JP2017068742A - Relevant document retrieval device, model creation device, method and program therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a relevant document retrieval technique capable of retrieving more accurately relative to prior arts.SOLUTION: A relevant document retrieval device comprises: a feature amount extraction part 3 for extracting a similarity feature amount group about an input sentence and each retrieval object sentence, in which a similarity feature amount group about a certain sentence and a certain retrieval object sentence is plural different feature amounts indicating a similarity between the certain sentence and the certain retrieval object sentence; a similarity score model storage part 4 for storing a similarity score model indicating a relationship between the similarity feature amount group and a similarity score corresponding to the similarity feature amount group; and a similarity score calculation part 5 for calculating the similarity score about the input sentence and each retrieval object sentence using the similarity score model stored in the similarity score model storage part 4 and the similarity feature amount group which is extracted.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a technique for searching for a document related to an input sentence.

  The techniques for searching for documents related to input sentences are roughly classified as follows: (1) A technique for extracting words (keywords) used for search from input natural sentences and searching for documents that match the keywords; ) The relationship between the entire input sentence and the search target document is calculated, and the technique is classified into a technique for searching for a highly relevant document. Hereinafter, (1) is a keyword matching method, and (2) is a sentence similarity determination method.

  In the keyword matching method, as in Non-Patent Document 1, a transposed index is prepared in advance, and a document including a keyword associated with an input sentence is obtained by subtracting the transposed index. At that time, the keyword is expanded in terms of similarity, character fluctuation, etc., and a document including the expanded keyword is obtained in addition to the original keyword, thereby enabling a search with a higher recall.

  In the sentence similarity determination method, as in Patent Document 1, a search is performed by converting an input sentence and a search target document into concept vectors and using the document similarity as the closeness between the concept vectors (which is obtained as a cosine measure).

Search engine mechanisms and technology development (Information Science and Technology 54 (2), 66-71, 2004-02-01)

JP 2007-317132 A

  However, in the keyword matching method, keywords included in the input sentence, expanded keywords, etc. are used, but since only the keywords are focused, only local words are seen without looking at the meaning of the whole document. Therefore, there is a case where it is determined that documents having low relevance are similar in terms of meaning.

  In the document similarity judgment method, the closeness of the whole document is seen, so the whole document has multiple meanings / topics, and there are descriptive sentences (greetings, introductions, explanations of special words, etc.). If it is included, the concept of the document may be blurred, and the relevance may not be determined correctly.

  An object of the present invention is to provide a related document search device, a model creation device, and a method and a program thereof that enable a search with higher accuracy than before.

  The related document search apparatus according to an aspect of the present invention provides a plurality of different feature amounts representing similarity between a certain sentence and the certain search target document as a similarity feature amount group between the certain sentence and the certain search target document. The relationship between the feature quantity extraction unit that extracts the similarity feature quantity group for the input sentence and each search target document, and the similarity score corresponding to the similarity feature quantity group and the similarity feature quantity group An input sentence using the similarity score model storage unit storing the similarity score model to be represented, the similarity score model stored in the similarity score model storage unit, and the extracted similarity feature amount group; A similarity score calculation unit that calculates a similarity score for each search target document.

  In the model creation device according to an aspect of the present invention, the similarity feature amount group for a certain sentence and a certain document is input as a plurality of different feature amounts representing the similarity between the certain sentence and the certain document. The feature amount extraction unit that extracts a similarity feature amount group for a sentence and each document, and the similarity feature amount extracted assuming that the similarity score between the input sentence and each document is predetermined. By performing regression analysis using the similarity score corresponding to the extracted similarity score group as the objective variable, the similarity score group corresponding to the similarity score group and the similarity score corresponding to the similarity score group And a creation unit for creating a similarity score model representing the relationship between the two.

  Search with higher accuracy than before is possible.

The block diagram for demonstrating the example of a related document search apparatus. The flowchart for demonstrating the example of a related document search method. The block diagram for demonstrating the example of a model production apparatus. The flowchart for demonstrating the example of a model creation method. 6 is a diagram showing an example of information about a search target document stored in a search target document storage unit 1. FIG. The figure which shows the example of the 1st feature-value. The figure which shows the example of the 2nd feature-value. The figure which shows the example of the similarity degree feature-value group extracted by the feature-value extraction part 3. FIG. The figure which shows the example of the pair of the word memorize | stored in the output condition memory | storage part. The figure which shows the example of the some document memorize | stored in the learning document memory | storage part 10, and a similarity score. The figure which shows the example of the similarity feature-value group extracted by the feature-value extraction part 8. FIG.

[Related Document Retrieval Apparatus and Method]
Hereinafter, an embodiment of the related document search apparatus and method will be described with reference to the drawings. As shown in FIG. 1, the related document search apparatus includes a search target document storage unit 1, a search target document narrowing unit 2, a feature amount extraction unit 3, a similarity score model storage unit 4, and a similarity score calculation. The unit 5, the output condition storage unit 6, and the output unit 7 are provided, for example. Each unit of the related document search apparatus performs the process of each step in FIG. 2 to realize the related document search method.

<Search target document storage unit 1>
The search target document storage unit 1 stores a plurality of search target documents.

  The search target document is stored in the search target document storage unit 1 together with a search target document ID that is an identifier for identifying the search target document.

<Search target document filtering unit 2>
The search target document narrowing unit 2 receives the text and the search target document read from the search target document storage unit 1.

  The search target document narrowing unit 2 determines the category of the input document, and selects the search target document of the determined category from the plurality of search target documents stored in the search target document storage unit 1 ( Step S2).

  First, the search target document narrowing unit 2 extracts the category of the input document using, for example, the following text pattern extraction technique, topic estimation technique, and multi-value classification technique. Of course, the search target document narrowing unit 2 may extract the category of the input document using another category determination technique.

  See Reference 1 for examples of text pattern extraction techniques.

[Reference 1] Nippon Telegraph and Telephone Corporation, “Text Knowledge Extraction Technology“ Rich Indexer ””, [online], [searched September 24, 2015], Internet <URL: http://www.ntt.co .jp / svlab / activity / category_2 / product2_07.html>
See Reference 2 for examples of topic estimation techniques (LDA, etc.).

[Reference 2] David M. Blei, two others, “Latent Dirichlet Allocation”, [online], Journal of Machine Learning Research 3 (2003) 993-1022, [searched September 24, 2015], Internet < URL: https://www.cs.princeton.edu/~blei/papers/BleiNgJordan2003.pdf>
See Reference 3 for examples of multilevel classification techniques (multilayer perceptron, SVC (SVM), etc.).

  [Reference 3] Asa Ben-Hur, three others, “Support Vector Clustering”, [online], Journal of Machine Learning Research 2 (2001) 125-137, [searched September 24, 2015], Internet < URL: http://www.jmlr.org/papers/volume2/horn01a/rev1/horn01ar1.pdf>

  When the processing of the search target document narrowing unit 2 is performed, as illustrated in FIG. 5, the search target document storage unit 1 stores the search target document category in advance and is stored together with the search target document ID. Suppose that

  The search target document narrowing unit 2 selects a search target document having the same category as the determined input document category from a plurality of search target documents stored in the search target document storage unit 1 and is selected. Outputs the search target document ID of the search target document.

  By narrowing down the search target, it is possible to efficiently search by reducing search processing for unnecessary documents, and it is expected to improve accuracy by excluding documents with different topics from the search target.

<Feature Extraction Unit 3>
The feature amount extraction unit 3 receives the input text and the search target document selected by the search target document narrowing unit 2. The search target document selected by the search target document narrowing unit 2 is specified by the search target document ID output by the search target document narrowing unit 2.

  The feature amount extraction unit 3 extracts a similarity feature amount group for the input sentence and each search target document selected by the search target document narrowing unit 2 (step S3). The extracted similarity feature quantity group is output to the similarity score calculation unit 5 together with the corresponding search target document ID, for example.

  Here, a similarity feature amount group for a certain sentence and a certain search target document is set as a plurality of different feature amounts representing the similarity between the certain sentence and the certain search target document. In other words, the similarity feature amount group is a set of feature amounts obtained by a plurality of techniques having different properties.

  For example, the inter-document feature amount obtained by the first technique (for example, keyword matching method) and the document amount obtained by the second technique (for example, sentence similarity determination method) having different properties from the first technique. A pair with the feature amount of the second feature amount can be used as the similarity feature amount. Hereinafter, the process of the feature amount extraction unit 3 will be described with reference to this example. Of course, this is merely an example, and a set of feature amounts between documents obtained by each of three or more technologies may be used as the similarity feature amount. Further, the first technique and the second technique are not limited to the keyword matching method and the document similarity determination method. If there is a technique having other properties capable of calculating information indicating the degree of similarity between documents, it may be used.

  The first calculation unit 31 of the feature quantity extraction unit 3 uses the first feature quantity representing the similarity between the input sentence and each search target document selected by the search target document narrowing unit 2 as the first feature quantity. Calculate based on technology.

  FIG. 6 shows an example of the first feature amount. FIG. 6 shows an example of the first feature amount between the input document and the search target document of each search target document ID. In the example of FIG. 6, the first feature amount is a vector composed of three elements. As in the example of FIG. 6, the feature quantity, the first feature quantity, and the second feature quantity may be vectors composed of a plurality of elements.

  In addition, the second calculation unit 32 of the feature amount extraction unit 3 determines the second feature amount representing the similarity between the input sentence and each search target document selected by the search target document narrowing unit 2. Calculate based on 2 techniques.

  FIG. 7 shows an example of the second feature amount. FIG. 7 shows an example of the second feature amount between the input document and the search target document of each search target document ID.

  Then, the feature quantity extraction unit 3 combines the first feature quantity and the second feature quantity into a similarity feature quantity group.

  At the time of combination, the feature quantity extraction unit 3 normalizes each of the first feature quantity and the second feature quantity, and uses the normalized first feature quantity and second feature quantity as a similarity feature quantity group. Good. For example, normalization is performed for each element of the first feature value and the second feature value.

  FIG. 8 shows an example of the similarity feature quantity group obtained by combining the feature quantity normalized from the first feature quantity in FIG. 6 and the feature quantity normalized from the second feature quantity in FIG. .

Hereinafter, an example of normalization processing will be described. Normalization is performed using, for example, one of the normalization functions f (x) exemplified below. Which normalization processing is performed for each feature amount (or whether normalization processing is not performed) is determined in advance. Further, when the feature amount is composed of one or more elements, normalization is performed by applying a normalization function to each element of the feature amount. In the following formula, x represents the element value of the feature quantity before normalization. a, σ ₁ , σ ₂ , and σ ₃ are constants.
f (x) = tanh (x)
f (x) = 1 / (1 + e ^-ax )
f (x) = σ ₁ / (σ ₂ + σ ₃ | x |)

<Similarity score model storage unit 4>
The similarity score model storage unit 4 stores a similarity score model representing the relationship between the similarity feature quantity group and the similarity score corresponding to the similarity feature quantity group.

  The similarity score model is, for example, a linear regression model. Of course, the similarity score model may be a regression model other than the linear regression model.

  The similarity score model is created by a model creation device and method described later with reference to FIGS. 3 and 4.

<Similarity score calculation unit 5>
The similarity score calculation unit 5 receives the similarity feature group corresponding to each search target document extracted by the feature amount extraction unit 3 and the similarity score model read from the similarity score model storage unit 4. . The search target document is specified by the search target document ID.

  The similarity score calculation unit 5 uses the similarity score model stored in the similarity score model storage unit 4 and the similarity feature amount group extracted by the feature amount extraction unit 3 to input text and each search A similarity score for the target document is calculated (step S5). The similarity score is an index representing the height of similarity between documents (including the case between sentences and documents) for which similarity is calculated. Further, it can be said that the similarity score is a combination of similarity feature groups calculated by a plurality of technologies having different properties.

  The similarity score model is composed of regression coefficients obtained by regression analysis, for example. In this case, the similarity score calculation unit 5 calculates an output value when the similarity feature amount group is input to the expression specified by the regression coefficient obtained by the regression analysis, and the calculation result is used as the similarity score. And When the similarity score model is a linear regression model, the similarity score calculation unit 5 calculates the inner product of the similarity feature quantity group that is a vector and the regression coefficient that is a vector, and the calculation result is used as the similarity. Score.

  The similarity score calculation unit 5 outputs a pair of a similarity score and a search target document ID corresponding to the similarity score to the output unit 7. At that time, the similarity score calculation unit 5 sorts the pairs of the similarity score and the search target document ID corresponding to the similarity score in descending or ascending order with respect to the similarity score, and in the sorted order. It may be output.

  The similarity score calculation unit 5 learns a ranking learner that is an existing technology in advance, inputs a similarity feature amount group to the learner, outputs the similarity, and uses it as a similarity score. Good. In this case, the ranking learner corresponds to the similarity score model.

  Further, the similarity score calculation unit 5 may manually determine the weights of the feature amounts constituting the similarity feature amount group, obtain the inner product of the weight vector and the similarity feature amount group, and use the similarity score as the similarity score. In this case, the weight of the feature amount configured manually corresponds to the similarity score model.

<Output condition storage unit 6>
The output condition storage unit 6 stores output conditions used by the output unit 7.

<Output unit 7>
The output unit 7 receives a pair of the similarity score output by the similarity score calculation unit 5 and the search target document ID corresponding to the similarity score.

  The output unit 7 outputs information about the search target document having a high similarity score (step S7). The information about the search target document is, for example, a search target document ID. The information about the search target document may be the search target document itself.

  The output unit 7 may read the output condition from the output condition storage unit 6 and output information about the search target document that satisfies the read output condition.

  For example, the output unit 7 may output information on search target documents having the highest α similarity score. In this case, the information that the “similarity score has the top α” is the output condition (hereinafter referred to as output condition (1)). α is an integer of 1 or more.

  Further, the output unit 7 may output information on a search target document whose similarity score calculated by the similarity score calculation unit 5 is equal to or greater than a predetermined threshold. In this case, the information “above a predetermined threshold” is the output condition (hereinafter referred to as output condition (2)). When the similarity score is set to be higher as the similarity score is smaller, the output unit 7 determines a search target document whose similarity score calculated by the similarity score calculation unit 5 is equal to or less than a predetermined threshold. May be output.

  Furthermore, it is assumed that a plurality of word pairs are stored in the output condition storage unit 6 as so-called NG words. An example of word pairs stored in the output condition storage unit 6 is shown in FIG.

  In this case, the output unit 7 seems to include one word constituting the word pair in the input document and the other word constituting the word pair in the search target document. If there is a pair of words in a plurality of pairs of words stored in the output condition storage unit 6, a process of not outputting the search target document may be performed. In other words, when a pair of a word included in the input sentence and a word included in the search target document is stored in the output condition storage unit 6, a process of not outputting the search target document may be performed. .

  In this case, “a pair of words in which the input document includes one word constituting a word pair and the search target document includes the other word constituting the word pair. However, if it is in a plurality of word pairs stored in the output condition storage unit 6, the condition that the search target document is not output is referred to as an output condition (hereinafter referred to as an output condition (3)). Become.

  The output condition (3) can be compatible with the output condition (1) or (2). That is, only search target documents that do not satisfy the output condition (3) are output among search target documents that satisfy the output condition (1) or (2), and search target documents that satisfy the output condition (3) are not output. Processing may be performed.

  In this way, by obtaining the final similarity score in consideration of two or more types of feature amounts extracted by a plurality of techniques having different properties, a search with higher accuracy than before can be performed.

[Model creation apparatus and method]
Hereinafter, an embodiment of a model creation apparatus and method will be described with reference to the drawings. As shown in FIG. 3, the model creation apparatus includes a learning document storage unit 10, a feature amount extraction unit 8, and a creation unit 9, for example. Each part of the model creation device performs the process of each step in FIG. 4 to realize a model creation method.

<Learning document storage unit 10>
The learning document storage unit 10 stores a plurality of documents. A plurality of documents are associated with a similarity score for every two different documents. This similarity score is predetermined by hand, for example. The plurality of documents may be the same as or different from the search target documents stored in the search target document storage unit 1. Documents having a high degree of similarity may be stored as a plurality of documents.

  FIG. 10 shows an example of a plurality of documents and similarity scores stored in the learning document storage unit 10. In the example of FIG. 10, a document ID that is an identifier is assigned to each of a plurality of documents. In the upper part of FIG. 10, the similarity score for the document with the document ID of qid00001 and the document with the document ID of pid00001 is 3, and the similarity score of the document with the document ID of qid00001 and the document with the document ID of pid00003 is 1. It represents something. In this example, the similarity is set to be higher as the similarity score is larger.

<Feature Extraction Unit 8>
The feature quantity extraction unit 8 extracts a similarity feature quantity group for two different documents (step 8). The extracted similarity feature quantity group is output to the creation unit 9.

  The process of the feature quantity extraction unit 8 is the same as the process of the feature quantity extraction unit 3 described above. That is, the feature amount extraction unit 3 performs a similarity feature amount group for each of two different documents by a process similar to the process of extracting the similarity feature amount group for the input sentence and each search target document. To extract. In other words, the feature quantity extraction unit 8 extracts the same similarity feature quantity group as the similarity feature quantity group extracted by the feature quantity extraction unit 3. The process of the first calculation unit 81 of the feature quantity extraction unit 8 is the same as the process of the first calculation unit 31 of the feature quantity extraction unit 3, and the process of the second calculation unit 82 of the feature quantity extraction unit 8 is the feature quantity extraction unit. 3 is the same as the processing of the second calculation unit 32.

  FIG. 11 shows an example of the similarity feature quantity group extracted by the feature quantity extraction unit 8.

<Creation unit 9>
The creation unit 9 uses the extracted similarity feature quantity group as an explanatory variable, and performs a regression analysis using the similarity score corresponding to the similarity feature quantity group extracted by the feature quantity extraction unit 8 as an objective variable. A similarity score model representing the relationship between the degree feature quantity group and the similarity score corresponding to the similarity feature quantity group is created (step S9).

  The creation unit 9 performs, for example, linear regression analysis. In this case, the creation unit 9 learns and obtains parameters of the linear regression model using, for example, SVM.

  The created similarity score model is stored in the similarity score model storage unit 4 of FIG.

[Program and recording medium]
The processing described in the related document search device and method and the model creation device and method is not only executed in time series according to the order of description, but also in parallel or individually according to the processing capability of the device that executes the processing or as necessary. May be executed.

  When each process in the related document search device or model creation device is realized by a computer, the processing contents of the functions that the related document search device or model creation device should have are described by a program. Then, by executing this program on a computer, each process is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  Each processing means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

[Modification]
The processing of the search target document narrowing unit 2 may not be performed. In this case, the feature amount extraction unit 3 extracts a similarity feature amount group for the input sentence and each search target document stored in the search target document storage unit 1.

  Needless to say, other modifications are possible without departing from the spirit of the present invention.

DESCRIPTION OF SYMBOLS 1 Search object document memory | storage part 2 Search object document narrowing-down part 3 Feature-value extraction part 31 1st calculation part 32 2nd calculation part 4 Similarity score model memory | storage part 5 Similarity score calculation part 6 Output condition memory | storage part 7 Output part 8 Feature amount extraction unit 81 First calculation unit 82 Second calculation unit 9 Creation unit 10 Document storage unit for learning

Claims (10)

A group of similarities between a certain sentence and a certain search target document is used as a plurality of different feature quantities representing the similarity between the certain sentence and the certain search target document. A feature amount extraction unit that extracts a similarity feature amount group for
A similarity score model storage unit storing a similarity score model representing a relationship between the similarity feature quantity group and the similarity score corresponding to the similarity feature quantity group;
Using the similarity score model stored in the similarity score model storage unit and the extracted similarity feature quantity group, a similarity score for calculating a similarity score for the input sentence and each search target document A calculation unit;
Related document search device including The related document search device according to claim 1,
The similarity score model is a linear regression model.
Related document retrieval device. The related document search device according to claim 1 or 2,
An output unit that outputs information about a search target document in which the calculated similarity score is greater than or less than a predetermined threshold;
Related document retrieval device. The related document search device according to claim 1 or 2,
An output unit for outputting a search target document having a high similarity score;
An output condition storage unit that stores a plurality of word pairs;
The output unit includes one word constituting a word pair included in the input document and the other word constituting the word pair included in the search target document. When the word pair is in the plurality of word pairs stored in the output condition storage unit, the search target document is not output.
Related document retrieval device. The related document search device according to any one of claims 1 to 4,
A search target document storage unit storing a plurality of search target documents; and
A search target document filtering unit that determines a category of the input document and selects a search target document of the determined category from a plurality of search target documents stored in the search target document storage unit; In addition,
The feature amount extraction unit extracts a similarity feature amount group for the input sentence and each of the selected search target documents.
Related document retrieval device. The sentence inputted by the feature quantity extraction unit as a plurality of different feature quantities representing the similarity degree between the certain sentence and the certain search target document, the similarity degree group of the certain sentence and the certain retrieval target document And feature amount extraction step for extracting a similarity feature amount group for each search target document,
The similarity score calculation unit uses the similarity score model representing the relationship between the similarity feature quantity group and the similarity score corresponding to the similarity feature quantity group and the extracted similarity feature quantity group to input A similarity score calculating step for calculating a similarity score for each of the retrieved sentences and each search target document;
Related document search method including   The program for functioning a computer as each part of the related document search apparatus in any one of Claim 1 to 5. Similarity feature quantities for an input sentence and each document are defined as a plurality of different feature quantities representing similarity between the sentence and the document. A feature extraction unit for extracting a group;
Assuming that the similarity score between the input sentence and each document is predetermined, the extracted similarity feature amount group is used as an explanatory variable, and the similarity corresponding to the extracted similarity feature amount group A creation unit that creates a similarity score model that represents a relationship between a similarity feature quantity group and a similarity score corresponding to the similarity feature quantity group by performing regression analysis with the degree score as an objective variable;
Model creation device including Similarity between two different documents in which the feature quantity extraction unit uses a similarity feature quantity group for a given sentence and a given document as a plurality of different feature quantities representing the similarity between the given sentence and the given document. A feature amount extraction step for extracting a degree feature amount group;
Assuming that the similarity score between the two different documents is determined in advance, the creation unit uses the extracted similarity feature amount group as an explanatory variable, and the similarity corresponding to the extracted similarity feature amount group A creation step for creating a similarity score model representing a relationship between a similarity feature group and a similarity score corresponding to the similarity feature group by performing regression analysis with the score as an objective variable;
Model creation method including   The program for functioning a computer as each part of the model production apparatus of Claim 8.

Images