JP2015200961A - Semantic relationship extraction apparatus and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a semantic relation extraction device and a program capable of estimating the relation between those nouns regarding a noun pair appearing in a sentence.SOLUTION: A noun pair extraction unit extracts a noun pair contained in the same sentence in document data. A third noun extraction unit extracts a third noun from the sentence containing the noun pair extracted by the noun pair extraction unit. A seed pattern storage unit stores a pattern of a sentence structure in which a seed word pair, which is a noun pair having the known relation, appears as a seed pattern associated with the relation. A score calculation unit calculates a score showing a degree in which the noun pair has the relation associated with the seed pattern, based on the noun pair extracted by the noun pair extraction unit, the third noun extracted by the third noun extraction unit and the seed pattern stored in the seed pattern storage unit. A specific relation noun pair extraction unit extracts a noun pair having the predetermined relation based on the score.

Description

  The present invention relates to a semantic relationship extraction apparatus and program for extracting a semantic relationship between words from an input document.

In one field of natural language processing technology, attempts have been made to extract information from text data.
When trying to automatically estimate the relationship between two nouns appearing in a text, one conventional technique is a method that uses a pattern connecting two nouns as a clue. Here, the pattern is a sentence pattern such as “A (noun) causes B (noun)”.

  Non-Patent Document 1 describes a technique for acquiring a semantic relationship between words on a large scale by learning processing using such a pattern as a clue.

Stijn De Saeger, Kentaro Torisawa, Junichi Kazama, Wataru Kuroda, Maki Murata, “Acquiring large-scale semantic relationships by pattern learning using word semantic classes”, Proc. Of the 16th Annual Conference of the Language Processing Society, D4 -2, 2010, p. 932-p. 935

  In the conventional technique described above, a relationship between words is estimated using a pattern connecting two words as a clue. The number of types of patterns for the estimation is enormous. Since such various patterns exist, there is a problem that it is difficult to use a pattern with a low appearance frequency or a pattern in which a noun appears as a pattern for relationship estimation.

  Here, the relationship between words (nouns) includes, for example, a causal relationship, an upper-lower relationship, a relationship between a disease and a treatment method, a relationship between a disease and a preventive method, a relationship between a place and a specialty.

  The present invention has been made based on the above problem recognition, and provides a semantic relationship extraction apparatus and program capable of estimating a relationship between nouns that has been difficult to estimate in the past.

  [1] In order to solve the above-described problem, a semantic relationship extraction device according to an aspect of the present invention extracts a noun pair extraction unit that extracts noun pairs included in the same sentence in the document data based on the document data. And a third noun extraction unit that extracts a third noun from the sentence in which the noun pair extracted by the noun pair extraction unit is included, and a seed word pair that is a noun pair having a known relationship appears. A pattern of sentence structure to be stored as a seed pattern associated with the relation, a noun pair extracted by the noun pair extraction unit, and a third noun extracted by the third noun extraction unit And a score calculation unit that calculates a score indicating a degree that the noun pair has the relationship associated with the seed pattern based on the seed pattern stored in the seed pattern storage unit; Wherein on the basis of the score calculated by the score calculating unit comprises a, a specific relationship noun pair extraction unit for extracting the noun pair is estimated to have the relationship.

  [2] Further, according to one aspect of the present invention, in the semantic relation extraction device, the seed pattern is a tree-structured data representing a dependency relation between phrases, and the phrase between phrases belonging to the seed word pair. It is characterized by being expressed as a path.

  [3] Further, according to one aspect of the present invention, in the semantic relation extraction device, the score calculation unit is a value representing a degree of appearance of a class pair of nouns corresponding to the noun pair in the relation. One score is calculated, and the score based on the first score is calculated.

  [4] Further, according to one aspect of the present invention, in the semantic relationship extraction apparatus, the score calculation unit includes a pattern between one noun included in the noun pair and the third noun. A second score, which is a value indicating a degree of ease of appearing in the seed pattern, is calculated, and the score based on the second score is calculated.

  [5] In addition, according to one aspect of the present invention, in the semantic relationship extraction apparatus, the score calculation unit is configured to display a pattern between one noun and the third noun included in the noun pair. A third score, which is a value representing the likelihood of relationship or the likelihood of parallel relationship, is calculated, and the score based on the third score is calculated.

  [6] According to another aspect of the present invention, a noun pair extraction unit that extracts noun pairs included in the same sentence in the document data based on the document data, and the noun pair extraction unit extracts the noun pairs. A third noun extraction unit that extracts a third noun from a sentence in which the noun pair was included, and a sentence structure pattern in which a seed word pair that is a noun pair having a known relation appears is associated with the relation. Stored in the seed pattern storage unit, the seed pattern storage unit that stores the seed pattern, the noun pair extracted by the noun pair extraction unit, the third noun extracted by the third noun extraction unit, and the seed pattern storage unit Based on the seed pattern, a score calculator that calculates a score indicating the degree that the noun pair has the relationship associated with the seed pattern; and the score calculated by the score calculator Based on A, is a program for causing a specific relation noun pair extraction unit, as a computer for extracting the noun pair is estimated to have the relationship.

  According to the present invention, it is possible to automatically estimate the relationship between two words. That is, information extraction can be performed based on document data.

It is the block diagram which showed schematic function structure of the semantic relationship extraction apparatus by one Embodiment of this invention. It is a flowchart which shows the operation | movement procedure of the process which the semantic relationship extraction apparatus by the embodiment extracts the relationship between words. It is related with the process by the noun pair extraction part by the embodiment, and is the schematic which shows the example of the dependency relationship in a noun pair and a sentence. It is a flowchart which shows the procedure of the process in which the score calculation part by the embodiment calculates the score used as the parameter | index for determining whether a word has a specific relationship. It is the schematic which shows an example of the seed word pair which the score calculation part by the same embodiment utilizes.

Next, an embodiment of the present invention will be described with reference to the drawings.
When analyzing text and evaluating the similarity between texts, related words are useful. In the present embodiment, a new word pair having a relationship is extracted by using three words appearing in the same sentence and a word pair (seed word pair) having a relationship prepared in advance.

  FIG. 1 is a block diagram showing a schematic functional configuration of the semantic relationship extraction apparatus according to the present embodiment. As shown in the figure, the semantic relationship extraction device 1 includes a document acquisition unit 10, a noun extraction unit 11, a noun pair extraction unit 12, a third noun extraction unit 13, a score calculation unit 14, and a specific related noun pair extraction. A unit 15 and a specific related noun pair output unit 16 are included. The functions of these units are realized, for example, by appropriately combining logic operations using electronic circuits. In each unit, information can be stored by using a semiconductor memory or a magnetic hard disk drive (HDD) as appropriate.

  The semantic relationship extraction apparatus 1 also includes a seed pattern storage unit (not shown). The seed pattern storage unit stores a sentence structure pattern in which a seed word pair, which is a noun pair having a known relationship, appears as a seed pattern associated with the relationship. Details of the seed pattern will be described later (see also FIG. 4).

The document acquisition unit 10 captures an input document to be processed in the following.
The noun extraction unit 11 divides the input document acquired by the document acquisition unit 10 into sentences, and extracts nouns from the obtained sentences. When extracting a noun, the noun extraction unit 11 performs a morphological analysis process using an existing technology.
The noun pair extraction unit 12 extracts an arbitrary pair from the noun set obtained by the processing of the noun extraction unit 11. That is, the noun pair extraction unit 12 extracts any two nouns from the noun set. That is, the noun pair extraction unit 12 extracts noun pairs included in the same sentence in the document data based on the input document data.
The third noun extraction unit 13 extracts a third noun deeply related to the noun pair extracted by the noun pair extraction unit 12 from the same sentence. That is, the third noun extraction unit 13 extracts the third noun from the sentence in which the noun pair extracted by the noun pair extraction unit 12 was included.

  Whether the score calculation unit 14 has a target relationship using the pattern and relationship between the noun pair extracted by the noun pair extraction unit 12 and the third noun extracted by the third noun extraction unit 13. The score (score) for judging whether or not is calculated. In other words, the score calculation unit 14 is based on the noun pair extracted by the noun pair extraction unit 12, the third noun extracted by the third noun extraction unit 13, and the seed pattern stored in the seed pattern storage unit. The score indicating the degree that the noun pair has a relationship associated with the seed pattern is calculated.

  The specific related noun pair extraction unit 15 extracts a noun pair having a high score as a noun pair having a target relationship based on the score calculated by the score calculation unit 14. That is, the specific related noun pair extraction unit 15 extracts a noun pair estimated to have a specific relationship based on the score calculated by the score calculation unit 14.

  FIG. 2 is a flowchart showing an operation procedure of the semantic relationship extraction apparatus 1. Hereinafter, the procedure of the process of extracting the semantic relationship between words will be described with reference to this flowchart. As a premise of this series of processing, the document acquisition unit 10 has already acquired document data composed of a large amount of sentences.

  First, in step S1-1, the noun extraction unit 11 divides the document data acquired by the document acquisition unit 10 for each sentence. At this time, the noun extraction unit 11 divides the sentence into sentences by using symbols (“!”, “?”, Etc.) that serve as clues for phrases and sentence breaks. And the noun extraction part 11 performs the morphological analysis process of a sentence using the morphological analyzer by the existing technique, etc., and extracts the noun which appears in each sentence. For example, Mecab or the like can be used as the morphological analyzer.

  Next, in step S1-2, the noun pair extraction unit 12 generates one noun pair by combining any two of the nouns extracted by the noun extraction unit 11. The noun pair extraction unit 12 performs noun pair generation processing based on syntactic constraints in the original sentence. Specifically, the noun pair extraction unit 12 refers to the dependency structure of the original sentence in which the noun was included. Then, the noun pair extraction unit 12 includes a phrase that is equal to or more than a predetermined number (for example, three phrases) in a path to a common dependency destination of two phrases including each of the two nouns. Those two nouns are not output as noun pairs. A specific example of selecting a noun pair based on the dependency structure will be described later with reference to FIG.

  It should be noted that a syntax analyzer based on an existing technology can be used for the processing of dividing the clauses in the sentence and the analysis of the dependency relation between the clauses. As an example, the dependency analyzer cabocha described in the document [Taku Kudo, Yuji Matsumoto, Dependency Analysis by Chunking Stage Application, SIG-NL-142, 2001] can be used.

  Next, in step S <b> 1-3, the third noun extraction unit 13 selects a third noun that is closely related to the noun pair for one of the noun pairs extracted by the noun pair extraction unit 12. Extract one sentence from the same sentence as

At this time, the third noun extraction unit 13 extracts a noun corresponding to one of the following two. First, it is a noun in a phrase directly related to a phrase including one word of the noun pair, or a noun in a phrase indirectly related to a phrase (for convenience, the first noun Called the rules). Second, if there is a phrase in parallel with the phrase containing one word of the noun pair, it is the noun in the phrase that appears last in the parallel relationship (referred to as the second rule for convenience). ). When the second rule is applicable, only the second rule is applied, and the first rule is not applied. There can be a plurality of nouns corresponding to either the first or the second. And the 3rd noun extraction part 13 produces | generates the 3 term group of two nouns contained in the original noun pair, and the noun corresponding to either said 1st or 2nd.
Each of these first or second patterns will be described later with reference to FIG.

  Next, in step S <b> 1-4, the score calculation unit 14 calculates a score for the noun ternary group generated by the third noun extraction unit 13. This score is an index for determining whether or not a noun pair (noun pair extracted by the noun pair extraction unit 12) that is a component of the three-tuple group has a specific relationship (for example, a causal relationship). It will be.

  Next, in step S1-5, the semantic relationship extraction apparatus 1 determines whether there is another (unprocessed) third noun corresponding to the currently processed noun pair. If it exists (step S1-5: YES), the process returns to step S1-3 to calculate a score for another third noun. If it does not exist (step S1-5: NO), the process proceeds to the next step S1-6.

  Next, in step S1-6, the semantic relationship extraction apparatus 1 determines whether or not all the extracted noun pairs have been processed. If all noun pairs have not been processed (if there are unprocessed noun pairs, step S1-5: NO), the process returns to step S1-2 to perform processing for other noun pairs. If all noun pairs have been processed (step S1-6: YES), the process proceeds to the next step S1-7.

  Next, in step S1-7, the semantic relationship extraction apparatus 1 determines whether or not the entire sentence has been processed. If the entire sentence has not been processed (if there is an unprocessed sentence, step S1-7: NO), the process returns to step S1-1 in order to perform processing for another sentence. If the entire sentence has been processed (step S1-7: YES), the process proceeds to the next step S1-8.

Next, in step S1-8, the specific related noun pair extraction unit 15 extracts a noun pair having a target relationship. Here, for the noun pair n _i , n _k , the one having the maximum score all value (see the flowchart in FIG. 4 and the description thereof) described later is extracted. That is, the specific related noun pair extraction unit 15 calculates the score score (n _i , n _k ) for the noun pair n _i , n _{k according} to the following equation (1).

That is, Score (n _i , n _k ) is a value indicating the degree to which the noun pair (n _i , n _k ) has a specific relationship corresponding to the seed pattern SP. The seed pattern will be described later.

The calculation of the score all scoreAll (n _i , n _k , SP, n _j ) in the equation (1) will be described later with reference to FIG.

  Then, the specific relation noun pair extraction unit 15 determines that the noun pair whose score calculated by the equation (1) is a predetermined threshold or more is a noun pair having a target relation (specific relation). .

  FIG. 3 relates to processing by the noun pair extraction unit 12, and is a schematic diagram illustrating an example of a noun pair and a dependency relation in a sentence. Each of (A) to (C) in the figure shows tree-structured data representing the dependency relationship of sentences. In each tree structure, nodes (rectangular boxes) correspond to clauses included in the sentence. An edge (arrow) corresponds to a dependency relationship between phrases (in the case of a solid line) or a parallel relationship (in the case of a broken line).

  In the example of FIG. 4A, there are four clauses (phrase 1, clause 2, clause 3, clause 4) in the path to the common destination of the clause containing noun A and the clause containing noun B (phrase 4). Is included. That is, since the path length is longer than the reference “3”, the pair of nouns A and B appearing in this sentence is not extracted as a noun pair and is excluded from the target of subsequent processing. In this example, the phrase directly related to the noun A is phrase 1. The phrase directly related to the noun B is the phrase 4. The phrases indirectly related to the noun A are phrase 2, phrase 3, and phrase 4.

  In the example of FIG. 5B, three clauses (sentence 1, clause 2, and clause 3) are included in the path to the common destination of the clause containing the noun A and the clause containing the noun B (phrase 3). . That is, since the path length is “3” or less which is the reference, the pair of nouns A and B appearing in this sentence is not excluded and is extracted as a noun pair.

  In the example shown in FIG. 5C, three clauses (phrase 2, clause 3, and clause 4) are included in the path to the common destination of the clause containing noun A and the clause containing noun B (phrase 4). . The phrase including the noun B, the phrase 1, and the phrase 2 are in a parallel relationship. Also in this case, since the dependency-related path length is equal to or less than the reference “3”, the pair of nouns A and B appearing in this sentence is not excluded and is extracted as a noun pair. The path length combining the dependency relationship and the parallel relationship is 5, but here, a determination is made based on whether or not the path length of the dependency relationship is less than or equal to the reference.

Next, nouns that can be extracted by the third noun extraction unit 13 will be described with reference to FIG.
In the example shown in FIG. 5B, when all the clauses that are directly or indirectly dependent on the clauses including each of the noun pairs (noun A and noun B) are extracted, clause 1, clause 2, clause 3 is extracted. Combining the nouns included in these phrases with the nouns A and B, the noun triplets are generated. That is, the third noun extraction unit 13 adds (noun included in noun A, noun B, phrase 1), (noun included in noun A, noun B, phrase 2), (noun A, noun B, phrase 3). A noun) is created.
In the example shown in FIG. 5C, there is a phrase in parallel with the phrase including the noun B. Therefore, a noun ternary group is generated by combining the nouns in the phrase 2 which is the last of the parallel relation, the nouns A and B. That is, the 3rd noun extraction part 13 produces | generates the 3 term group called (the noun A, the noun B, the noun contained in the phrase 2).

  FIG. 4 is a flowchart showing a procedure of processing in which the score calculation unit 14 calculates a score.

Prior to performing the processing shown in the figure, a seed word pair is generated in advance as an example of a word pair having a specific relationship to be extracted. The seed word pair is data composed of a first word and a second word and information (relation name) representing the relationship between the two words. The relationship name is, for example, “causal relationship”. The process itself for generating seed word pairs is described in the literature [Stijn, Torizawa, Kazama, Kuroda, Murata, acquisition of large-scale semantic relations by pattern learning using word semantic classes, and 16th Annual Conference of the Association for Natural Language Processing. The method described in the collection of papers, D4-2, 2010] can be used. In addition, the seed word pair data may be created by enumerating word pairs that a person can think of (word pairs having a specific relationship).
An example of the seed word pair will be described later with reference to FIG.

  In step S3-1, the score calculation unit 14 reads seed word pair data and performs seed pattern generation processing based on the read data. Details of seed pattern generation are as described below.

  Here, a pattern is a path to a common dependency point between clauses to which two nouns belong in a dependency structure of an input sentence (expressed as a directed graph having a tree structure). Does not include the phrase to which the noun belongs). Among such patterns, a pattern that expresses a specific relationship between nouns or a pattern that expresses such a specific relationship is a seed pattern. That is, the seed pattern is represented as a path of phrases between words belonging to the seed word pair in the tree-structured data representing the dependency relationship between phrases.

  Examples of specific patterns are as follows. For example, as shown in FIG. 3A, a phrase including the noun A is related to another phrase 1, the phrase 1 is related to another phrase 2, the phrase 2 is related to another phrase 3, and the phrase 3 Is related to clause 4. A phrase including the noun B is related to the phrase 4 described above. At this time, the common destination of the noun A and the noun B is the phrase 4. Therefore, the path from the phrase containing noun A to the phrase containing noun B is {sentence 1, phrase 2, phrase 3, phrase 4} (sentences 1 to 4 are connected in series), This is a pattern.

  Of the patterns extracted in this way, a seed pattern is generated according to a predetermined criterion. For example, one of the first to third seed pattern generation methods described below is used.

  In the first seed pattern generation method, all patterns in which seed word pairs appear are generated as seed patterns. Equation (2) below represents a set of seed patterns generated by the first seed pattern generation method.

Here, (n _i , pat, n _j ) represents an event in which the noun pair (n _i , n _j ) co-occurs with the pattern pat. In the first seed pattern generation method, as shown in Expression (2), all patterns (pat) in which noun pairs included in seed word pairs (noun pairs, seedpair) co-occur are used as seed patterns.

  In the second seed pattern generation method, only patterns with high appearance frequency among patterns in which seed word pairs appear are generated as seed patterns. Equation (3) below represents a set of seed patterns generated by the second seed pattern generation method.

  Here, Freq (pat) is the number of appearances of the pattern pat. Freq (seedpair) is the number of appearances of the seed word pair. The numerator on the left side of the inequality in Expression (3) is the number of appearances of the pattern pat that co-occurs with two words included in the seed word pair in the input sentence. A value obtained by dividing this value by the number of occurrences of the seed word pair is larger than a predetermined threshold (β) is used as a seed pattern.

  In the third seed pattern generation method, among all patterns in which a seed word pair appears, a pattern having a high ratio of the corresponding pattern co-occurring with the seed word pair is generated as a seed pattern. Equation (4) below represents a set of seed patterns generated by the third seed pattern generation method.

  Here, the numerator on the left side of the inequality included in Equation (4) represents a pattern pat that co-occurs with two words included in the seed word pair. Similarly, the denominator on the left side indicates the number of appearances of the pattern pat. A seed pattern having a value on the left side larger than a predetermined threshold value (γ) is used.

  The score calculation unit 14 generates a set of a predetermined amount of seed patterns by any one of the seed pattern generation methods described above. Note that a set of seed patterns may be generated only by hand without using the seed pattern generation method. By storing the seed pattern set generated in this manner in a storage device (seed pattern storage unit), the seed pattern set can be used in common in the calculation of the score for each noun ternary group. The seed pattern storage unit is configured by a semiconductor memory, a hard disk device, or the like.

In the following processing, the score calculation unit 14 uses the generated seed pattern set to calculate a score that serves as an index for determining whether a certain word pair represents a target relationship. Here, it is assumed that the word pairs are n _i and _nk . Further, a third noun closely related to this noun pair is assumed to be n _j .
In the following, n _j may have a direct or indirect dependency relationship from n _k, or, the case where n _j is in parallel relationship with the n _k. However, if n _j has a directly or indirectly dependency relation from n _i, or, in the case where n _j is in parallel relationship with the n _i also exchanging the n _i and n _k in the equation below The score can be obtained by

Next, in step S3-2, the score calculation unit 14 calculates score 1 (score 1, first score). The score 1 is the ease of appearance of the class (c _i , c _k ) to which the two words (n _i , n _k ) belong to the seed pattern representing the target relationship. Here, a class is a broad word or a broad concept to which a word belongs. For example, the word “hypertension” corresponds to the class “disease”. A word may correspond to multiple classes. The relationship between words and classes can be extracted from document data using existing technology, or can be obtained using thesaurus data. The relationship between words and classes is stored in advance in the semantic relationship extraction apparatus 1 as data. Specifically, the score calculation unit 14 calculates the score 1 by the following equation (5).

Here, SP is a seed pattern set. N _i and n _k are word pairs. C _i and c _k are classes to which the words n _i and n _k belong, respectively. The numerator on the right side in Equation (5) is the number of co-occurrence with SP of words belonging to the class (c _i , c _k ). The denominator on the right side in the equation (5) is the number of co-occurrence with an arbitrary pattern of words belonging to the class (c _i , c _k ).

That is, the score 1 is a value representing the degree (possibility, probability) that the class pair (c _i , c _k ) corresponding to the noun pair (n _i , n _k ) appears in a specific relationship (target relationship). is there.

Next, in step S3-3, the score calculation unit 14 calculates score 2 (score 2, second score). This score 2 represents the ease of appearance of a pattern (with class restriction) between two words (n _i , n _j ) in the seed pattern. Specifically, the score calculation unit 14 calculates the score 2 by the following equation (6).

Here, I (P _ij , c _i , c _k ) is a pattern P between the word pairs n _i and n _{j in} the noun ternary group (n _i , n _j , n _k ) to be processed. _It represents a set of noun pairs belonging to c _i and c _k co-occurring with _ij (classes of n _i and n _k , respectively). I (SP (c _i , c _k )) represents a set of noun pairs belonging to classes c _i and c _k that co-occur with any of the seed patterns. I (P _ij ) is a noun pair set that co-occurs with the pattern P _ij between n _i and n _j . I (SP) is a noun pair set that co-occurs with any of the seed patterns.

That is, score 2 is a value calculated by {(A) × (B)} / {(C) × (D)} using the following (A) to (D).
(A) Noun pairs belonging to classes c _i and c _k co-occurring with either a set of noun pairs belonging to classes c _i and c _k co-occurring with pattern P _ij between word pairs n _i and n _j The number of elements in the intersection with the set.
(B) The number of elements in the intersection of the noun pair set co-occurring with the pattern P _ij between the word pairs n _i and n _j and the noun pair set co-occurring with any of the seed patterns.
(C) Noun pairs belonging to classes c _i and c _k co-occurring with either a noun pair set belonging to classes c _i and c _k co-occurring with pattern P _ij between word pairs n _i and n _j or seed patterns The number of elements in the union with the set.
(D) The number of elements in the union of the noun pair set co-occurring with the pattern P _ij between the word pairs n _i and n _j and the noun pair set co-occurring with any of the seed patterns.

That is, the score 2 is a value representing the degree (probability, probability) of the ease of appearance of the pattern P _ij between the two words (n _i , n _j ) in the noun ternary group in the seed pattern.
Is a value representing

Next, in step S3-4, the score calculation unit 14 calculates score 3 (score 3, third score). This score 3 represents the upper and lower relations of the pattern between two words (n _k , n _j ) and the likelihood of parallel relations. Specifically, the score calculation unit 14 calculates the score 3 by the following equation (7).

Here, probNoun _hyp (n _k , n _j ) indicates the probability that the two words (n _k , n _j ) are in the upper and lower relationship, and this probability is calculated from the co-occurrence probability with the upper and lower relationship pattern given in advance. I can do it. ProbPattern _hyp (p _kj ) is a probability indicating the likelihood of an upper-lower relationship of the pattern p _kj between two words (n _k , n _j ), and this probability is a co-occurrence with a higher-lower relationship word pair given in advance. It can be calculated from the probability. probNoun _para (n _k , n _j ) indicates the probability that two words (n _k , n _j ) are in a parallel relationship, and this probability can be calculated from the co-occurrence frequency with the parallel relationship pattern. probPattern _para (p _kj ) indicates the parallelism of the pattern of two words (n _k , n _j ), which extracts sibling word pairs from the upper and lower relation word pair sets and co-occurs with the sibling word pairs It can be obtained from the frequency.

  The reason for calculating the score 3 using the upper / lower relationship or the parallel relationship is that the relationship between the word contained in the original word pair and the third noun is an upper / lower relationship or a parallel relationship. This is because the third noun has an important effect for specifying the relationship between word pairs.

That is, a score of 3, 2 words _(n k, n _j) and Is ness Upper Lower relationship pattern _{p kj} between the two words, or two words _(n k, n _j) and the pattern _{p kj} between the two words This is the value that represents the greater of the parallelism.

  Next, in step S3-5, the score calculation unit 14 calculates score all. This score all is a value calculated by taking the product of score 1, score 2, and score 3 calculated in the above steps. Specifically, the score calculation unit 14 calculates the score all value by the following equation (8).

That is, the score ol _{ScoreAll (n i, n k,} SP, n j) is a noun pair _(n i, n _k) and, when the third noun _{n j} is given, noun pair _(n i, _{n k} ) Is a value representing the degree of having a specific relationship corresponding to the seed pattern SP.
And, as represented by the equation (8), the score all ScoreAll (n _i , n _k , SP, n _j ) is a product value of the score 1, the score 2, and the score 3. However, it is the value of the product when the product value is selected to be the maximum among the class pairs (c _i , c _k ) to which the noun pair (n _i , n _k ) belongs. As a premise thereof, each of the class sets class (n _i ) and class (n _k ) to which the nouns n _i and n _k belong is a set that can have a plurality of elements.

  FIG. 5 is a schematic diagram illustrating an example of a seed word pair described in the processing by the score calculation unit 14. As shown in the figure, the seed word pair includes a plurality of word pairs, and is represented, for example, as tabular data. The seed word pair data has data items (digits) of word 1, relation name, and word 2, and is composed of a plurality of rows. For example, the data in the first row indicates that the word “cholesterol” (word 1) and the word “hypertension” (word 2) have a relationship name “causal”. The data in the second row indicates that the word “dextrin” (word 1) and the word “cavities” (word 2) have a relation name “causal relationship”. In the example shown in the figure, word 1 corresponds to the cause and word 2 corresponds to the result, but the direction of the relationship may be reversed and arbitrary.

[Example of processing and operation of the configuration of this embodiment]
Here, an example of processing by the semantic relationship extraction apparatus 1 will be briefly described. As an example, consider the sentence "A substance such as cholesterol that causes high blood pressure ...". When this part of the sentence is subjected to dependency analysis, the result includes a dependency relation of [hypertension]-[cause]-[substance is] and a dependency relation of [such as cholesterol]-[substance is]. A tree structure is obtained. An example of a noun pair extracted based on such a sentence is (high blood pressure, cholesterol). An example of the extracted third noun is “substance”. One of the patterns obtained from this is [sentence including noun A] − “cause” − [sentence including noun B]. Such a pattern is included in the seed pattern corresponding to a specific relationship called a causal relationship in order to express the causal relationship well. Such a seed pattern is obtained from a seed word pair as described above.

  If the third noun is not used and a pattern (sentence path to the common contact point) is obtained only from the noun pair, (sentence including the noun "hypertension")-"cause"-"substance"-( Phrase containing the noun “cholesterol”).

  When the pattern using the third noun is compared with the pattern not using the third noun, the former is shorter in this way and has a specific relationship in statistical processing (in this example, a causal relationship). ) Is good for calculating the co-occurrence degree. One reason for this is that the pattern when the third noun is not extracted and used (phrase containing the noun “hypertension”)-“cause”-“substance is”-(phrase containing the noun “cholesterol”) is the path This is because the noun “substance” is included in the middle of this, which increases the number of types of patterns.

  The semantic relationship extraction apparatus 1 of the present embodiment includes a third noun extraction unit 13, and the score calculation unit 14 calculates a pattern between the extracted third noun and a word included in the original word pair. Calculate the score used. With this configuration, it is possible to obtain better results than in the prior art. In addition, the noun “cholesterol” and the third noun “substance” included in the noun pair are in a high-order subordinate relationship, and the value of the score 3 is high for such a third noun.

  In other words, the semantic relationship extraction device 1 uses the information of the upper and lower relations between the nouns included in the noun pair and the third noun (the upper part of cholesterol is a substance), and thereby the information of the noun pair is obtained. It has a technical configuration for estimating the relationship (in other words, calculating the degree of having a specific relationship (target relationship)).

  Note that the function of the semantic relationship extraction device in the above-described embodiment, or a part of the function, may be realized by a computer. In that case, the program for realizing the control function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible disk, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, a “computer-readable recording medium” dynamically holds a program for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included, and a program that holds a program for a certain period of time. The program may be a program for realizing a part of the above-described functions, or may be a program that can realize the above-described functions in combination with a program already recorded in a computer system.

  Although the embodiment has been described above, the present invention can also be implemented in the following modified example.

Modified example: In the above embodiment, when calculating the value of the score all ScoreAll (n _i , n _k , SP, n _j ) in the equation (8), the product of score 1, score 2 and score 3 is I used it. In this modification, instead of the product of score 1, score 2, and score 3, any one of the following values (A) to (F) is used.
(A) Product of score 1 and score 2 only (B) Product of score 2 and score 3 only (C) Product of score 1 and score 3 only (D) Value of score 1 itself (E) Value of score 2 itself ( F) Score 3 itself

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

  The present invention can be used to estimate the relationship between documents. For example, if the document is a web page (web document), the present invention can be used to estimate the association between web pages. If a document (for example, a program summary sentence document in EPG) is associated with a broadcast program, the present invention can be used to estimate the relationship between the broadcast programs. That is, it can be used to realize an associative search function that presents a broadcast program related to a certain broadcast program (for example, a program being viewed by a viewer). That is, the present invention can also be used for an on-demand distribution service for content such as video.

DESCRIPTION OF SYMBOLS 1 Semantic relationship extraction apparatus 10 Document acquisition part 11 Noun extraction part 12 Noun pair extraction part 13 3rd noun extraction part 14 Score calculation part 15 Specific relation noun pair extraction part 16 Specific relation noun pair output part

Claims (6)

Based on the document data, a noun pair extraction unit that extracts noun pairs included in the same sentence in the document data;
A third noun extraction unit for extracting a third noun from the sentence including the noun pair extracted by the noun pair extraction unit;
A seed pattern storage unit that stores a pattern of a sentence structure in which a seed word pair, which is a noun pair having a known relationship, appears as a seed pattern associated with the relationship;
Based on the noun pair extracted by the noun pair extraction unit, the third noun extracted by the third noun extraction unit, and the seed pattern stored in the seed pattern storage unit, the noun pair is A score calculator that calculates a score indicating the degree of having the relationship associated with the seed pattern;
Based on the score calculated by the score calculation unit, a specific related noun pair extraction unit that extracts the noun pair estimated to have the relationship;
A semantic relationship extraction apparatus comprising: The seed pattern is represented as a path of a phrase between words belonging to the seed word pair in a tree-structured data representing a dependency relation between phrases.
The semantic relationship extraction apparatus according to claim 1, wherein: The score calculation unit calculates a first score that is a value representing a degree that a class pair of nouns corresponding to the noun pair appears in the relationship, and calculates the score based on the first score.
The semantic relationship extraction device according to claim 1, wherein the semantic relationship extraction device is a device. The score calculation unit calculates a second score that is a value indicating a degree of ease of appearing in the seed pattern of a pattern between one noun included in the noun pair and the third noun. Calculating and calculating the score based on the second score;
The semantic relationship extraction device according to any one of claims 1 to 3, wherein The score calculation unit calculates a third score, which is a value representing the likelihood of an upper-lower relationship or a parallel relationship, of a pattern between one noun included in the noun pair and the third noun, Calculating the score based on 3 scores;
The semantic relationship extraction device according to any one of claims 1 to 4, wherein: Based on the document data, a noun pair extraction unit that extracts noun pairs included in the same sentence in the document data;
A third noun extraction unit for extracting a third noun from the sentence including the noun pair extracted by the noun pair extraction unit;
A seed pattern storage unit that stores a pattern of a sentence structure in which a seed word pair, which is a noun pair having a known relationship, appears as a seed pattern associated with the relationship;
Based on the noun pair extracted by the noun pair extraction unit, the third noun extracted by the third noun extraction unit, and the seed pattern stored in the seed pattern storage unit, the noun pair is A score calculator that calculates a score indicating the degree of having the relationship associated with the seed pattern;
Based on the score calculated by the score calculation unit, a specific related noun pair extraction unit that extracts the noun pair estimated to have the relationship;
As a program to make the computer function as.

Images