JP5503632B2 - Feature word extraction method, apparatus, and program - Google Patents

Description

  The present invention relates to a feature word extraction method, apparatus, and program, and more particularly, to a feature word extraction method, apparatus, and program for extracting feature words in each topic of a hierarchical topic model in which topics have a hierarchical structure.

  Conventionally, there is a topic model in which words included in a document set are feature amounts and probabilistic clustering is performed using one document as a data point. There is also a hierarchical topic model in which topics have a hierarchical structure. A method of extracting feature words in each topic of the topic model based on the ratio between the parameter average of the entire topic model and the parameter in each topic has been proposed (see, for example, Non-Patent Document 1).

D. Blei, A. Y. Ng, M. Jordan, "Latent Dirichlet Allocation", Journal of Machine Learning Research 3 (2003) 993-1022

  However, when the method described in Non-Patent Document 1 is applied to a topic in an intermediate hierarchy of the hierarchical topic model, a word characteristic of any topic located in the lower level of the intermediate hierarchy is There is a problem that feature words that are easily selected as feature words in a topic and suitable for a topic in an intermediate hierarchy, that is, feature words common to concepts in the intermediate hierarchy and lower cannot be extracted.

  The present invention has been made in view of the above circumstances, and a feature word extraction method capable of extracting a feature word common to a concept below the intermediate hierarchy as a feature word in a topic in an intermediate hierarchy of the hierarchical topic model An object is to provide a device and a program.

  To achieve the above object, the feature word extraction method of the present invention includes a parameter representing each of a plurality of topics for probabilistic clustering of document data, and the topic has a hierarchical topic model having a hierarchical structure. A target parameter extraction step for extracting a target parameter representing a topic to be processed, a peer hierarchy parameter extraction step for extracting a peer hierarchy parameter representing a topic of the target subject and a peer hierarchy from the hierarchical topic model, A lower layer parameter extracting step for extracting a lower layer parameter representing a lower layer topic corresponding to the processing target topic from the hierarchical topic model, and using the target parameter, the peer layer parameter, and the lower layer parameter , Included in the target parameter For each word, a score calculating step for calculating a score based on the intra-topic variance in the lower hierarchy and the inter-topic variance in the peer hierarchy, and based on the score of each word included in the calculated target parameter, A feature word extraction step for extracting feature words in the topic to be processed.

  According to the feature word extraction method of the present invention, the target parameter extraction step includes a parameter representing each of a plurality of topics for probabilistic clustering of document data, and the topic has a hierarchical topic model having a hierarchical structure. Extract the target parameter that represents the topic to be processed, extract the peer hierarchy parameter that represents the topic to be processed and the topic of the peer hierarchy from the hierarchical topic model in the peer hierarchy parameter extraction step, From the hierarchical topic model, a lower hierarchy parameter representing a lower hierarchy topic corresponding to the topic to be processed is extracted. Then, in the score calculation step, using the target parameter, peer hierarchy parameter, and lower hierarchy parameter, for each word included in the target parameter, a score based on the intra-topic variance in the lower hierarchy and the inter-topic variance in the peer hierarchy is calculated. In the feature word extraction step, feature words in the processing target topic are extracted based on the score of each word included in the calculated target parameter.

  In this way, feature words are extracted by calculating the score based on the intra-topic variance in the lower hierarchy and the inter-topic variance in the peer hierarchy. It is possible to extract feature words common to the concepts.

  Further, the score can be increased as the variance of each word in the lower-level topics is smaller and the variance among the topics in the peer hierarchy is larger. In this way, by directly reflecting the intuitive features of the feature words common to the concepts below the intermediate hierarchy in the score, it is possible to extract an appropriate feature word as the feature word in the topic of the intermediate hierarchy. .

  The feature word extraction device of the present invention represents a topic to be processed from a hierarchical topic model that includes parameters representing each of a plurality of topics for probabilistic clustering of document data, and the topics have a hierarchical structure. From the target parameter extracting means for extracting the target parameter, from the hierarchical topic model, from the hierarchical topic model, the isotopic hierarchy parameter extracting means for extracting the isotopic hierarchy parameter representing the topic to be processed and the topic of the peer hierarchy, and the hierarchical topic model, Included in the target parameter using lower layer parameter extraction means for extracting a lower layer parameter representing a lower layer topic corresponding to the topic to be processed, the target parameter, the peer layer parameter, and the lower layer parameter For each word Score calculation means for calculating a score based on variance within the pick and variance between topics in the peer hierarchy, and feature words in the processing target topic based on the score of each word included in the calculated target parameter And feature word extraction means for extracting.

  The feature word extraction program of the present invention is a program for causing a computer to execute each step of the above feature word extraction method.

  As described above, according to the feature word extraction method, apparatus, and program of the present invention, in order to extract a feature word by calculating a score based on the intra-topic distribution in the lower hierarchy and the inter-topic variance in the peer hierarchy, As a feature word in a topic in an intermediate hierarchy of the hierarchical topic model, it is possible to extract a feature word common to concepts in the intermediate hierarchy and lower.

It is a functional block diagram which shows the structure of the feature word extraction apparatus which concerns on this Embodiment. It is the schematic which shows a hierarchical topic model. It is a flowchart which shows the content of the feature word extraction process routine in the feature word extraction apparatus which concerns on this Embodiment.

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

  The feature word extraction device 10 according to the present embodiment extracts feature words in each topic of the hierarchical topic model. The feature word extraction device 10 can be configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a feature word extraction processing routine described later. As shown in FIG. 1, this computer functionally includes a target component model extraction unit 12, a peer layer component model extraction unit 14, a lower layer component model extraction unit 16, a score calculation unit 18, a feature word It can be expressed by a configuration including the extraction unit 20.

  Here, the topic model is a model for performing probabilistic clustering using a word (v) included in a document set as a feature amount and one document as a data point (d), and using topic model parameters, It is a model represented by the following formula (1).

In the equation (1), the topic model parameters are p (z) (1 × Kmatrix) and p (v | z) (K × Vmatrix). z is a random variable representing a topic, p (z) is a prior probability for the random variable z, p (v | z) is a probability based on a multinomial distribution under the random variable z, and n _dv is a word v in the document data d. The number of appearances.

  A hierarchical topic model is a topic model in which topics have a hierarchical structure. For example, as shown in FIG. 2, topic 2, topic 2-2 including topic 1, topic 2, topic 3, and topic 4 in the first hierarchy and dividing topic 2 into three as the second hierarchy. , And a topic 2-3. Each topic in each hierarchy is represented by the above topic model parameters. For the topic 2 of the first hierarchy, the topic model parameters of the topic 2 before the division are separated from the topic model parameters representing the topics 2-1, 2-2, and 2-3 of the second hierarchy after the division. As it is.

  The target component model extraction unit 12 receives the hierarchical topic model 22 as an input, and extracts a topic model component (parameter) (hereinafter referred to as “target component model”) of the target topic number k from the hierarchical topic model 22. Here, p (v | k) (1 × Vmatrix) is extracted as the target component model 24. In the following, a case will be described in which k = 2 and topic 2 in FIG. 2 is the target topic.

The peer hierarchy component model extraction unit 14 receives the hierarchical topic model 22 as an input, and the topic model component (parameter) in the topic existing in the peer hierarchy with the topic of the topic number k from the hierarchical topic model 22 (hereinafter, “ "Isotopic hierarchy component model"). Here, p (v | t _B ) is extracted as the peer hierarchy component model. t _B is the topic number of the topic in the peer hierarchy, and when there are a plurality of topics in the peer hierarchy, a plurality of peer hierarchy component models are extracted and set as the peer hierarchy component model group 26. For example, if Topic 2 is the target topic k, the topic model component model in each of Topic 1, Topic 3, and Topic 4 in the first hierarchy is extracted as the peer hierarchy component model.

The lower hierarchical component model extraction unit 16 receives the hierarchical topic model 22 as an input, and uses topic model components (parameters) (hereinafter referred to as “a topic topic component”) in a topic existing in the lower hierarchy of the topic number k that is the target from the hierarchical topic model 22. "Lower layer component model"). Here, p (v | t _W ) is extracted as the lower layer component model. t _W is the topic numbers in the lower hierarchy of topics, if the lower layer of the topic there are multiple, extracts a plurality of lower layer component models, a lower layer component model group 28. For example, if Topic 2 is the target topic k, the topic model component model in each of Topic 2-1, Topic 2-2, and Topic 2-3 in the second hierarchy is extracted as the lower-layer component model.

  The score calculation unit 18 calculates a score for all words included in the target component model 24, and extracts words having a score equal to or higher than a predetermined threshold as feature words.

Here, the principle of the score calculated by the score calculation unit 18 will be described. The object of the present invention is to extract feature words common to the concepts of the lower hierarchy in the topic of the intermediate hierarchy. Such feature words include:
1. Less distributed among lower-level topics (commonality)
2. Large variance among topics in peer hierarchy (characteristic)
This feature is intuitive. Putting this feature directly in the score function leads to extracting desired common feature words. Therefore, a score function is used in consideration of the intra-topic variance in the lower hierarchy and the inter-topic variance in the peer hierarchy.

In accordance with the above principle, the score calculation unit 18 calculates, for example, the intra-topic variance σ _{W (v)} in the lower hierarchy for each word by the following equation (2), and the inter-topic variance σ _{B (v)} in the peer hierarchy: It is possible to calculate using the equation (3) and the ratio (equation (4)) as a score.

Here, v is a word, k is a topic number of a target topic, t _W is a topic number of a topic in a lower hierarchy, t _B is a topic number of a topic in a peer hierarchy, and c _W is the number of topics in a lower hierarchy (here, 3) , C _B is the number of topics in the peer hierarchy (here, 3). Note that the score used in the score calculation unit 18 is not limited to the above case, and may be any score as long as the features of the common feature words described above can be expressed.

  The feature word extraction unit 20 sorts the words based on the score of each word calculated by the score calculation unit 18, and extracts and outputs the top N words as feature words in the topic k. Moreover, you may make it extract a feature word according to the comparison result of a score and a threshold value using a predetermined threshold value.

  Next, a feature word extraction processing routine executed by the feature word extraction device 10 of the present embodiment will be described with reference to FIG.

In step 100, the hierarchical topic model 22 is acquired, and the target component model 24 (p (v | k)) of the target topic number k is extracted from the hierarchical topic model 22. Next, in step 102, a peer hierarchical component model group 26 (p (v | t _B )) is extracted from the hierarchical topic model 22 in the topic having the topic number k and the topic existing in the peer hierarchy. Next, in step 104, a lower layer component model group 28 (p (v | t _W )) in the topic existing in the lower layer of the topic with the topic number k is extracted from the hierarchical topic model 22.

  Next, in step 106, scores are calculated for all words included in the target component model 24 using a score function that takes into account the intra-topic variance in the lower hierarchy and the inter-topic variance in the peer hierarchy. Next, in step 108, the words are sorted based on the score of each word calculated in step 106, the top N words are extracted and output as feature words in topic k, and feature word extraction processing is performed. finish.

  As described above, according to the feature word extraction device of the present embodiment, the topic (target model) of the feature word, the topic of the peer hierarchy, and the parameters (component models) of the topics of the lower hierarchy are extracted, Calculates the score considering the intra-topic variance of the hierarchy and the inter-topic variance of the peer hierarchy, and extracts words with low variance in the lower-level topics and large variance between topics in the peer hierarchy as feature words in the target topic Therefore, it is possible to extract a feature word common to the concepts below the intermediate hierarchy as a feature word in the topic of the intermediate hierarchy of the hierarchical topic model.

  The present invention is not limited to the above embodiment, and various modifications and applications can be made without departing from the gist of the present invention.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium. Further, the feature word extraction device of the present invention may be configured by hardware such as a semiconductor integrated circuit for realizing the above processing.

DESCRIPTION OF SYMBOLS 10 Feature word extraction apparatus 12 Target component model extraction part 14 Isotopic hierarchy component model extraction part 16 Lower hierarchy component model extraction part 18 Score calculation part 20 Feature word extraction part

Claims (4)

A target parameter extracting step for extracting a target parameter representing a processing target topic from a hierarchical topic model including a parameter representing each of a plurality of topics for probabilistic clustering of document data and the topic having a hierarchical structure;
A peer hierarchy parameter extracting step for extracting a peer hierarchy parameter representing a topic of the processing target and a peer hierarchy from the hierarchical topic model;
A lower layer parameter extracting step for extracting, from the hierarchical topic model, a lower layer parameter representing a lower layer topic corresponding to the topic to be processed;
Using the target parameter, the peer hierarchy parameter, and the lower hierarchy parameter, for each word included in the target parameter, a score based on the intra-topic variance of the lower hierarchy and the inter-topic variance of the peer hierarchy is calculated. A score calculation step;
A feature word extraction step of extracting a feature word in the processing target topic based on the score of each word included in the calculated target parameter;
A feature word extraction method.   The feature word extraction method according to claim 1, wherein the score is increased as the variance of topics in a lower hierarchy of each word is smaller and as the variance between topics in the peer hierarchy is greater. A target parameter extracting means for extracting a target parameter representing a processing target topic from a hierarchical topic model including a parameter representing each of a plurality of topics for probabilistic clustering of document data and the topic having a hierarchical structure;
A peer hierarchy parameter extracting means for extracting a peer hierarchy parameter representing a topic of the processing target and a peer of the peer hierarchy from the hierarchical topic model;
A lower layer parameter extracting means for extracting a lower layer parameter representing a lower layer topic corresponding to the processing target topic from the hierarchical topic model;
Using the target parameter, the peer hierarchy parameter, and the lower hierarchy parameter, for each word included in the target parameter, a score based on the intra-topic variance of the lower hierarchy and the inter-topic variance of the peer hierarchy is calculated. A score calculation means;
Feature word extraction means for extracting a feature word in the processing target topic based on the score of each word included in the calculated target parameter;
A feature word extraction device.   A feature word extraction program for causing a computer to execute each step of the feature word extraction method according to claim 1.

Images