JP7110554B2 - Ontology generation device, ontology generation program and ontology generation method - Google Patents

Description

The present invention relates to an ontology generation device, an ontology generation program, and an ontology generation method.

In order to construct a specific ontology of specialized knowledge, efforts are being made to extract word pairs that have relationships that can express specialized knowledge from a group of documents. Regarding the extraction of related word pairs from a group of documents, there is a method of assigning a small number of word pairs as clues (seeds) and extracting them by pattern matching using the extraction patterns of parts of speech and syntactic structures.

However, in the case of the above method, word pairs that are not in the intended relationship may be extracted. Therefore, we classify whether the extraction pattern can extract an accurate word pair, and use a supervised classifier for judging the appropriateness of the extraction pattern as learning data to judge the appropriateness of the new extraction pattern, and do not use inappropriate extraction patterns. There is a technique for suppressing the extraction of unintended word/phrase pairs by means of the above (see Patent Literature 1 below).

Patent No. 5430989

However, since a supervised classifier is constructed using the extracted pattern appropriateness classification results, there is a problem that subsequent determination accuracy is reduced due to the influence of extracted pattern appropriateness determination errors.

Accordingly, it is an object of one aspect of the present invention to improve the accuracy of extracting word/phrase pairs having an intended relationship without being affected by the results of intermediate processing.

In one example of the aspect, an ontology generation device that generates an ontology dictionary, which is a semantic database, is a word pair that is a pair of words that are related in meaning, and is composed of a front word and a back word. The rear word/phrase is obtained by: a storage unit that stores the word/phrase pair that is arranged at a position behind the front word/phrase in the word meaning sentence containing the word/phrase pair as an item of the ontology dictionary; and the word/phrase pair that is input from the outside. An acquisition unit that acquires a certain seed word pair and stores it in the storage unit as the item, a division unit that divides the acquired document data into a plurality of word sense sentences, and from the plurality of word sense sentences obtained by the division, a word sense sentence extracting unit for extracting a word sense sentence containing the word pair stored as the item in the storage unit; and the word sense sentence pair included in the word sense sentence extracted by the word sense sentence extracting unit. using a predetermined calculation formula to calculate a score representing the degree of relevance in meaning between each word and phrase that constitutes the seed word and phrase pair, and the score calculated for the word meaning sentence and a predetermined threshold value an evaluation unit for determining whether or not the relevance is low by comparing the magnitudes, and excluding the word meaning sentence from the result of the extraction when the relevance is determined to be low; and the extraction without being excluded by the evaluation unit. Generating a relationship extraction pattern by replacing each word and phrase constituting the word and phrase pair contained in the word meaning sentence remaining as a result of the above with a variable having attribute information for each word and phrase and a word/phrase in which a word/phrase other than the variable among the plurality of word meaning sentences obtained by the division matches the relation extraction pattern and is arranged at the position of the variable in the relation extraction pattern. a word/phrase pair extracting unit for extracting a pair of words/phrases placed at the position from a word meaning sentence whose attribute information matches the attribute information of the variable, and storing the word/phrase pair in the storage unit as the item. and the predetermined calculation formula is, in the word-meaning sentence extracted by the word-meaning sentence extracting unit, the theme distribution and the seed word-phrase pair for the preceding word/phrase of the word/phrase pair included in the word-sense sentence. the subject distribution for the back term of the word pair contained in the semantic sentence and the subject for the back term of the seed word pair; It is a formula for calculating the weighted average value of the vector distance from the distribution as the score, and the subject distribution for the words in the word meaning sentence is the word is represented by a vector obtained by multiplying a vector representing the word topic distribution for a word with a vector representing the document topic distribution for the semantic sentence, and It is a probability distribution that expresses the probability that the word or phrase is included for each subject. It is a probability distribution that represents

It is possible to improve the accuracy of extracting word/phrase pairs having an intended relationship without being affected by the results of intermediate processing.

FIG. 4 is a diagram for explaining an automatic extraction framework for extraction patterns and word/phrase pairs; FIG. 4 is a diagram for explaining word pair extraction by a supervised classifier that determines pattern suitability; FIG. 10 is a diagram for explaining that an error in determination of pattern appropriateness affects subsequent determinations and lowers the determination accuracy; 2 is a diagram showing the functional configuration of the ontology generation device of the first embodiment; FIG. FIG. 4 is a diagram showing an example of words and phrases having a high-level and low-level relationship that are the basis for constructing an ontology; FIG. 3 is a diagram showing an example of a constructed tree-structured ontology; FIG. 10 is a diagram showing the result of processing the text by the parser; FIG. 10 is a diagram showing the results of morphological analysis and word/phrase extraction on text; FIG. 10 is a diagram showing a word/phrase topic distribution and a document topic distribution of learning results; FIG. 10 is a diagram for explaining how to create a combination of positive and negative examples in automatic determination of weight; FIG. 10 is a diagram showing an example of weights and semantic relevance scores in automatic determination of weights; FIG. 11 is a flow chart showing an example of a relationship extraction pattern extraction processing flow in the ontology generation devices of the first, second, and fourth embodiments; FIG. 4 is a subroutine showing an example of the processing flow of subject distribution learning by word/phrase document in the ontology generation device of the first embodiment; 4 is a subroutine showing an example of a pattern information extraction processing flow in the ontology generation device of the first embodiment; FIG. 10 is a diagram for explaining the concept of determination of deletion of a candidate for generating a relationship extraction pattern; FIG. 2 is a diagram showing an example of a hardware configuration of an ontology generation device according to each embodiment; FIG. FIG. 4 is a diagram for explaining the effect of the ontology generation device of each embodiment; FIG. 4 is a diagram for explaining that an intended word/phrase pair can be extracted by the ontology generation device of each embodiment; FIG. 4 is a diagram showing the result of word pairs extracted by a conventional ontology generation method; FIG. 10 is a diagram showing the results of word/phrase pairs extracted by the ontology generation device of each embodiment; It is a figure which shows the extraction result of both word-phrase pairs. FIG. 11 is a diagram showing the functional configuration of ontology generation devices of the second and fourth embodiments; FIG. 11 is a diagram showing word/phrase topic distribution and document topic distribution in the second embodiment; FIG. 11 is a subroutine showing an example of the processing flow of subject distribution learning by word/phrase document in the ontology generation device of the second and third embodiments; FIG. FIG. 11 is a subroutine showing an example of the pattern information extraction processing flow in the ontology generation devices of the second and third embodiments; FIG. FIG. 13 is a diagram showing the functional configuration of an ontology generation device according to the third embodiment; FIG. FIG. 14 is a flowchart showing an example of a relationship extraction pattern extraction processing flow in the ontology generation device of the third embodiment; FIG. FIG. 10 is a diagram showing the text after synonym replacement is performed on the acquired text; FIG. 10 is a diagram showing a learning result when learning is performed without performing synonym replacement on an acquired text and a learning result when learning is performed on an acquired text with synonym replacement; FIG. 10 is a diagram showing word topic distributions and document topic distributions including combinations in the third embodiment; FIG. 14 is a subroutine showing an example of the processing flow of subject distribution learning by word/phrase document in the ontology generation device of the fourth embodiment; FIG. FIG. 13 is a subroutine showing an example of the pattern information extraction processing flow in the ontology generation device of the fourth embodiment; FIG. FIG. 10 is a diagram showing an example of a search system using ontologies generated by the ontology generation devices of the first to fourth embodiments; FIG. FIG. 4 is a diagram for explaining automatic extension by a keyword extension device;

EMBODIMENT OF THE INVENTION Hereinafter, it demonstrates, referring drawings for the form for implementing this invention.
First, a relationship extraction pattern (hereinafter also referred to as an extraction pattern) and an automatic extraction framework for word pairs will be described with reference to FIG. The automatic extraction framework analyzes the document group 11 on a sentence-by-sentence basis and extracts pairs of words having a binary relationship as word pairs. That is, a relation extraction pattern is acquired from a sentence (text) analyzed using a seed word/phrase pair of a predetermined seed (clue), and recorded in a relation extraction pattern DB (Data Base) 12 . A word/phrase pair is obtained from the text using the obtained relationship extraction pattern and recorded in the word/phrase pair DB 13 . An ontology is generated based on the information of word pairs acquired sequentially in this way and recorded in the word pair DB 13 , provided to the user 14 , and further fed back by the user 14 .

A phrase pair is a pair of phrases, and can be expressed as (A, B), for example. A and B are words and phrases respectively. Here, the word/phrase includes not only the word/phrase A extracted by morphological analysis, but also the word/phrase A, such as A Corporation, which is a combination of the word A and the word corporation.

A relationship extraction pattern refers to all or part of a text in which a seed word/phrase pair appears, in which the occurrence locations of the seed word/phrase pair are converted into variables. For example, consider the case where the seed phrase pair is (X, Y), where X is the manufacturer name and Y is the smartphone model name. At this time, if the text in which the seed word/phrase pair (X, Y) appears is "X's latest product Y has been announced", the relationship extraction pattern is The name is converted into a variable, that is, replaced with “(manufacturer name)’s latest product (smartphone model name) has been announced”. Note that “smartphone” is an abbreviation for smartphone, and the same applies hereinafter.

An ontology is a type of dictionary that systematically organizes the concepts of words.
In the above automatic extraction framework, word pairs with unintended relationships may be extracted. For example, the relationship extraction pattern is "(manufacturer name)'s latest product (smartphone model name) has been announced", and the word/phrase pair obtained from other text based on the relationship extraction pattern is (XX, YY). Suppose there is In this case, there is no problem if XX is the manufacturer name and YY is the smartphone model name. However, while XX is the name of the manufacturer, YY may be the name of clothes instead of the model name of the smartphone. In this case, a word/phrase pair having an unintended relationship is extracted.

The above-mentioned Patent Document 1 discloses a technique for suppressing the extraction of unintended word/phrase pairs. In the technique disclosed in Patent Document 1, a supervised classifier is constructed that classifies whether or not a relationship extraction pattern can extract an accurate word pair, and determines pattern appropriateness as learning data. As a result, the suitability of the new relationship extraction pattern is determined, and extraction of unintended relationship word pairs is suppressed by not using inappropriate relationship extraction patterns.

Specifically, as shown in FIG. 2, pattern extraction (relationship extraction pattern extraction) is performed from the text based on the seed phrase pair (sedan, sky XX), and the first extraction pattern is, for example, "X's new car Y was announced" is extracted. Here, it is assumed that (vehicle type, product name) is extracted as the desired relationship.

Appropriateness of the extracted pattern is determined by a supervised classifier that uses the past pattern determination results as learning data. In this example, the relationship extraction pattern "Y which is one of X's representative cars" and "X's work Y" are used as learning data.

As a result of the determination, the extracted pattern "X's new car, Y has been announced" is determined to be a suitable pattern, and this pattern is used to extract a new word/phrase pair (sedan, foogu). When new word pairs are extracted, the second extraction pattern is the same as the first extraction pattern: "X's new model Y announced", "Y's new model announced at X", "Y is a new product of the X series There is" is extracted, and the suitability of the pattern is determined.

As a result of the determination, it was determined that the extraction pattern "Y is a new product of the X series" is not appropriate. ), (Arlows, M03) are extracted.

However, the extracted word/phrase pairs may include word/phrase pairs that do not have the desired relationship. This is because when a supervised classifier is constructed using the classification results of pattern appropriateness so far, if an error occurs in determining pattern appropriateness, subsequent determinations are affected, and the determination accuracy decreases. This is because

A specific description will be given with reference to FIG. Inappropriate word/phrase pair 31 is extracted from pattern 30 due to an error in determination of appropriateness of the pattern, and inappropriate word/phrase pair 33 is again extracted from pattern 32 extracted by the extracted word/phrase pair 31 . In this way, once an error in determining the suitability of a pattern occurs, it affects subsequent determinations and lowers determination accuracy. The white frames (white areas) indicate all the word/phrase pairs and patterns to be extracted.

The ontology generation device of each embodiment solves the above problem and suppresses the extraction of unintended relational word pairs.

The ontology generation device of each embodiment will be described below.
<First embodiment>
The configuration of the ontology generation device of the first embodiment will be described with reference to FIG. The ontology generation device 40 includes a seed phrase input unit 401, a text acquisition unit 402, a syntax analysis unit 403, an extraction processing control unit 404, an extraction pattern generation candidate extraction unit 405, a pattern extraction unit 406, a phrase extraction unit 407, and an ontology generation unit 408. , and an output unit 409 . The ontology generation device 40 further includes a syntactic structure storage unit 410, a word/phrase storage unit 411, an extraction pattern storage unit 412, an extraction pattern generation candidate storage unit 413, a phrase document-by-word topic distribution learning unit 414, and a phrase/document-by-word topic distribution storage unit. 415 and an extraction pattern generation candidate evaluation unit 416 .

The seed word/phrase input unit 401 accepts in advance a word/phrase pair (seed word/phrase pair) that is specified by the user and has a desired relationship, and uses the input word/phrase pair as a relationship extraction pattern and a seed for the framework for automatically extracting the word/phrase pair. do. For example, when the user inputs a word/phrase pair (BBB, Arrows) as a seed word/phrase pair for a relationship example (smartphone manufacturer, smartphone series) desired by the user, the seed word/phrase input unit 401 inputs the word/phrase pair (BBB, Arrows). Here, BBB is, for example, the name of a smartphone manufacturer, and Arrows is a series name indicating a series of smartphones.

A text acquisition unit 402 acquires a group of documents (document data) from a database that records the group of documents, etc., divides the acquired group of documents into a plurality of texts (also called semantic sentences), and extracts the texts. Here, the document group is a collection of a plurality of documents, for example, documents in text format or HTML (Hyper Text Markup Language) format.

The syntactic analysis unit 403 performs morphological analysis and syntactic analysis on the text extracted by the text acquisition unit 402 , or semantic structure analysis of the text, and stores the results in the syntactic structure storage unit 410 . For example, when the extracted text is "This year's best smartphone award was awarded to Arrows M03, the latest product of BBB", the syntactic analysis unit 403 performs morphological analysis and syntactic analysis, and the result is "now Arrows M03, the latest product of BBB, won the best smartphone award of the year." is stored in the syntax structure storage unit 410.

The extraction processing control unit 404 controls the execution of the framework for automatically extracting relationship extraction patterns and word pairs, and stops the automatic extraction processing of related word pairs when new related word pairs are not extracted from the text, The word/phrase pairs stored in the word/phrase storage unit 411 are transmitted to the ontology generation unit 408 .

The extraction pattern generation candidate extraction unit 405 extracts the locations (sentences) where the word/phrase pairs stored in the word/phrase storage unit 411 appear from the syntactic structure stored in the syntactic structure storage unit 410, Send to For example, if the word/phrase pair stored in the word/phrase storage unit 411 is (CCC, zenfo3), the first appearance (sentence) "CCC's latest product, zenfo3." 20% increase in sales forecast for zenfo3" is extracted from the annual shipping plan. Here, CCC is, for example, the name of a smartphone manufacturer, and zenfo3 is the model name of the smartphone. The extraction pattern generation candidate extraction unit 405 transmits the extracted sentence to the extraction pattern generation candidate storage unit 413 to store it.

The pattern extraction unit 406 generates relationship extraction patterns for extracting relationship extraction patterns and related word/phrase pairs in automatic extraction frames of word/phrase pairs from the sentences stored in the extraction pattern generation candidate storage unit 413 . For example, if the sentence is "zenfo3, the latest product of CCC." Generate patterns. The relationship extraction pattern generated in this case is "(manufacturer name) latest product (smartphone model name)".

The word/phrase extraction unit 407 extracts words/phrases corresponding to variableization using the relationship extraction patterns recorded in the extraction pattern generation candidate storage unit 413, that is, extracts the words/phrases of the divided plural word meaning sentences from the lexical concepts. Words (that is, word/phrase pairs) forming a set of concept pairs are specified and recorded in the word/phrase storage unit 411 as related word/phrase pairs.

For example, consider the case where the relationship extraction pattern is "(manufacturer name) latest product (smartphone model name)". The word/phrase extraction unit 407 uses the relationship extraction pattern to extract the variable part from "This year's best smartphone award was awarded to Arrows M03, the latest product of BBB", stored in the syntactic structure storage unit 410. Extract the corresponding word pairs (BBB, Arrows M03). The word/phrase extraction unit 407 transmits the extracted word/phrase pair to the word/phrase storage unit 411 for storage.

The ontology generation unit 408 organizes and connects related word/phrase pairs recorded in the word/phrase storage unit 411 to construct an ontology. For example, the ontology generation unit 408 constructs a tree-structured ontology shown in FIG. 5B based on the upper-level and lower-level relationships shown in FIG. 5A. For example, if the upper level is a stadium, the lower level is an individual stadium name that is one of the stadiums. In FIG. 5B, the upper ones are indicated by solid line circles, and the lower ones are indicated by dashed line circles.

The output unit 409 converts the ontology construction result generated by the ontology generation unit 408 into a predetermined format, displays it on a display device (not shown), or transmits the content after conversion to another device.

The syntax structure storage unit 410 stores the results processed by the syntax analysis unit 403 . For example, as described above, the structure "This year's best smartphone award was won by BBB's latest product Arrows M03" is recorded (see FIG. 6). At this time, the arrows shown in FIG. 6 indicate the modification destination of the clause.

The word/phrase storage unit 411 stores seed word/phrase pairs received by the seed word/phrase input unit 401 and word/phrase pairs extracted by the word/phrase extraction unit 407 in the automatic word/phrase pair extraction process.

The extracted pattern storage unit 412 stores the relationship extraction patterns extracted by the pattern extraction unit 406 . For example, it stores the above-mentioned "(manufacturer name) latest product (smartphone model name)".

The extraction pattern generation candidate storage unit 413 stores the occurrence locations (sentences) of the word/phrase pairs extracted by the extraction pattern generation candidate extraction unit 405 . As described above, for example, "CCC's latest product, zenfo3."

The word/phrase document-by-document topic distribution learning unit 414 performs morphological analysis on the text acquired by the text acquisition unit 402, learns the document-by-document topic distribution (Topic Model) of the analysis result word/phrase set, and obtains the learning result (vector). It is stored in the phrase-document-by-word-theme-distribution storage unit 415 . In other words, there is a word/phrase theme distribution in which words and phrases are extracted by performing morphological analysis on a word meaning sentence (text) and a combination word/phrase combining the extracted words and phrases refers to the topic, and the extracted words/phrases exist. Learn the document topic distribution in which the semantic sentence refers to the topic.

Topic Model is a clustering (unsupervised classification) that mainly assumes document data, ignores the order of words and considers only the frequency of words in the document, and models the topic of words without giving supervised data. It is something that becomes. That is, the Topic Model is a model that assumes that a document is probabilistically generated from a plurality of potential themes, and that each word in the document appears according to the probability distribution of the topic. Topic Model can analyze the similarity between topics and their meaning by assuming the appearance frequency distribution of words for each topic.

Phrases do not appear independently, they have potential subjects, and phrases with the same subject tend to appear in the same document. For example, there is a sentence that reads, ``Mt. Here, it is assumed that ○○yama is a famous yokozuna and △△maru is a famous ozeki. In this case, there is no word "sumo" in the sentence, but looking at this sentence, it can be understood that the sentence is related to "sumo". That is, this sentence has one subject, "sumo."

For example, as shown in FIG. 7A, there are N acquired texts, and morphological analysis and phrase extraction are performed on these texts to obtain extracted phrases for each text. Assuming that the number of themes is K, the word/phrase document topic distribution learning unit 414 uses the word/phrase topic distribution (vector V _w ) and the document topic distribution (vector V _D ) shown in FIG. Stored in the document-by-document topic distribution storage unit 415 . The phrase-subject distribution is the probability distribution that mentions the subject of the phrase itself, and the document-subject distribution is the probability distribution that mentions the subject of the document.

The numerical values in the word/phrase theme distribution indicate the probability distribution of occurrence of each word/phrase in each subject. The sum of these numbers equals 1. Also, the numerical values in the document subject distribution indicate the probability distribution of appearance of each document in each subject. there is The sum of these numbers equals 1.

Although the number of themes is K here, it is not limited to K, and other numbers are possible. The word aaa shown in FIG. 7A is, for example, a product name, and the words AAA and BBB are, for example, manufacturer names.

Note that the learning method is a conventional technique, and although detailed description is omitted here, the following documents can be referred to. For example, Biei, David M. (2012). Probabilistic topic models, Commun. ACM, 55(4), 77-84. Also, Liu, Jun S, (1994). The Collapsed Gibbs Sampler in Bayesian Computations with Association, 89(427), 958-966. Also, Mikolov, Tomas; et al, Efficient Estimation of Word Representations in Vector Space.

The phrase document-by-word topic distribution storage unit 415 records the learning result (vector) by the phrase-by-word document topic distribution learning unit 414 . For example, record the phrase topic distribution and the document topic distribution shown in FIG. 7A. This distribution is a probability distribution.

The extraction pattern generation candidate evaluation unit 416 compares the occurrence location of the word/phrase pair recorded in the extraction pattern generation candidate storage unit 413 with the seed word/phrase pair recorded in the word/phrase storage unit 411, and compares the occurrence location with the word/phrase pair. Determine (evaluate) whether the semantic relevance score with the seed phrase pair exceeds a threshold. If the semantic relevance score does not exceed the threshold, the occurrence location (sentence) is deleted from the extraction pattern generation candidate storage unit 413 . That is, among the plurality of word-sense sentences, a first word (seed word-phrase pair) forming a concept pair corresponding to the first word-sense sentence and a second word forming a concept pair corresponding to the second word-sense sentence (word/phrase pairs at appearance locations) are similar, and if it is determined that they are not similar, the words (word/phrase pairs) that make up the concept pair are excluded from being registered in the ontology dictionary. It should be noted that "similar" means that there is a relationship between word/phrase pairs.

The semantic relevance score is a value that evaluates the extent to which the respective meanings of contrasting word pairs are related.

For example, it is assumed that there are two occurrence locations (sentences) of the word/phrase pair (sedan, fugu) extracted in the case of the seed word/phrase pair (sedan, sky XX). One is "Fugu, which is a kind of sedan, was announced", and the other is "Fugu, a new sedan, was announced". It is assumed that the first sentence relates to (car model, product name) and the second sentence relates to (place name, music type).

In this case, the extraction pattern generation candidate evaluation unit 416 records the semantic relevance scores of the seed word/phrase pair in the document of the appearance location and the extracted word/phrase pair in the document of the appearance location in the word/phrase-by-document subject distribution storage unit 415. Calculate based on the learned results. In this example, since the relevance between (sedan, fugu) and (sedan, sky○○) in the second sentence is low, the calculated semantic relevance score does not exceed the threshold, and the second occurrence ( sentence) is deleted from the extraction pattern generation candidate storage unit 413 .

That is, the extracted pattern generation candidate evaluation unit 416 calculates a semantic relevance score indicating the degree of similarity (relevance) between the first word (seed word pair) and the second word (word/phrase pair at the occurrence location) using a predetermined calculation formula. is calculated by the following (1). Then, when the calculated semantic relevance score is equal to or less than a predetermined threshold, it is determined that the first word and the second word are not similar (relevance is low).

Calculation of semantic relevance scores for seed phrase pairs in the document of occurrence and extracted phrase pairs in the document of occurrence will now be described.

In the following, we describe the computation of the semantic relevance score for two phrase pairs i, j in document d. Phrase pair i is the seed phrase pair and phrase pair j is the extracted phrase pair. In document d, a semantic relevance score SRScore _{i, j, d} is calculated by the following formula (1) for a phrase pair j=(j ^A , j ^B ) and a phrase pair i=(i ^A , i ^B ) . j ^A is phrase A of phrase pair j (previous phrase A), j ^B is phrase B of phrase pair j (back phrase B), i ^A is phrase A of phrase pair i (previous phrase A), i ^B is phrase pair Term B of i (rear term B) is shown respectively. For example, if the word/phrase pair j is (BBB, Zinrei), then word/phrase A is BBB and word/phrase B is Zinrei. A forward phrase refers to a phrase that appears earlier in a phrase pair, and a backward phrase refers to a phrase that appears later in a phrase pair.

はシード語句対ｉの前方語句Ａの文書ｄにおける主題分布を表すベクトルである。これは、例えば主題数を３個（次元Ｋ＝３）とした場合の語句Ａの主題分布Ａ_ｋと文書ｄの主題分布Ｄ_ｋの掛け算によって求められる。例えば、語句Ａの主題分布Ａ_ｋが｛０．３、０．１、０．６｝、文書ｄの主題分布Ｄ_ｋが｛０．４、０．４、０．２｝とした場合、語句Ａの文書ｄにおける主題分布は｛０．３×０．４、０．１×０．４、０．６×０．２｝＝｛０．１２、０．０４、０．１２｝となる。 here,

is a vector representing the subject distribution in document d of the predecessor phrase A of seed phrase pair i. This is obtained, for example, by multiplying the subject distribution Ak of the word/phrase A and the subject distribution _Dk of the document _d when the number of themes is three (dimension K=3). For example, if the subject distribution A _k of word A is {0.3, 0.1, 0.6} and the subject distribution D _k of document d is {0.4, 0.4, 0.2}, the word The subject distribution in document d of A is {0.3×0.4, 0.1×0.4, 0.6×0.2}={0.12, 0.04, 0.12}.

The subject distribution A _k {0.3, 0.1, 0.6} of the word A indicates the probability distribution of the appearance of the word A in each of the three themes. The probability of doing so is 0.3. The sum of these numbers equals 1. Note that the subject distribution D _k {0.4, 0.4, 0.2} of document d can also be considered in the same way.

A specific example of the topical distribution of phrases in documents is given. For example, document 1 is "Tomorrow is the day of the machine learning hackathon. My preparation is perfect." When morphological analysis is performed on document 1 and phrases are extracted, it becomes "tomorrow, machine learning, hackathon, day, preparation, perfect". When morphological analysis is performed on document 2 and phrases are extracted, "machine learning, hackathon, staying up late, scary" is obtained.

Here, it is assumed that the number of themes is K=3, and the topic distribution of each word and the topic distribution of each document are known. In this case, based on the topic distribution of each word and the topic distribution of each document, the topic distribution of each word in document 1 is as follows. That is, "tomorrow {0.1, 0.1, 0.8}, machine learning {0.2, 0.3, 0.5}, hackathon {0.4, 0.3, 0.3}, day { 0.2, 0.2, 0.6}, preparation {0.3, 0.5, 0.2}, perfect {0.2, 0.6, 0.2}". Also, the subject distribution of each word in document 2 is "machine learning {0.4, 0.3, 0.3}, hackathon {0.9, 0.05, 0.05}, late night {0.5, 0.3, 0.2}, scary {0.4, 0.3, 0.3}". In this way, since the themes of documents are different, the topic distribution of words and phrases is also different for each document.

は、シード語句対ｉの前方語句Ａと、抽出された語句対ｊの前方語句Ａの文書ｄにおけるベクトル距離を示しており、例えば

のコサイン類似度で示される。 again,

denotes the vector distance in document d between the prefix A of seed phrase pair i and the prefix A of extracted phrase pair j, e.g.

is represented by the cosine similarity of .

In addition, α _l indicates the weighting factor of the front phrase A and the rear phrase B, and may be set manually in advance, or may be set by automatic judgment. In the first embodiment, there are two weights, but in other embodiments described later, there are three weights (forward term A, backward term B, combination), and the like.

Here, setting by automatic determination of weights will be described. First, a correct word/phrase pair (positive example word/phrase pair list) that matches the user's intention and an absolutely incorrect word/phrase pair (negative example word/phrase pair list) are prepared. For the document group to be extracted, a combination of all words in the positive example word pair list and the negative example word pair list is created, and l=A, B, C in the above equation (1), and the denominator is 3. Calculate the semantic relevance score as Adjust _αl and repeat the score calculation until the semantic relevance score distinguishes between positive and negative examples. Set the optimized α _l as the weight used in actual processing. Note that the above process may be performed by the extraction pattern generation candidate evaluation unit 416, for example, or may be performed by another component.

For example, as shown in FIG. 8A, positive examples (positive example word/phrase pair list) and negative examples (negative example word/phrase pair list) are provided manually. In the positive example here, the front term is the smartphone manufacturer and the back term is the smartphone model name. The words BBB, DDD, and EEE are smartphone manufacturers, and the words Arrows, G Note, and Xpier are smartphone model names. On the other hand, the negative examples are not smartphone manufacturers and smartphone model names, but the antecedent phrase AAA is automobile manufacturers, the antecedent phrase FFF is a general electric manufacturer, the suffix term Prio is a car product name, and the suffix Biri is a TV model name.

Create a combination by mixing the positive and negative examples. Combinations in this case include combinations of positive cases, combinations of positive cases and negative cases, and combinations of negative cases. For example, a combination of positive cases has an ID of 1-2, a combination of a positive case and a negative case has an ID of 1-R, and a combination of negative cases has an ID of RS, for example.

Using the created combinations, a semantic relevance score is calculated while adjusting the weights (weight 1, weight 2, weight 3, . . . ) as shown in FIG. 8B. The results are shown in FIG. 8B. As an example of weights to be adjusted, weight 1 is 0.3, 0.1, and 0.6. A threshold for determining the weight is set to 0.5, for example. From the semantic relevance score results shown in FIG. 8B, only when the weight is 2, the semantic relevance score between positive examples is as high as 0.8 or 0.9, and the combination of positive and negative examples is 0.4, 0.3. , 0.1, which is low without exceeding the threshold. Based on this result, the weighting factor α _l is set to the weight 2 value. A weight that gives a high score for a combination of positive examples and a low score for a combination of positive and negative examples is suitable as a weight for extracting word/phrase pairs that match the user's intention.

Next, a flow of relation extraction pattern extraction processing in the ontology generation device of the first embodiment will be described with reference to FIG.

The ontology generation device 40 receives and acquires seed word/phrase pairs (step S901). Also, the ontology generation device 40 acquires a group of documents from a database or the like that records the group of documents, and extracts and acquires text from the documents of the acquired group of documents (step S902).

The ontology generation device 40 performs morphological analysis and syntactic analysis on the acquired text, or semantic structure analysis of the text, that is, syntactic structure analysis (step S903). The ontology generation device 40 performs morphological analysis on the acquired text, and learns the document-by-document theme distribution (Topic Model) of the analysis result word/phrase set, that is, word/phrase-document-by-word topic distribution learning (step S904). The details of the processing flow of topic distribution learning for each phrase document will be described later.

The ontology generation device 40 compares the occurrence location of the word/phrase pair recorded in the extraction pattern generation candidate storage unit 413 with the seed word/phrase pair recorded in the word/phrase storage unit 411, and compares the occurrence location word/phrase pair with the seed word/phrase pair. It is determined (evaluated) whether or not the semantic relevance score of exceeds the threshold. If the semantic relevance score does not exceed the threshold, the occurrence location (sentence) is deleted from the extraction pattern generation candidate storage unit 413 . That is, the ontology generation device 40 performs pattern information extraction processing (step S905). The details of the flow of pattern information extraction processing will be described later.

The ontology generating device 40 uses the generated relationship extraction pattern to extract the word/phrase corresponding to the variableization, that is, performs word/phrase information extraction processing (step S906).

Here, the processing flow of the topic distribution learning by word/phrase document will be described with reference to FIG. The ontology generation device 40 identifies words in the acquired text (step S1001). That is, words are extracted from the text and the words to be learned are specified. The ontology generation device 40 learns the document-by-document theme distribution of the identified phrase set, that is, performs model learning (step S1002).

Also, the pattern information extraction processing flow will be described with reference to FIG. The ontology generation device 40 first extracts the occurrence location D of the word/phrase pair j recorded in the extraction pattern generation candidate storage unit 413 (step S1101). For occurrence dj of phrase pair j={ ^jA , ^jB }, and for seed phrase pair i= ^{ iA,iB}, phrase { ^iA , ^jA } and phrase ^{{ iB} , ^jB } is calculated (step S1102). In other words, the score is calculated by comparing the preceding words and phrases and the following words and phrases of the first word (seed word and phrase pair) and the second word (word and phrase pair of appearance location).

The ontology generation device 40 determines whether or not the calculated semantic relevance score exceeds the threshold (step S1103). Delete (exclude) from the pattern generation candidates (step S1104), and perform the same processing for the next word/phrase pair. On the other hand, if it exceeds (Yes in step S1103), the ontology generation device 40 performs similar processing for the next seed word/phrase pair.

When the processing for all the word/phrase pairs is completed, the ontology generation device 40 generates a relationship extraction pattern of the word/phrase pairs from the occurrence locations (filtered) of the word/phrase pair j (step S1105).

In the first embodiment, forward words and phrases and backward words and phrases are evaluated. In the second embodiment, which will be described later, words and phrases that are combined as one word and phrase are further evaluated. . The concept of deletion determination of the relationship extraction pattern generation candidate when these three evaluation patterns are used will be described with reference to FIG. 12 .

For example, a seed phrase pair (A, B) that is anterior phrase A, an posterior phrase B, a combination phrase (AB) and a phrase pair (C, D) that is an anterior phrase C, an posterior phrase D, a combination phrase (CD). Consider the case. As shown in FIG. 12, in Document 2, front words and phrases (A, C), back words and phrases (B, D), and combination words and phrases (AB, CD) tend to represent the same subject. That is, the front words (A, C) represent the topic T1, the rear words (B, D) represent the topic T2, and the combination words (AB, CD) represent the topic T3.

On the other hand, in document 1, the theme distribution tendencies of the combination words (AB, CD) are different. That is, the words (AB) represent the subject T2, the words (CD) represent the subject T1, and the subjects of the respective words are different. Since the words and phrases are the same in the document but have different themes, the locations of appearance in Document 1 are deleted from the relationship extraction pattern generation candidates, and a relationship extraction pattern is generated only from the locations of appearance in Document 2.

Next, the hardware configuration of the ontology generation device of the first embodiment will be explained using FIG. Since the same applies to ontology generation devices of other embodiments, which will be described later, a description of the hardware configuration of the ontology generation devices of other embodiments will be omitted.

The hardware of the ontology generation device 40 includes a CPU (Central Processor Unit) 130, a HDD (Hard Disk Drive) 131, a RAM (Random Access Memory) 132, a ROM (Read Only Memory) 133, an input device 134, a communication interface 135, a bus 136. The CPU 130, HDD 131, RAM 132, ROM 133, input device 134, and communication interface 135 are connected to each other via a bus 136, for example.

The CPU 130 reads a program for performing various processing (for example, pattern information extraction processing) stored in the HDD 131 or the like via the bus 136, temporarily stores the read program in the RAM 132, and executes the program according to the program. Perform various processing. The CPU 130 mainly includes the text acquisition unit 402, the syntax analysis unit 403, the extraction processing control unit 404, the extraction pattern generation candidate extraction unit 405, the pattern extraction unit 406, the phrase extraction unit 407, the ontology generation unit 408, the output unit 409, the phrase document It functions as a separate subject distribution learning unit 414 and an extraction pattern generation candidate evaluation unit 416 .

The HDD 131 stores application programs for performing various processes of the ontology generation device 40, data necessary for processing of the ontology generation device 40, and the like. The HDD 131 mainly functions as various storage units described above.

The RAM 132 is a volatile memory that temporarily stores a part of an OS (Operating System) program and application programs to be executed by the CPU 130 . Also, the RAM 132 stores various data necessary for processing by the CPU 130 .

The ROM 133 is a non-volatile memory and stores programs such as a boot program and BIOS (Basic Input/Output System).

The input device 134 is for inputting information, and is, for example, a keyboard.
The communication interface 135 transmits and receives data to and from the outside (such as an external device that requests an ontology construction result, etc.) via a network.

A bus 136 is a path that mediates transfer of control signals, data signals, and the like between devices.
According to the ontology generation device of the first embodiment, the relevance to the seed word/phrase pair for each word/phrase pair appearance is evaluated using the document-by-document thematic distribution of words/phrases, and inappropriate appearances are deleted. Therefore, extraction of inappropriate relationship extraction patterns can be suppressed. That is, according to the ontology generation device of the first embodiment, as shown in the dot-marked area in FIG. Extraction accuracy of word pairs is improved.

Also, as shown in FIG. 15, by excluding inappropriate relationship extraction patterns, intended word pairs can be extracted, that is, extraction of inappropriate word/phrase pairs can be avoided.

Here, it was verified that word pairs that do not have the relationship expressed by seed word pairs can be deleted by using the document-by-document theme distribution of words. As shown in FIGS. 16A and 16B, the target document group here is 706 press release documents, and the seed phrase pairs are {product, service type} and {product, service name}. (smartphone, ARLOWS Z IS W11P). Also, here, the result of word pair extraction by the conventional ontology generation technology, which is not the technology of Patent Document 1, is compared with the result of word pair extraction by the ontology generation device of the embodiment.

FIG. 16A shows the results of word pairs extracted by the conventional ontology generation method, and FIG. 16B shows the results of word pairs extracted by the ontology generation device of the embodiment. It should be noted that the ◯ mark described in the table showing the extraction result in FIG. 16B indicates that the intended word/phrase pair was extracted. FIG. 16C shows a table summarizing both results.

The number of extractions shown in FIG. 16C is the number of extracted word/phrase pairs, and the number of correct answers is the number of extracted word/phrase pairs having the intended relationship. The recall rate indicates how many correct word pairs exist. In the conventional technology, the recall rate is set to 1 because word pairs are extracted regardless of the presence or absence of an intended relationship. The precision rate is obtained by dividing the number of correct answers by the number of extractions, that is, the correct answer rate. The F value is an evaluation scale for prediction results.

In the prior art, the number of extractions is 243, the number of correct answers is 88, the recall is 1, the precision is 0.36, and the F value is 0.53. On the other hand, in the ontology generation device of the embodiment, the number of extractions is 135, the number of correct answers is 81, the recall is 0.82, the precision is 0.62, and the F value is 0.72. From this result, although the number of correct answers is smaller in the ontology generation device of the embodiment than in the conventional technology, the number of extractions is also reduced, so the matching rate (correct answer rate) is improved.

<Second embodiment>
The configuration of the ontology generation device of the second embodiment will be explained using FIG. As shown in FIG. 17, the ontology generation device 1700 of the second embodiment differs from the ontology generation device of the first embodiment only in that it has a phrase pair combination generation unit 1701, and the other components are the same. be. Therefore, only the word/phrase pair combination generation unit 1701 will be described, description of the same components will be omitted, and the same reference numerals as in the first embodiment will be used to indicate the same components.

The word/phrase pair combination generation unit 1701 analyzes the text acquired from the text acquisition unit 402, extracts words, and generates a word/phrase combination (also referred to as a combination word/phrase) using words that appear in the same sentence. For example, if the text is as shown in FIG. 7A, then sedan and aaa are in the same sentence, so a phrase combination (such as sedan-aaa) is generated as shown in FIG. Words such as "sedan" and "aaa" indicate words in the set of words of the analysis result in the first embodiment.

The phrase-document-by-document topic distribution learning unit 414 learns the document-by-document topic distribution including the above-mentioned phrase combination as the phrase set in the first embodiment, and stores the learning result (vector) in the phrase-by-document topic distribution storage unit 415. .

The extraction pattern generation candidate evaluation unit 416 adds the word/phrase combination, considers the word/phrase combination, and determines (evaluates) whether or not the semantic relevance score between the word/phrase pair at the appearance location and the seed word/phrase pair exceeds the threshold. do. If the semantic relevance score does not exceed the threshold, the occurrence location (sentence) is deleted from the extraction pattern generation candidate storage unit 413 .

Specifically, the extraction pattern generation candidate evaluation unit 416 uses the word/phrase theme distribution and the document theme distribution shown in FIG.

Here, we describe the computation of the semantic relevance score for two phrase pairs i, j in document d. Phrase pair i is the seed phrase pair and phrase pair j is the extracted phrase pair. In the document d, the semantic relevance score SRScore _{i, j, d} is calculated by the following formula (2) for the phrase pair j={j ^A , j ^B } and the phrase pair i={i ^A , i ^B } . j ^A denotes word A of word pair j (previous word A), j ^B denotes word B of word pair j (back word B), and j ^C denotes combination word C of words A and B of word pair j. Also, iA indicates the word/phrase ^A of the word/phrase pair i (preceding word/phrase A), iB indicates the word/phrase ^B of the word/phrase pair i (backward word/phrase B), and iC indicates the combined word/phrase ^C of the words/phrases A and B of the word/phrase pair i. For example, if the word/phrase pair j is (BBB, Zinrei), then the word/phrase A is BBB, the word/phrase B is Zinrei, and the combined word/phrase C is BBB-Zinrei.

はシード語句対ｉの組み合わせ語句Ｃの文書ｄにおける主題分布を表すベクトルである。これは、例えば主題数を３個（次元Ｋ＝３）とした場合の組み合わせ語句Ｃの主題分布Ｃ_ｋと文書ｄの主題分布Ｄ_ｋの掛け算によって求められる。組み合わせ語句Ｃの主題分布Ｃ_ｋは図１８に示す語句主題分布のうちの語句組み合わせ部分であり、文書ｄの主題分布Ｄ_ｋは図１８に示す文書主題分布である。 here,

is a vector representing the thematic distribution in document d of combination phrase C of seed phrase pair i. This can be obtained, for example, by multiplying the subject distribution _Ck of the combined word/phrase C and the subject distribution _Dk of the document d when the number of themes is three (dimension K=3). The topic distribution _Ck of the combined word/phrase C is the word/phrase combination part of the word/phrase topic distribution shown in FIG. 18, and the topic distribution Dk of the document _d is the document topic distribution shown in FIG.

For example, if the subject distribution C _k of the combination word C is {0.3, 0.1, 0.6} and the subject distribution D _k of the document d is {0.4, 0.4, 0.2}, The subject distribution of combination phrase C in document d is {0.3×0.4, 0.1×0.4, 0.6×0.2}={0.12, 0.04, 0.12} Become.

Also, α _l indicates the weighting factor of the front phrase A, the rear phrase B, and the combination phrase C, and may be set manually in advance, or may be automatically determined and set as described above. good.

The combination words are also taken into consideration in this way because the relationship expressed by the word pair can sometimes be expressed by the combination of the constituent words. For example, in the phrase pair (BBB, Arrows M03) and the phrase pair (CCC, Xpier), "BBB" and "CCC" are smartphone manufacturers, and "Arlows M03" and "Xpier" are smartphones. However, the word pair (BBB, Arrows M03) represents the relationship between the manufacturer's name and its own product, while the word pair (CCC, Xpier) represents the relationship between the manufacturer's name and the products of other companies.

Therefore, by taking combination words into consideration, it is possible to further prevent the extraction of unintended word/phrase pairs.

Next, the flow of extraction processing of relationship extraction patterns in the ontology generation device of the second embodiment will be described. Since the basic processing flow is the same as the processing flow in the first embodiment shown in FIG. 9, only different processing will be described. The different processes are topic distribution learning for each word/phrase document and pattern information extraction process.

A processing flow of subject distribution learning by word/phrase document will be described with reference to FIG. The ontology generator 1700 identifies words in the acquired text (step S1901). That is, words are extracted from the text and the words to be learned are specified. The ontology generation device 1700 generates word/phrase combinations within the same sentence (step S1902). The ontology generation device 1700 learns the document-by-document theme distribution of the identified phrase set (including phrase combinations), that is, performs model learning (step S1903).

Also, the pattern information extraction processing flow will be described with reference to FIG. The ontology generation device 1700 first extracts the occurrence location D of the word/phrase pair j recorded in the extraction pattern generation candidate storage unit 413 (step S2001). For phrase pair j= ^{{ jA} , ^jB } occurrence dj, and for seed phrase pair i= ^{{ iA} , ^iB }, combine phrases ^iC = ^iAiB and ^jC = ^jAjB is generated (step S2002).

The ontology generation device 1700 calculates the semantic relevance scores of the phrases {i ^A , j ^A }, the phrases {i ^B , j ^B }, and the phrases {i ^C , j ^C } in the document group (step S2003). That is, the first word (seed word pair) and the second word (word/phrase pair at the location of occurrence) are compared with each other and after each other, and furthermore, the words and phrases in which the first word and the after word are combined. The semantic relevance score is calculated by comparing the words and phrases that combine the forward and backward phrases of the second word with each other.

The ontology generation device 1700 determines whether or not the calculated semantic relevance score exceeds the threshold (step S2004). Delete (exclude) from the pattern generation candidates (step S2005), and perform the same processing for the next word/phrase pair. On the other hand, if it exceeds (Yes in step S2004), the ontology generation device 1700 performs similar processing for the next seed word pair.

When the processing for all word/phrase pairs is completed, the ontology generation device 1700 generates word/phrase pair relationship extraction patterns from the occurrence locations (filtered) of word/phrase pair j (step S2006).

<Third Embodiment>
The configuration of the ontology generation device of the third embodiment will be explained using FIG. As shown in FIG. 21, the ontology generation device 2100 of the third embodiment differs from the ontology generation device of the second embodiment only in that it has a synonym replacement unit 2101 and a synonym dictionary 2102. are the same. Therefore, only the synonym replacement unit 2101 and the synonym dictionary 2102 will be explained, and the explanation of the same components will be omitted. use.

The synonym replacement unit 2101 unifies words and phrases having the same meaning into one expression regardless of the context or document. That is, when the synonym replacement unit 2101 determines that there is a synonymous word in the word meaning sentence (text) based on predetermined information (synonym dictionary 2102) for replacing synonyms, the synonym replacement unit 2101 replaces the word with a predetermined word. .

For example, since the words "BBB Corporation" and "BBB Corporation" are synonymous, they are unified into "BBB Corporation". At this time, the unification does not change the meaning of the words. As a result, even if the number of documents is small, the learning effect of the word/document-based topic distribution learning unit 414 is improved.

The synonym dictionary 2102 stores dictionary information for performing the synonym replacement. This dictionary information is, for example, prepared in advance by the user, or prepared by some other automatic extraction method. For example, automatic maintenance is possible by the following method.

The first is based on the semantics of phrase constituents. For example, the phrase "color name" and the phrase "color name" both have the same meanings as "color" and "name", so they are considered synonymous.

The second method is based on statistical information of appearance contexts from a large-scale document group. For example, a predetermined tool learns from a large-scale corpus that "BBB" and "BBB Co., Ltd." have the same semantic vector, and regards them as synonyms. A corpus is a large database of linguistic data that is searchable by a computer.

Next, the flow of extraction processing of relationship extraction patterns in the ontology generation device of the third embodiment will be described with reference to FIG. The relationship extraction pattern extraction processing flow in the ontology generation device of the third embodiment differs from the relationship extraction pattern extraction processing flow in the ontology generation devices of the first and second embodiments in that the processing in step S2203 There is one point.

The ontology generation device 2100 receives and acquires seed word/phrase pairs (step S2201). Also, the ontology generation device 2100 acquires a document group from a database that records the document group or the like, and extracts and acquires text from the documents of the acquired document group (step S2202).

The ontology generation device 2100 performs synonym replacement for the acquired text (step S2203). The details of synonym replacement will be described later.

The ontology generation device 2100 performs morphological analysis and syntactic analysis on the synonym-replaced text, or semantic structure analysis of the text, that is, syntactic structure analysis (step S2204). The ontology generation device 2100 performs morphological analysis on the replaced text, and learns the document-by-document theme distribution (Topic Model) of the analysis result word/phrase set, that is, word/phrase-by-document topic distribution learning (step S2205). The processing flow of topic distribution learning by word/phrase document is the same as that of the second embodiment, so the explanation is omitted.

The ontology generation device 2100 compares the occurrence location of the word/phrase pair recorded in the extraction pattern generation candidate storage unit 413 with the seed word/phrase pair recorded in the word/phrase storage unit 411, and compares the occurrence location word/phrase pair with the seed word/phrase pair. It is determined (evaluated) whether or not the semantic relevance score of exceeds the threshold. If the semantic relevance score does not exceed the threshold, the occurrence location (sentence) is deleted from the extraction pattern generation candidate storage unit 413 . That is, the ontology generation device 2100 performs pattern information extraction processing (step S2206). The flow of the pattern information extraction process is the same as that of the second embodiment, so description thereof will be omitted.

The ontology generation device 2100 extracts the word/phrase corresponding to the variable using the generated relationship extraction pattern, that is, performs word/phrase information extraction processing (step S2207).

Here, the substitution of synonyms will be described with reference to FIGS. 23A and 23B. FIG. 23A shows the text after synonym replacement is performed on the acquired text. Also, FIG. 23B shows the learning result when learning is performed without subjecting the acquired text to synonym replacement, and the learning result when learning is performed after subjecting the acquired text to synonym replacement. there is

As shown in FIG. 23A, the acquired text includes "BBB" and "BBB Co., Ltd.", which are synonymous and unified into "BBB" to replace the text. Since synonyms are replaced and unified, as shown in FIG. 23B, the learning result when learning with replacement is less than the learning result when learning without replacement.

<Fourth Embodiment>
The configuration of the ontology generation device of the fourth embodiment will be described. Since the configuration of the ontology generation device of the fourth embodiment is the same as that of the second embodiment, the configuration diagram shown in FIG. explain. The fourth embodiment is characterized in that when the word/phrase pair combination generation unit 1701 generates a combination of words/phrases, the combinations are generated in consideration of the modification relation of the words/phrases.

That is, the word/phrase pair combination generation unit 1701 considers the modification relationship in the occurrence location (sentence) and processes a combination of other words/phrases close to the analysis target word/phrase as one word/phrase. For example, in the appearance (sentence) of the phrase pair (BBB, Arrows M03), "This year's best smartphone award was awarded to Arrows M03, the latest product of BBB", the phrase modification relationship is shown in FIG. .

A combination of words and phrases that are closely related to a word and phrase are registered as one word and phrase. The modification structures within the same sentence are treated as an undirected graph, and the search distance between each word is obtained. For example, the phrase “Arlows M03” can be searched up to “current year” by following the path (search distance is 3) of “Arlows M03”→“award”→“best smartphone award”→“current year”. A search distance is restricted, and a modifier group within that distance is registered as one word. For example, if the search distance is 1, consider "latest product--Arlows M03--awarded" as one phrase. Also, for example, when the search distance is the maximum, "BBB-latest product-Arlows M03-award-best smartphone award-this year" is regarded as one phrase.

Therefore, in the occurrence of the phrase pair (BBB, Arrows M03) (when the search distance is 1), the modifier-related phrases of “BBB” (aggregate of phrases close to BBB) “BBB-latest product” and “ "Arlows M03" is generated as a modification relational phrase (aggregate of phrases close to Arrows M03) "latest product-Arlows M03-award". The search distance may be set in advance by the user, or may be set by a predetermined automatic generation method.

Here, FIG. 24 shows an example of learning results of the theme distribution of words and phrases including combinations generated in consideration of the modification relation and the document theme distribution. The original text is the text shown in FIG. 7A. The bold-framed portion of the word/phrase theme distribution shown in FIG. 24 indicates word/phrase combinations (such as “BBB-AAA-new business”) with a search distance of 1 in the modification relation within the same sentence.

Calculation of the semantic relevance score in the fourth embodiment will be described below using Equation (3). Specifically, the extraction pattern generation candidate evaluation unit 416 uses the word/phrase topic distribution and the document topic distribution shown in FIG.

Here, we describe the computation of the semantic relevance score for two phrase pairs i, j in document d. Phrase pair i is the seed phrase pair and phrase pair j is the extracted phrase pair. In document d, the semantic relevance score SRScore _{i, j, d} is calculated for the word pair j={j ^A , j ^B } and the word pair i={i ^A , i ^B } by the following equation (3). . j ^A denotes word A of word pair j (previous word A), j ^B denotes word B of word pair j (back word B), and j ^C denotes combination word C of words A and B of word pair j. In addition, j ^AG is a modifier set of forward modifiers A (forward modifiers AG), j ^BG is a modifier set of backward modifiers B (back modifiers BG), and j ^CG is a combination of forward modifiers AG and backward modifiers BG. A set of modifiers (combined modifiers CG) for the phrase CG is shown.

Also, iA indicates the word/phrase ^A of the word/phrase pair i (preceding word/phrase A), iB indicates the word/phrase ^B of the word/phrase pair i (backward word/phrase B), and iC indicates the combined word/phrase ^C of the words/phrases A and B of the word/phrase pair i. In addition, i ^AG is a modifier set of forward modifiers A (forward modifiers AG), i ^BG is a modifier set of backward modifiers B (back modifiers BG), and i ^CG is a combination of forward modifiers AG and backward modifiers BG. A set of modifiers (combined modifiers CG) for the phrase CG is shown. For example, taking the above example, if the phrase pair j is (BBB, Arrows M03), then phrase A is "BBB", phrase B is "Arlows M03", and combination phrase C is "BBB-Arlows M03". Also, the phrase AG is "BBB-latest product", the phrase BG is "latest product-Arlows M03-awarded", and the phrase CG is "BBB-latest product-Arlows M03-awarded".

はシード語句対ｉの前方修飾語句ＡＧの文書ｄにおける主題分布を表すベクトルである。 here,

is a vector representing the subject distribution in document d of the prefix modifiers AG of seed word pair i.

α _l indicates the weighting factor of the forward phrase A, the backward phrase B, the combination phrase C, the front modifier AG, the rear modifier BG, and the combination modifier CG, and may be set manually in advance, as described above. may be determined and set automatically.

Next, the flow of extraction processing of relationship extraction patterns in the ontology generation device of the fourth embodiment will be described. Since the basic processing flow is the same as the processing flow in the first embodiment shown in FIG. 9, only different processing will be described. The different processes are topic distribution learning for each word/phrase document and pattern information extraction process.

A processing flow of subject distribution learning by word/phrase document will be described with reference to FIG. The ontology generation device 1700 identifies words in the acquired text (step S2501). That is, words are extracted from the text and the words to be learned are specified. The ontology generation device 1700 generates combinations considering the modification relationships of words in the same sentence (step S2502). The ontology generation device 1700 learns the document-by-document theme distribution of the identified phrase sets (including phrase combinations), that is, performs model learning (step S2503).

Also, the pattern information extraction processing flow will be described with reference to FIG. The ontology generation device 1700 first extracts the occurrence location D of the word/phrase pair j recorded in the extraction pattern generation candidate storage unit 413 (step S2601). The ontology generation device 1700 considers the word modification relation in dj for the occurrence location dj of the word/phrase pair j={ ^jA , ^jB }, and considers the word/phrase modification relation in dj, and the aggregate of words/phrases close to ^jAG = ^jA and ^jBG = A set of words close to j ^B and j ^CG =j ^AG j ^BG are generated (step S2602).

The ontology generation device 1700 generates i ^AG , i ^BG , and i ^CG for the seed word/phrase pair i={i ^A , i ^B } (step S2603).

The ontology generation device 1700 generates the terms {i ^A ,j ^A }, terms {i ^B ,j ^B }, terms {i ^C ,j ^C }, terms {i ^AG ,j ^AG }, terms {i ^BG ,j ^BG }. , the semantic relevance score in the document group of the phrase {i ^CG ,j ^CG } is calculated (step S2604). That is, the modifier sets of forward phrases of the first word (seed word pair) and the second word (word phrase pair of occurrence location), the modifier sets of backward phrases, and the modifiers of the forward phrases of the first word A semantic relevance score is calculated including a comparison between a phrase combining the set and the modifier set of the backward phrase and a phrase combining the modifier set of the forward phrase of the second word and the modifier set of the backward phrase.

The ontology generation device 1700 determines whether or not the calculated semantic relevance score exceeds the threshold (step S2605). Delete (exclude) from the pattern generation candidates (step S2606), and perform the same processing for the next word/phrase pair. On the other hand, if it exceeds (Yes in step S2605), the ontology generator 1700 performs similar processing for the next seed word/phrase pair.

When the processing for all the word/phrase pairs is completed, the ontology generation device 1700 generates a relationship extraction pattern of the word/phrase pairs from the occurrence locations (filtered) of the word/phrase pair j (step S2607).

<Fifth Embodiment>
In the fifth embodiment, a search system using ontologies generated by the ontology generation devices of the first to fourth embodiments will be described. As shown in FIG. 27, in this search system, ontology generators 40 (1700, 2100) are arranged on a network 2701, and ontology generators 40 are created and updated daily by users of a specific community (XX department, company). It acquires from the department document DB 2702 storing the document group. The ontology generation device 40 generates an ontology from the acquired document group, and records the generated ontology in the XX department ontology DB 2703 . The keyword expansion device 2704 automatically expands the input keyword based on the input keyword input to the search server 2705 and the ontology of the XX department ontology DB 2703 . The search server 2705 searches based on the automatically expanded keywords. Note that a system composed of the ontology generation device 40 and the keyword expansion device 2704 is also called an automatic keyword expansion system.

Automatic extension by the keyword extension device 2704 will be specifically described with reference to FIG. Consider a case where the input keyword is, for example, "carbon". Keywords are automatically expanded based on the input keyword "carbon" and the ontology. For the input keyword "carbon", if the ontology is used, for example, the English name "carbon" and the Japanese name "carbon" as well as words such as "carbon paper/take a copy" are extracted. In addition, "C" is extracted as a chemical formula, and "carbon fiber", "carbon fiber", "acrylic", "pitch" and the like are extracted as related terms. Further, as subordinate concepts (solid lines) and related concepts (dashed lines) of related terms, "PAN-based carbon fiber", "carbonized fiber", "graphite fiber", "pitch-based carbon fiber", etc. are extracted.

In this way, if the input keyword "carbon" is expanded based on the ontology, carbon; carbon/carbon paper/take a copy; C; carbon fiber; carbon fiber; Keywords are extended, such as fiber. By searching based on this expanded keyword, it becomes possible to search for many related documents as search results.

According to one aspect of the ontology generation device of each embodiment, it is possible to improve the accuracy of extracting word/phrase pairs having an intended relationship without being affected by the results of intermediate processing.

The following notes are further disclosed with respect to the above embodiments.
(Appendix 1) An ontology generation program executed by an ontology generation device that generates an ontology dictionary that is a semantic database,
Divide the acquired document data into multiple word meaning sentences,
Identifying words that constitute concept pairs, which are sets of lexical concepts, for the divided multiple word meaning sentences,
Whether the first word forming the concept pair corresponding to the first word meaning sentence and the second word forming the concept pair corresponding to the second word meaning sentence are similar among the plurality of word meaning sentences is determined, and if it is determined that they are not similar, the words that make up the concept pair are excluded from being registered in the ontology dictionary;
An ontology generation program characterized by causing a computer of the ontology generation device to execute processing.
(Supplementary Note 2) If the non-similarity is determined, delete the word meaning sentence in which the second word appears from the extraction pattern candidates for extracting words having a similar relationship with the first word, 1. The ontology generation program according to appendix 1, wherein the second words forming the concept pair are excluded from targets to be registered in the ontology dictionary.
(Appendix 3) A score indicating the degree of similarity between the first word and the second word is calculated by a predetermined calculation formula, and if the calculated score is equal to or less than a predetermined threshold, the first word and the second word are 3. The ontology generation program according to appendix 1 or 2, characterized by determining that two words are dissimilar.
(Appendix 4) A morphological analysis is performed on the word-sense sentence to extract words, and there is a word/phrase topic distribution in which the extracted words/phrases and combination words/phrases combining the extracted words/phrases refer to the topic, and the extracted words/phrases exist. learning the document subject distribution in which the semantic sentences refer to the subject,
3. The ontology generation program according to appendix 3, wherein the score is calculated using the word/phrase subject distribution and the document subject distribution.
(Additional remark 5) The score is calculated by comparing the forward phrases and the backward phrases of the first word and the second word, or a phrase in which the forward phrases and the backward phrases of the first word are combined. 5. The ontology generation program according to appendix 4, wherein the score is calculated including a comparison of words and phrases that are a combination of a prefix phrase and a posterior phrase of the second word.
(Appendix 6) Modifier sets of forward phrases of the first word and second word, modifier sets of backward phrases of each other, modifier set of forward phrases of the first word and modifiers of backward phrases The score is calculated according to Supplementary Note 5, wherein the score is calculated including a comparison between a phrase combining a set and a modifier set of prefix phrases of the second word and a modifier set of backward phrases of the second word. An ontology generator.
(Appendix 7) When it is determined that a synonymous word exists in the word-meaning sentence, it is replaced with a predetermined word based on predetermined information for replacing the synonym, and after the replacement, the word-thematic distribution is learned. 7. An ontology generation program according to any one of appendices 4 to 6, characterized in that:
(Appendix 8) An ontology generation method for generating an ontology dictionary that is a semantic database,
Divide the acquired document data into multiple word meaning sentences,
Identifying words that constitute concept pairs, which are sets of lexical concepts, for the divided multiple word meaning sentences,
Whether the first word forming the concept pair corresponding to the first word meaning sentence and the second word forming the concept pair corresponding to the second word meaning sentence are similar among the plurality of word meaning sentences is determined, and if it is determined that they are not similar, the words that make up the concept pair are excluded from being registered in the ontology dictionary;
An ontology generation method characterized by:
(Supplementary note 9) when it is determined that the second word is not similar, deleting the word meaning sentence in which the second word appears from the extraction pattern candidates for extracting words having a relationship similar to the first word, 8. The ontology generating method according to appendix 8, wherein the second words forming the concept pair are excluded from targets to be registered in the ontology dictionary.
(Appendix 10) A score indicating the similarity between the first word and the second word is calculated by a predetermined calculation formula, and if the calculated score is equal to or less than a predetermined threshold, the first word and the second word are 10. The ontology generation method according to appendix 8 or 9, wherein the two words are determined to be dissimilar.
(Appendix 11) A morphological analysis is performed on the word-sense sentence to extract words, and a word/phrase topic distribution in which the extracted words/phrases and combination words/phrases combining the extracted words/phrases refer to the topic, and the extracted words/phrases exist. learning the document subject distribution in which the semantic sentences refer to the subject,
11. The ontology generating method according to appendix 10, wherein the score is calculated using the word topic distribution and the document topic distribution.
(Appendix 12) The score is calculated by comparing forward phrases and backward phrases of the first word and the second word, or a phrase combining forward phrases and backward phrases of the first word. 12. The ontology generating method according to Supplementary Note 11, wherein the score is calculated including a comparison of words and phrases that are a combination of a prefix phrase and a posterior phrase of the second word.
(Appendix 13) Modifier sets of forward phrases of the first word and second word, modifier sets of backward phrases of each other, modifier set of forward phrases of the first word and modifiers of backward phrases 13. The score according to Supplementary Note 12, wherein the score is calculated including a comparison between a phrase combining a set and a modifier set of prefix phrases of the second word and a modifier set of backward phrases of the second word. Ontology generation method.
(Appendix 14) When it is determined that a synonymous word exists in the word-sense sentence, it is replaced with a predetermined word based on predetermined information for replacing the synonym, and after the replacement, the word-thematic distribution is learned. 14. An ontology generation method according to any one of appendices 11 to 13, characterized in that:
(Appendix 15) An ontology generation device that generates an ontology dictionary that is a semantic database,
a dividing unit that divides the acquired document data into a plurality of word meaning sentences;
an identification unit that identifies words that form concept pairs, which are sets of lexical concepts, for the plurality of divided word meaning sentences;
Whether the first word forming the concept pair corresponding to the first word meaning sentence and the second word forming the concept pair corresponding to the second word meaning sentence are similar among the plurality of word meaning sentences and, if it is determined that they are not similar, a processing unit that excludes the words that make up the concept pair from being registered in the ontology dictionary;
a keyword expansion device that automatically expands the input keyword based on the inputted input keyword and the ontology dictionary generated by the ontology generation device;
A keyword automatic expansion system comprising:

11 document group 12 relationship extraction pattern DB
13 Phrases vs. DB
14 users 30, 32 patterns 31, 33 phrase pairs 40, 1700, 2100 ontology generator 130 CPU
131 HDDs
132 RAMs
133 ROMs
134 input device 135 communication interface 136 bus 401 seed phrase input unit 402 text acquisition unit 403 syntax analysis unit 404 extraction processing control unit 405 extraction pattern generation candidate extraction unit 406 pattern extraction unit 407 phrase extraction unit 408 ontology generation unit 409 output unit 410 syntax Structure storage unit 411 Word/phrase storage unit 412 Extraction pattern storage unit 413 Extraction pattern generation candidate storage unit 414 Word/phrase document-based theme distribution learning unit 415 Word/phrase document-based subject distribution storage unit 416 Extraction pattern generation candidate evaluation unit 1400 Inappropriate relationship extraction pattern 1701 Phrase pair combination generator 2101 Synonym replacement unit 2102 Synonym dictionary 2701 Network 2702 Department document DB
2703 XX Department Ontology DB
2704 Keyword expansion device 2705 Search server

Claims (6)

An ontology generation device that generates an ontology dictionary that is a semantic database,
A phrase pair, which is a pair of phrases having semantic relevance, is composed of a pair of anterior phrase and a posterior phrase, and the posterior phrase is posterior to the anterior phrase in a semantic sentence containing the phrase pair. a storage unit that stores the word/phrase pair arranged at the position as an item of the ontology dictionary;
an acquisition unit that acquires a seed word/phrase pair that is the word/phrase pair input from the outside and stores the seed word/phrase pair in the storage unit as the item;
a dividing unit that divides the acquired document data into a plurality of word meaning sentences;
a word sense sentence extracting unit for extracting a word sense sentence including the word/phrase pair stored as the item in the storage unit from the plurality of word sense sentences obtained by the division;
Predetermining a score representing a degree of semantic relevance between each word forming the word pair and each word forming the seed word pair included in the word meaning sentence extracted by the word meaning sentence extracting unit It is determined whether the relevance is low by comparing the score calculated for the word sense sentence with a predetermined threshold, and if the relevance is determined to be low, the word sense sentence is an evaluation unit that excludes from the results of the extraction;
In the word-sense sentence remaining as a result of the extraction without being excluded by the evaluation unit, each word and phrase constituting the word-and-phrase pair included in the word-sense sentence is provided with attribute information about each word and phrase. a generation unit that generates a relationship extraction pattern by replacing with variables;
Attribute information about a word or phrase whose word or phrase other than the variable matches the relation extraction pattern and is arranged at the position of the variable in the relation extraction pattern, among the plurality of word meaning sentences obtained by the division. a word/phrase pair extracting unit for extracting a word/phrase pair located at the position from a word meaning sentence matching attribute information of the variable and storing the word/phrase pair in the storage unit as the item;
has
The predetermined calculation formula is:
a vector distance between the subject distribution for the forward term of the word pair contained in the semantic sentence and the subject distribution for the forward term of the seed word pair;
A weighted average of vector distances between the subject distribution for the backward term of the word pair included in the semantic sentence and the subject distribution for the backward term of the seed word pair, the score is a formula to calculate as
a subject distribution for a word in a semantic sentence is represented by a vector obtained by multiplying a vector representing the word subject distribution for the word and a vector representing the document subject distribution for the semantic sentence;
A word/phrase theme distribution for a word is a probability distribution representing the probability that the word/phrase is included in a document related to the subject for each subject,
A document topic distribution for a semantic sentence is a probability distribution that expresses the probability that the semantic sentence is included in documents related to the subject for each subject.
An ontology generation device characterized by: An ontology generation device that generates an ontology dictionary that is a semantic database,
A phrase pair, which is a pair of phrases having semantic relevance, is composed of a pair of anterior phrase and a posterior phrase, and the posterior phrase is posterior to the anterior phrase in a semantic sentence containing the phrase pair. a storage unit that stores the word/phrase pair arranged at the position as an item of the ontology dictionary;
an acquisition unit that acquires a seed word/phrase pair that is the word/phrase pair input from the outside and stores the seed word/phrase pair in the storage unit as the item;
a dividing unit that divides the acquired document data into a plurality of word meaning sentences;
a word sense sentence extracting unit for extracting a word sense sentence including the word/phrase pair stored as the item in the storage unit from the plurality of word sense sentences obtained by the division;
Predetermining a score representing a degree of semantic relevance between each word forming the word pair and each word forming the seed word pair included in the word meaning sentence extracted by the word meaning sentence extracting unit It is determined whether the relevance is low by comparing the score calculated for the word sense sentence with a predetermined threshold, and if the relevance is determined to be low, the word sense sentence is an evaluation unit that excludes from the results of the extraction;
In the word-sense sentence remaining as a result of the extraction without being excluded by the evaluation unit, each word and phrase constituting the word-and-phrase pair included in the word-sense sentence is provided with attribute information about each word and phrase. a generation unit that generates a relationship extraction pattern by replacing with variables;
Attribute information about a word or phrase whose word or phrase other than the variable matches the relation extraction pattern and is arranged at the position of the variable in the relation extraction pattern, among the plurality of word meaning sentences obtained by the division. a word/phrase pair extracting unit for extracting a word/phrase pair located at the position from a word meaning sentence matching attribute information of the variable and storing the word/phrase pair in the storage unit as the item;
has
The predetermined calculation formula is:
a vector distance between the subject distribution for the forward term of the word pair contained in the semantic sentence and the subject distribution for the forward term of the seed word pair;
a vector distance between the subject distribution for the back term of the word pair contained in the semantic sentence and the subject distribution for the back term of the seed word pair;
the vector distance between the subject distribution of the word combinations of the word pairs included in the word meaning sentence and the subject distribution of the word combinations of the seed word pairs; A weighted average value is a formula for calculating the score,
a subject distribution for a word in a semantic sentence is represented by a vector obtained by multiplying a vector representing the word subject distribution for the word and a vector representing the document subject distribution for the semantic sentence;
a subject distribution for a word combination in a sense sentence is represented by a vector obtained by multiplying a vector representing a word subject distribution for the word combination and a vector representing a document subject distribution for the word sense sentence;
A word/phrase theme distribution for a word is a probability distribution representing the probability that the word/phrase is included in a document related to the subject for each subject,
The word/phrase theme distribution for a combination word is a probability distribution representing the probability that all the words/phrases combined as the combination word/phrase are included in a document related to the subject for each subject,
A document topic distribution for a semantic sentence is a probability distribution that expresses the probability that the semantic sentence is included in documents related to the subject for each subject.
An ontology generation device characterized by: An ontology generation device that generates an ontology dictionary that is a semantic database,
A phrase pair, which is a pair of phrases having semantic relevance, is composed of a pair of anterior phrase and a posterior phrase, and the posterior phrase is posterior to the anterior phrase in a semantic sentence containing the phrase pair. a storage unit that stores the word/phrase pair arranged at the position as an item of the ontology dictionary;
an acquisition unit that acquires a seed word/phrase pair that is the word/phrase pair input from the outside and stores the seed word/phrase pair in the storage unit as the item;
a dividing unit that divides the acquired document data into a plurality of word meaning sentences;
a word sense sentence extracting unit for extracting a word sense sentence including the word/phrase pair stored as the item in the storage unit from the plurality of word sense sentences obtained by the division;
Predetermining a score representing a degree of semantic relevance between each word forming the word pair and each word forming the seed word pair included in the word meaning sentence extracted by the word meaning sentence extracting unit It is determined whether the relevance is low by comparing the score calculated for the word sense sentence with a predetermined threshold, and if the relevance is determined to be low, the word sense sentence is an evaluation unit that excludes from the results of the extraction;
In the word-sense sentence remaining as a result of the extraction without being excluded by the evaluation unit, each word and phrase constituting the word-and-phrase pair included in the word-sense sentence is provided with attribute information about each word and phrase. a generation unit that generates a relationship extraction pattern by replacing with variables;
Attribute information about a word or phrase whose word or phrase other than the variable matches the relation extraction pattern and is arranged at the position of the variable in the relation extraction pattern, among the plurality of word meaning sentences obtained by the division. a word/phrase pair extracting unit for extracting a word/phrase pair located at the position from a word meaning sentence matching attribute information of the variable and storing the word/phrase pair in the storage unit as the item;
has
The predetermined calculation formula is:
a vector distance between the subject distribution for the forward term of the word pair contained in the semantic sentence and the subject distribution for the forward term of the seed word pair;
a vector distance between the subject distribution for the back term of the word pair contained in the semantic sentence and the subject distribution for the back term of the seed word pair;
A vector distance between a theme distribution for a combination of words that combine the words that make up the word pair included in the word meaning sentence and a theme distribution for a combination of words that make up the seed word pair;
A vector distance between a subject distribution for a modifier set for the preceding term of the word pair contained in the semantic sentence and a subject distribution for a modifier set for the preceding term of the seed word pair. When,
A vector distance between a subject distribution for a modifier set for the backward term of the word pair included in the semantic sentence and a subject distribution for a modifier set for the backward term of the seed word pair. When,
Thematic distribution of a modifier set for a combined word/phrase combining each word/phrase constituting the word/phrase pair contained in the word meaning sentence and a modifier set for a combined word/phrase combining each word/phrase constituting the seed word/phrase pair A formula for calculating the weighted average value of the vector distance from the subject distribution for the score as the score,
A modifier set for words and phrases forming a word pair included in a word meaning sentence is a set whose elements are the words and phrases that modify the words and phrases in the word meaning sentence,
A modifier set for a combination word that combines word pairs included in a meaning sentence is a set whose elements are each word that is combined as the combination word and a word that modifies each word in the meaning sentence. and
a subject distribution for a word in a semantic sentence is represented by a vector obtained by multiplying a vector representing the word subject distribution for the word and a vector representing the document subject distribution for the semantic sentence;
a subject distribution for a combination word in a semantic sentence is represented by a vector obtained by multiplying a vector representing the word theme distribution for the combination word by a vector representing the document subject distribution for the semantic sentence;
The subject distribution for the modifier set in the word sense sentence is represented by a vector obtained by multiplying a vector representing the word theme distribution for the modifier set by a vector representing the document subject distribution for the word sense sentence,
A word/phrase subject distribution for a word is a probability distribution representing the probability that the word/phrase is included in a document related to the subject for each subject,
The word/phrase theme distribution for a combination word is a probability distribution representing the probability that all the words/phrases combined as the combination word/phrase are included in a document related to the subject for each subject,
The word/phrase subject distribution for a modifier set is a probability distribution that expresses the probability that all the words that are elements of the modifier set are included in documents related to the subject for each subject,
The document topic distribution for a semantic sentence is a probability distribution representing the probability that the semantic sentence is included in the document related to the subject for each subject.
An ontology generation device characterized by: It is determined based on predetermined information for substituting synonyms whether or not synonymous words exist in the plurality of word meaning sentences obtained by the division, and when it is determined that they exist, it is determined that they exist. further comprising a synonym replacement unit that replaces the synonymous words in the word meaning sentence with a predetermined word,
The word sense sentence extraction unit extracts the word sense sentences from the word sense sentences replaced by the synonym replacement unit and the remaining word sense sentences among the plurality of word sense sentences obtained by the division.
4. The ontology generation device according to any one of claims 1 to 3 , characterized by: An ontology generation program executed by an ontology generation device that generates an ontology dictionary that is a semantic database,
A phrase pair, which is a pair of phrases having semantic relevance, is composed of a pair of anterior phrase and a posterior phrase, and the posterior phrase is posterior to the anterior phrase in a semantic sentence containing the phrase pair. Obtaining a seed word pair, which is the word pair input from the outside, from the word pairs arranged at the position ;
storing the seed word pair as an item in a word/phrase pair storage unit that stores the word/phrase pair as an item of the ontology dictionary;
Divide the acquired document data into multiple word meaning sentences,
Extracting a word meaning sentence containing the word pair stored as the item in the word pair storage unit from the plurality of word meaning sentences obtained by the division,
Predetermining a score representing a degree of semantic relevance between each word and phrase constituting the word pair and each word and phrase making up the seed word pair included in the word meaning sentence extracted by the word meaning sentence extraction Calculated using the formula,
determining whether the relevance is low by comparing the score calculated for the word-sense sentence extracted by the word-sense sentence extraction with a predetermined threshold;
excluding the word meaning sentence from the result of extracting the word meaning sentence when it is determined that the relevance is low;
In the word meaning sentence that is not excluded by the above-mentioned exclusion and remains as a result of the extraction of the word meaning sentence, each word and phrase that constitutes the word and phrase pair included in the word meaning sentence is obtained by obtaining attribute information for each word and phrase. Generate a relationship extraction pattern by replacing variables with
Attribute information about a word or phrase whose word or phrase other than the variable matches the relation extraction pattern and is arranged at the position of the variable in the relation extraction pattern, among the plurality of word meaning sentences obtained by the division. extracts a pair of words and phrases arranged at the position from the word meaning sentence that matches the attribute information of the variable, and stores the word and phrase pairs as the items in the word and phrase pair storage unit. let
The predetermined calculation formula, in the word sense sentence extracted by the word sense sentence extraction,
a vector distance between the subject distribution for the forward term of the word pair contained in the semantic sentence and the subject distribution for the forward term of the seed word pair;
a vector distance between the subject distribution for the back term of the word pair contained in the semantic sentence and the subject distribution for the back term of the seed word pair;
is a formula for calculating the weighted average value of as the score,
a subject distribution for a word in a semantic sentence is represented by a vector obtained by multiplying a vector representing the word subject distribution for the word and a vector representing the document subject distribution for the semantic sentence;
A word/phrase theme distribution for a word is a probability distribution representing the probability that the word/phrase is included in a document related to the subject for each subject,
A document subject distribution for a semantic sentence is a probability distribution representing the probability that the semantic sentence is included in a document related to the subject for each subject,
An ontology generator. An ontology generation method for generating an ontology dictionary, which is a semantic database, comprising:
A phrase pair, which is a pair of phrases having semantic relevance, is composed of a pair of anterior phrase and a posterior phrase, and the posterior phrase is posterior to the anterior phrase in a semantic sentence containing the phrase pair. Obtaining a seed word pair, which is the word pair input from the outside, from the word pairs arranged at the position ;
storing the seed word pair as an item in a word/phrase pair storage unit that stores the word/phrase pair as an item of the ontology dictionary;
Divide the acquired document data into multiple word meaning sentences,
Extracting a word meaning sentence containing the word pair stored as the item in the word pair storage unit from the plurality of word meaning sentences obtained by the division,
Predetermining a score representing a degree of semantic relevance between each word and phrase constituting the word pair and each word and phrase making up the seed word pair included in the word meaning sentence extracted by the word meaning sentence extraction Calculated using the formula,
determining whether the relevance is low by comparing the score calculated for the word-sense sentence extracted by the word-sense sentence extraction with a predetermined threshold;
excluding the word meaning sentence from the result of extracting the word meaning sentence when it is determined that the relevance is low;
In the word meaning sentence that is not excluded by the above-mentioned exclusion and remains as a result of the extraction of the word meaning sentence, each word and phrase that constitutes the word and phrase pair included in the word meaning sentence is obtained by obtaining attribute information for each word and phrase. Generate a relationship extraction pattern by replacing variables with
Attribute information about a word or phrase whose word or phrase other than the variable matches the relation extraction pattern and is arranged at the position of the variable in the relation extraction pattern, among the plurality of word meaning sentences obtained by the division. The computer of the ontology generation device executes a process of extracting the pair of words and phrases arranged at the position from the word meaning sentence matching the attribute information of the variable and storing them as the items in the word and phrase pair storage unit. ,
The predetermined calculation formula, in the word sense sentence extracted by the word sense sentence extraction,
a vector distance between the subject distribution for the forward term of the word pair contained in the semantic sentence and the subject distribution for the forward term of the seed word pair;
a vector distance between the subject distribution for the back term of the word pair contained in the semantic sentence and the subject distribution for the back term of the seed word pair;
is a formula for calculating the weighted average value of as the score,
a subject distribution for a word in a semantic sentence is represented by a vector obtained by multiplying a vector representing the word subject distribution for the word and a vector representing the document subject distribution for the semantic sentence;
A word/phrase theme distribution for a word is a probability distribution representing the probability that the word/phrase is included in a document related to the subject for each subject,
A document subject distribution for a semantic sentence is a probability distribution representing the probability that the semantic sentence is included in a document related to the subject for each subject,
Ontology generation method.

Images