each noun belonging to the first feature noun set based on the first predicate, and a vector representing the appearance frequency for each noun belonging to the second feature noun set based on the second predicate Is derived.

Each term in the vector corresponds to a noun and is expressed as follows:
freq (p, n): Frequency of occurrence of noun n co-occurring with predicate p
freq_pn = [freq (p, n1), freq (p, n2) .....] ^T

The vector for each noun is expressed as follows:
Vector for predicate "walking": freq_pn1 '= [128,60] ^T
Vector for the predicate “look”: freq_pn2 ′ = [48,122] ^T

The generated vectors freq_pn1 ′ and freq_pn2 ′ are combined so that their dimensions are different.
f (walk, look) = [park, road, flower, forest] ^T
f (walking, looking) = [128, 60, 48, 122] ^T

(S37) Predicate pair clusters are generated by split optimization clustering based on the similarity between vectors derived in S36.

In S37, for the predicate pair (p1, p2), the cluster ID and the contribution degree in the cluster to which the predicate pair belongs are acquired. Here, it is assumed that the predicate pair (walking, seeing) and (going, seeing) belong to the same cluster by clustering based on the similarity between vectors. Similar to S27, when the same ID is obtained as the cluster ID to which the cluster belongs, the cluster contribution is obtained for each noun pair.
Noun pair Vector Cluster ID Cluster contribution
f (walking, looking) = [128,60,48,122] ^T r1 0.9
f (go, see) = [130, 60, 45, 121] ^T r1 0.7

  There is no significant difference in processing between the clustering of noun pairs in S27 and the clustering of predicate pairs in S37. The difference is that, in S27, noun pairs are clustered based on the vector similarity based on the appearance frequency for each predicate, whereas in S37, the predicate pairs are clustered based on the vector similarity based on the appearance frequency for each noun.

  The word pair clustering (S7), the noun pair clustering (S27), and the predicate pair clustering (S37) described above are limited to hard clustering in which one element must correspond one-to-one with one cluster. I can't. Soft clustering in which one element belongs to a plurality of clusters can also be used. In the case of hard clustering, the similarity is 0 when noun pairs or predicate pairs belong to different clusters between two vectors. On the other hand, in the case of soft clustering, since one noun pair can belong to a plurality of clusters, the number of vector pairs whose similarity is 0 can be reduced.

  FIG. 4 is a functional configuration diagram of the binary relation classification apparatus according to the present invention.

  According to FIG. 4, the binary relation classification apparatus 1 includes a sentence set accumulation unit 10, a word pair extraction unit 11, a similar word pair extraction unit 12, a dependency word set extraction unit 13, and a feature dependency word set. An extraction unit 14, a word appearance frequency counting unit 15, a word pair similarity calculation unit 16, and a word pair cluster generation unit 17 are included. These functional components other than the sentence set accumulating unit 10 may be realized by executing a binary relation classification program that causes a computer installed in the apparatus to function.

  The sentence set storage unit 10 stores a large amount of sentence information.

  The word pair extraction unit 11 extracts a plurality of word pairs composed of the first word and the second word from the sentence set storage unit 10 (see S1 in FIG. 1 described above). The extracted word pairs are output to the similar word pair extraction unit 12.

  The similar word pair extraction unit 12 inputs the extracted word pair (see S2 in FIG. 1 described above). The similar word pair extraction unit 12 extracts word pairs in which the similarity indicating the ease of co-occurrence among word pairs is equal to or greater than a predetermined threshold. The extracted word pairs are output to the dependency word set extraction unit 13.

  The dependency word set extraction unit 13 inputs the extracted word pairs (see S3 in FIG. 1 described above). The dependency word set extraction unit 13 refers to the sentence set accumulation unit 10, and for the input word pair, from the sentence set accumulation unit 10, the first dependency word set co-occurs with the first word, and the second And a second dependency word set that co-occurs with the word. The extracted first dependency word set and second dependency word set are output to the feature dependency word set extraction unit 14.

  The feature dependency word set extraction unit 14 inputs the first dependency word set and the second dependency word set (see S4 in FIG. 1 described above). The feature dependency word set extraction unit 14 includes a first feature dependency word set composed of dependency words that appear in the first dependency word set and do not appear in the second dependency word set, and a second dependency. A second characteristic dependency language set consisting of dependency words that appear in the response language set and do not appear in the first dependency language set is extracted. The extracted first feature dependency word set and second feature dependency word set are output to the dependency word appearance frequency counting unit 15.

  The word appearance frequency counting unit 15 inputs the first feature dependency word set and the second feature dependency word set (see S5 in FIG. 1 described above). The dependency word appearance frequency counting unit 15 refers to the sentence set storage unit 10 for each dependency word belonging to the first characteristic dependency word set, and the dependency word is the first in the stored sentence set. Count the frequency of occurrences with the words. Similarly, the dependency word appearance frequency counting unit 15 refers to the sentence set storage unit 10 for each dependency word belonging to the second characteristic dependency word set, and the dependency word is included in the stored sentence set. Counts the frequency of occurrence of co-occurring with the second word. The counted appearance frequency of each dependency word belonging to the first feature dependency word set and each dependency word belonging to the second feature dependency word set is output to the word pair similarity calculation unit 16.

  The word pair similarity calculation unit 16 inputs the counted appearance frequency (see S6 in FIG. 1 described above). The word pair similarity calculation unit 16 includes a first vector representing an appearance frequency for each dependency word belonging to the first feature dependency word set based on the first word, and a second feature based on the second word. A vector obtained by combining the second vector representing the appearance frequency for each dependency word belonging to the dependency word set is derived. The generated dependency word vector is output to the word pair cluster generation unit 17.

  The word pair cluster generation unit 17 inputs a dependency word vector (see S7 in FIG. 1 described above). The word pair cluster generation unit 17 refers to the similar word pair extraction unit 12 and is stored in the similar word pair extraction unit 12 based on the input dependency word vector by division optimization clustering based on the similarity between vectors. Cluster word pairs. The clustered word pairs are accumulated in the word pair cluster generation unit 17.

  As described above in detail, according to the binomial relationship classification program, method, and apparatus of the present invention, a pair of terms that are semantically similar can be binarized without predefining the inter-word relationship to be acquired. Can be classified.

  According to the present invention, it is easy to extract a user's intention and discover a user's hidden behavior by acquiring various semantic relationships. Therefore, the present invention can provide, for example, a search keyword expansion function of a question answering system.

  For example, it is assumed that “beer party” is input as a search keyword by the user. From the acquired semantic relationship, it can be seen that the relationship between the nouns “beer party” and “edamame” is the “event-event essential tool” relationship. Thereby, it can be extracted that “edamame” is an essential tool for “beer party”. Therefore, by adding “edamame” to the search query, noise can be reduced from the search results.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Binary relation classification apparatus 10 Text set accumulation | storage part 11 Word pair extraction part 12 Similar word pair extraction part 13 Dependence word set extraction part 14 Feature dependency word set extraction part 15 Word appearance frequency counting part 16 Word pair similarity calculation part 17 word pair cluster generator

Claims (6)

In order to classify word pairs into semantic binary relations, a binary relation classification program that causes a computer installed in a device to execute
It has a sentence set storage unit that stores a large number of document information,
A first step of extracting a plurality of word pairs consisting of a first word and a second word from the sentence set storage unit;
A second step of extracting a word pair having a degree of similarity that is equal to or greater than a predetermined threshold among the word pairs;
A first dependency word set that co-occurs in the first word and a second dependency co-occurs in the second word from the sentence set accumulation unit for the word pair extracted in the second step. A third step of extracting a word set;
A first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set; and a first feature dependency word set that appears in the second dependency word set and the first A fourth step of extracting a second characteristic dependency word set consisting of dependency words that do not appear in the dependency word set;
For each dependency word belonging to the first feature dependency word set, the appearance frequency in the document set co-occurring with the first word and for each dependency word belonging to the second feature dependency word set, A fifth step of counting the frequency of occurrence in the document set that co-occurs with two words;
A first vector representing the appearance frequency of each dependency word belonging to the first feature dependency word set based on the first word, and a dependency word belonging to the second feature dependency word set based on the second word A sixth step of deriving a vector obtained by combining the second vector representing the appearance frequency of each;
A binary relation classification program that causes a computer to function as a seventh step of generating a word pair cluster by division optimization clustering based on similarity between vectors. The word is a noun,
The word pair is a noun pair,
The dependency set is a predicate set;
Predescriptor is a verb or adjective,
The binary relation classification program according to claim 1, further causing the computer to function so that the feature dependency word set is a feature predicate set. The word is a predicate that is a verb or an adjective;
The word pair is a predicate pair;
The dependency language set is a noun set,
The binary relation classification program according to claim 1, further causing the computer to function so that the feature dependency word set is a feature noun set.   4. The computer according to claim 1, wherein the computer further functions to cluster only pairs having a predetermined threshold value or more by using mutual information as the similarity in the second step. The binary relation classification program according to item 1. In a binary relation classification method in a device for classifying word pairs into semantic binary relations,
It has a sentence set storage unit that stores a large number of document information,
A first step of extracting a plurality of word pairs consisting of a first word and a second word from the sentence set storage unit;
A second step of extracting a word pair having a degree of similarity that is equal to or greater than a predetermined threshold among the word pairs;
A first dependency word set that co-occurs in the first word and a second dependency co-occurs in the second word from the sentence set accumulation unit for the word pair extracted in the second step. A third step of extracting a word set;
A first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set; and a first feature dependency word set that appears in the second dependency word set and the first A fourth step of extracting a second characteristic dependency word set consisting of dependency words that do not appear in the dependency word set;
For each dependency word belonging to the first feature dependency word set, the appearance frequency in the document set co-occurring with the first word and for each dependency word belonging to the second feature dependency word set, A fifth step of counting the frequency of occurrence in the document set that co-occurs with two words;
A first vector representing the appearance frequency of each dependency word belonging to the first feature dependency word set based on the first word, and a dependency word belonging to the second feature dependency word set based on the second word A sixth step of deriving a vector obtained by combining the second vector representing the appearance frequency of each;
And a seventh step of generating a word-pair cluster by division optimization clustering based on similarity between vectors. In a binary relation classifier in a device that classifies word pairs into semantic binary relations,
A sentence set storage means for storing a large number of document information;
Word pair extraction means for extracting a plurality of word pairs consisting of a first word and a second word from the sentence set storage unit;
Similar word pair extraction means for extracting word pairs whose similarity indicating the ease of co-occurrence among the word pairs is equal to or greater than a predetermined threshold;
For the word pairs extracted by the similar word pair extraction means, a first dependency word set that co-occurs in the first word and a second co-occurrence in the second word from the sentence set storage unit. A dependency word set extraction means for extracting a dependency word set;
A first feature dependency word set consisting of dependency words that appear in the first dependency word set and not appear in the second dependency word set; and a first feature dependency word set that appears in the second dependency word set and the first Feature dependency word set extraction means for extracting a second feature dependency word set consisting of dependency words that do not appear in the dependency word set;
For each dependency word belonging to the first feature dependency word set, the appearance frequency in the document set co-occurring with the first word and for each dependency word belonging to the second feature dependency word set, A dependency word appearance frequency counting means for counting the appearance frequency in the document set co-occurring with two words;
A first vector representing an appearance frequency for each dependency word belonging to the first feature dependency word set based on the first word, and a word for each word belonging to the second feature dependency word set based on the second word Word pair similarity calculating means for deriving a vector obtained by combining the second vector representing the appearance frequency;
A binary relation classifying apparatus comprising: word pair cluster generation means for generating word pair clusters by division optimization clustering based on similarity between vectors.

Images