JP4146361B2 - Label display type document search apparatus, label display type document search method, computer program for executing label display type document search method, and computer readable recording medium storing the computer program - Google Patents

Description

  The present invention relates to a label display type document search apparatus that displays a document label and then displays the document.

  In a search system in a computer network, the following is known as a search system for efficiently narrowing down search results.

・ Search system with ranking
In a keyword input type search system typified by google (registered trademark), content including the input keyword is compared with the input keyword (Non-Patent Document 1) and PageRank indicating the importance of the content (non- By ranking the contents in order (described in Patent Document 2), the desired contents can be reached more efficiently.

・ Relevance Feedback system A system that presents search results to the user, returns the user's evaluation for the search results to the search system once, and returns the search results by changing the search condition formula based on the information. is there. Thereby, the user can obtain a search result close to what the user intends (described in Non-Patent Document 3).

Clustering system A method of generating clusters based on the similarity between documents based on the assumption that “relevant documents are similar”, and classifying and presenting search results to the user. The user can efficiently reach desired information without evaluating all the contents included in the search result (described in Non-Patent Document 4).

Query expansion system A method for presenting keywords related to a query input by the user, and the user interactively correcting and modifying the query to efficiently obtain desired content. There are a method of acquiring a related word from a text corpus in advance and a method of using data obtained by analyzing a search result obtained from an input search request (described in Non-Patent Document 5).
tf-idf; Salton, G. et al. “Introduction to Modern Information Retrieval” McGraw-Hill Book Company, 1983 Brin, S. and Page, L., “The Anatomy of a Large-Scale Hypertextual Web Search Engine” Proceedings of 7th WWW Conference, 1998. Ricardo Baeza-Yates and Berthier Ribeiro-Neto, “Modern Information Retrieval”, 1999 Anton Leuski, “Evaluating Document Clustering for Interactive Information Retrieval”, Proceedings of the 2001 ACM CIKM International Conference on Information and Knowledge Management, 2001. H. Sakai, K. Ohtake and S. Masuyama, A retrieval support system by suggesting terms to a user, in Proceedings 2001 International Conference on Chinese Language Computing, 2001.

  However, each search system described above has the following disadvantages.

  In a search system with ranking, search results are presented in a list with priorities. If the search results are not sufficiently narrowed down by the search request, does the user search for the desired content from a huge list of search results? New search conditions for narrowing down must be prepared and re-searched. Of course, the former is expensive and the latter is generally known to be difficult.

  The Relevance Feedback system can improve the search request by evaluating the conformity and nonconformity for the top few to several tens of search results, and can obtain the search result closer to the user's request. Therefore, although it may be intuitive in terms of technique, in practice, a user must determine many documents in order to determine conformity / nonconformity of documents. This technique is useful as an approach that emphasizes recall so that it can cost more for a single search and ensure that all relevant documents are found, but it can be as good as selecting a TV channel. In the approach of finding out, the cost to the user is too great.

  A clustering system can help a user reach a desired search result by classifying search results. Generally, a clustering system trades off with the quality of clustering due to constraints on the processing time of clustering. Must be taken into account. Therefore, a method such as the K-Means method that predetermines the number of clusters is taken. However, if the number of actual topic classifications does not match the determined value, an unclear cluster is generated, making it difficult to label the contents of each cluster. There is a problem that the contents may be unclear and cannot be understood.

  Although there is a categorizing system, that is, a system that solves the problem with labels by putting documents that fit the label in a basket with labels in advance, it is assumed that the category is generated manually. In addition, there is a problem that the category definition and the category maintenance associated with the content update are costly for the information retrieval system administrator.

  In the query expansion system, it is possible for the user to efficiently expand the search expression by using a word that co-occurs in the query and the document, and this makes it possible to narrow down the search result easily. If the attributes of the query candidates are not taken into consideration, the level of the words that are the query candidates becomes uneven, and it becomes difficult to select information from the entire search results.

  In other words, in the conventional search system, the user is forced to search for content from an enormous list, and in the means for solving this, information that compels the user and the system administrator is costly or is provided. There is a problem that the method itself is insufficient.

  SUMMARY An advantage of some aspects of the invention is that it provides a document search apparatus that eliminates the need for a setting change according to a document update by an administrator and a complicated operation by a user.

In order to solve the above problems, the first aspect of the present invention is a document including a body composed of a character string, further including a title of the document and document identification information indicating the document, and a predetermined character string. A document storage unit storing a plurality of documents whose attribute values are included in the text, and a document vector corresponding to each document stored in the document storage unit, the number of attribute values included in the document being an attribute value Document vector generation means for generating a document vector included for each of the plurality of documents, and for each attribute value included in at least one of the plurality of documents stored in the document storage means, the number of appearances of the attribute value in the plurality of documents A statistical processing unit that generates recorded statistical information and stores the statistical information in a storage unit provided in advance, a document search unit that searches a plurality of documents from the document storage unit, and any one of the searched plurality of documents. Search result statistical information generating means for generating and storing search result statistical information in which the number of occurrences of the attribute value in the plurality of documents is recorded for each included attribute value, and any one of the plurality of searched documents For each attribute value included in at least the attribute value, the attribute value for calculating the degree of suitability indicating the degree to which the attribute value is suitable as a label that is a character string representing a plurality of documents forming a part of the plurality of documents. A degree-of-fit calculation means for calculating the degree of match of the attribute value using a calculation formula using the number of appearances in the statistical information of the value and the search result statistical information, and any one of the plurality of retrieved documents Selecting an attribute value corresponding to the fitness level from the higher fitness level corresponding to each of a plurality of attribute values consisting of at least included attribute values as long as the fitness level satisfies a preset condition. By Label information generating means for selecting a plurality of attribute values forming a part of the plurality of attribute values, using the selected attribute values as labels, and generating label information including the plurality of labels, and the label information For each label included in the document, the document identification information indicating the document that includes the character string that is the label and that is one of the retrieved documents, and the cluster information that includes the titles as many as the number of the documents A cluster information generating means for generating cluster information including the label and a vector sum of the document vectors corresponding to each of the plurality of documents including the document identification information included in the cluster information for each of the cluster information. A document that generates a certain cluster vector, includes document identification information that is not document identification information included in the cluster information, and is stored in the document storage unit. A cosine scale between the document vector corresponding to the stored document and the cluster vector is calculated. If the cosine scale exceeds a preset threshold value, the document identification information and title of the document are included in the cluster information. The cluster information changing means including the label information, and each label included in the label information is displayed, and one set of the document identification information and the title included in the cluster information including the label is displayed by selecting the label. And a document display control means for reading out and displaying a document including the document identification information and title set from the document storage means when one set of the document identification information and title is selected. A label display type document retrieval apparatus that performs the above processing is used as a solution.

A second aspect of the present invention is a label display type document search method performed by a label display type document search device, wherein the label display type document search device includes a document storage unit, a document vector generation unit, a statistical processing unit, and a document search unit. Search result statistical information generation means, fitness calculation means, label information generation means, cluster information generation means, cluster information change means, and document display control means, and the document storage means includes a text string. The label display type document retrieval method, wherein a plurality of documents, each of which includes a title of the document and document identification information indicating the document, and a plurality of attribute values that are predetermined character strings are included in the body, , the document vector generation means, a document vectors for each document stored in the document storage means, the number of attribute values included in the document for each attribute value A document vector generating step for generating a document vector, and the statistical processing means for each of the plurality of documents having the attribute value for each attribute value included in any of the plurality of documents stored in the document storage means. A statistical processing step for generating statistical information in which the number of occurrences is recorded and storing the statistical information in a storage unit provided in advance, a document search step in which the document search unit searches a plurality of documents from the document storage unit, and the search result A search in which statistical information generation means generates and stores search result statistical information in which the number of appearances of the attribute value in the plurality of documents is recorded for each attribute value included in at least one of the searched plurality of documents. The result statistical information generation step and the fitness level calculation means, for each attribute value included in at least one of the searched plurality of documents, The number of appearances of the attribute value in the statistical information and the search result statistical information for calculating the degree of suitability indicating the degree of suitability as a label that is a character string representing a plurality of documents forming a part of a document The fitness level calculating step for calculating the fitness level of the attribute value using the used calculation formula, and the label information generating means are composed of attribute values included in at least one of the retrieved plurality of documents. A part of the plurality of attribute values is selected by selecting an attribute value corresponding to the fitness level from the higher fitness level corresponding to each of the attribute values as long as the fitness level satisfies a preset condition. A plurality of attribute values forming a label, each of the selected attribute values as a label, a label information generating step for generating label information including the plurality of labels, and the cluster information generating means, For each label included in the label information, document identification information and titles indicating documents that include the character string that is the label and that are any of the plurality of retrieved documents are included for the number of the documents. A cluster information generating step for generating cluster information that is cluster information and includes the label, and each of the plurality of documents in which the cluster information changing means includes document identification information included in the cluster information for each of the cluster information A document that generates a cluster vector that is a vector sum of the document vectors corresponding to the document vector, includes document identification information that is not document identification information included in the cluster information, and corresponds to a document stored in the document storage unit Calculate the cosine scale of the vector and the cluster vector, and the cosine scale exceeds the preset threshold. Then, the cluster information changing step for including the document identification information and the title of the document in the cluster information, and the document display control means displays each label included in the label information, and one corresponding label is selected. If the document identification information and title combination included in the cluster information including the label are respectively displayed and one document identification information and title combination is selected, the document including the document identification information and title combination is selected. And a document display control step for reading and displaying the document from the document storage means .

  According to the present invention, a document is retrieved from a document storage unit that stores the document, and a fitness level when the attribute value included in the retrieved document is used as a document label is calculated. Select an attribute value that is smaller than the number of the attribute value as a label, store the label information indicating the selected label, read the stored label information, display the label with the label information, and the label indicates In this case, since the document including the label and also included in the retrieved document is read from the document storage means and displayed, it is not necessary to prepare the label in advance, and the number of labels can be reduced. Therefore, it is not necessary to change the settings according to the document update by the administrator and troublesome operations by the user.

  In addition, a document is retrieved from the document storage unit that stores the document, and a degree of conformity when the attribute value included in the retrieved document is used as a document label is calculated. A cluster that selects a smaller number of attribute values than the number as a label, stores label information indicating the selected label, and indicates a document that includes one of the selected labels and is also included in the retrieved document Information is generated, the similarity between the document indicated by the cluster information and the document not indicated by the cluster information is calculated, and if this similarity is high, the cluster information is changed and stored so that the latter document is indicated. When the label information is read out, the label is displayed according to the label information, the label is instructed, the existence of the document indicated by the changed cluster information is displayed, and the document is instructed. Since the document is read from the document storage means and displayed, the setting change according to the document update by the administrator and the complicated operation by the user are not required, and the document displayed by the instructions of the few labels displayed Since the number can be increased, the possibility that a desired document can be displayed increases.

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

[First embodiment]
FIG. 1 is a block diagram showing a device configuration of the first embodiment. From here, it demonstrates, showing the example when a document is made into a news article.

  The retrieval apparatus 1 is a server computer that retrieves a document, and corresponds to a label display type document retrieval apparatus that executes the label display type document retrieval method of the present invention. The search device 1 can communicate with a browser 2 of a client computer connected via a network (not shown).

  The browser 2 includes a keyword input unit 21 for inputting a keyword via an input device such as a keyboard and a mouse, and a CRT (Cathode Ray Tube), an LCD (Liquid Crystal Display), or the like (not shown) for a document searched by the keyword. And a document display control unit 22 to be displayed on the display device.

  Before displaying the retrieved document, the browser 2 displays the attribute value included in the document as a label and classified by the attribute name of the attribute value. When any label is specified by clicking, the title of the document including the label (attribute value) is displayed, and when any title is specified, the document with the title is displayed. It is like that.

  The search device 1 includes a communication unit 101 that receives a keyword from the browser 2 and transmits the searched document to the browser 2, and a request processing unit 102 that controls document search using the keyword given from the communication unit 101.

  In addition, the search device 1 includes a setting file 103 in which information used for document search is written. In the setting file 103, attribute names “genre” and “organization” used in the search are written. In the setting file 103, the number of labels to be selected is written. The setting file 103 stores parameters α, β, and γ that are used when determining the fitness as the attribute name to be displayed (referred to as attribute name fitness). Various threshold values and the like are written in the setting file 103.

  In addition, the search device 1 includes an attribute value classified as one of the attribute names written in the setting file 103, but no attribute name (tag) is given to the attribute value. A document is input, and a news article manager or the like inputs a document by generating a document by so-called manual tagging, and an untagged document, and the attribute value is automatically tagged. , A document generation unit 105 that generates a tagged document (also simply referred to as a document), and a document database that stores the documents generated by the document generation unit 104 and the document generation unit 105 (hereinafter, the database is referred to as DB). (Abbreviated) 106.

  In addition, the search device 1 includes a normalization unit 107 that normalizes attribute values included in documents stored in the document DB 106.

  Further, the search device 1 associates a word (which may be an attribute value) included in the document stored in the document DB 106 with document identification information (hereinafter, identification information is referred to as ID) indicating the document including the word. The index generation unit 108 that generates the index 107 and the document search unit 109 that searches the document DB 106 based on the keyword and the index 107 are provided.

  Further, the search device 1 stores a first statistical information DB 110 that stores first statistical information generated for each attribute name of the setting file 103, and second statistical information generated for each first statistical information. A second statistical information DB 111 and a statistical processing unit 112 that generates first statistical information and second statistical information are provided.

  The search device 1 also includes a label candidate selection unit 113 that selects a plurality of attribute values as label candidates for each attribute name of the setting file 103, and a fitness level (referred to as a label fitness level) when the label candidate is a document label. ), And a label determination unit 115 that determines a label based on the calculated label suitability.

  In addition, the search device 1 includes a cluster information generation unit 116 that generates cluster information for each determined label. In the present embodiment, a cluster means one or more searched documents including one label.

[Pre-search process]
Next, processing performed by the search device 1 before searching will be described.

  FIG. 2 is a flowchart showing processing performed by the search device 1 before search.

  The document generation unit 104 receives, for example, an untagged document as shown in FIG. 3, and further, for example, designates that “International Atomic Energy Agency has determined ***” as a title, or has an attribute value “International Atomic Energy Agency”. 4 is classified into the attribute name “organization” and the attribute value “economic” is classified into the attribute name “genre”, as shown in FIG. A document is generated by giving the contents and a document ID such as “001”, for example, and stored in the document DB 106 (S101).

  On the other hand, when an untagged document is input and there is a title designation or the like, the document generation unit 105 generates a document by automatically attaching a tag to the attribute value, and further assigns a document ID. It is stored in the document DB 106 (S101). The process of automatically attaching a tag will be described later in detail.

  By such processing, a large number of documents are stored in the document DB 106.

  Next, the normalization unit 107 normalizes attribute values included in the document stored in the document DB 106 (S103). Normalization means, for example, conversion of an attribute value “IAEA” expressed by an abbreviation into an attribute value “International Atomic Energy Agency” expressed in Japanese that is not abbreviated.

  In other words, the normalization unit 107 detects attribute values having synonyms with different expressions while having the same meaning in the document, and makes them the same expression.

  Although there are several methods for detecting synonyms, a method using the co-occurrence pattern shown in FIG. 5 can be adopted.

  By such processing, the attribute value of the document in the document DB 106 is normalized.

  Next, the index generation unit 108 generates an index 107 in which a word included in a document stored in the document DB 106 is associated with a document ID of a document including the word (S105).

  As shown in FIG. 6, in the index 107, for example, the document ID “001” of the document including the word is associated with the word “nuclear power”.

  Next, the statistical processing unit 112 generates first statistical information for each attribute name of the setting file 103 based on the document DB 106 and stores the first statistical information in the first statistical information DB 110 (S107).

  As shown in FIGS. 7A and 7B, one attribute name is assigned to one first statistical information.

  One piece of first statistical information is information including one or more pieces of information in which a document ID is associated with an attribute value included in a document with this ID and classified as an attribute name.

  FIG. 7A shows that, for example, the document with the document ID “001” includes an attribute value “economic” or the like classified into the attribute name “genre”. FIG. 7B shows that the document with the document ID “001” includes an attribute name “International Atomic Energy Agency” classified into the attribute name “organization”.

  Next, the statistical processing unit 112 generates second statistical information for each first statistical information and stores it in the second statistical information DB 111 (S109).

  As shown in FIGS. 8A and 8B, one second statistical information attribute name is assigned to one second statistical information.

  One piece of second statistical information is information including one or more items in which attribute values classified into attribute names are associated with the number of appearances of the attribute value in the first statistical information DB 110.

  FIG. 8A shows that the number of appearances of the attribute value “economy” classified into the attribute name “genre” is 100, for example. FIG. 8B shows that the number of appearances of the attribute value “International Atomic Energy Agency” classified into the attribute name “organization” is 70 times.

  The second statistical information may be generated by detecting the correspondence between the attribute value and the document ID in the first statistical information and counting up the number of appearances for each detection.

  The second statistical information may be composed of the attribute value itself and the document ID of the document in which the attribute value appears. Moreover, you may comprise 2nd statistical information by the co-occurrence frequency between each attribute value. The co-occurrence frequency at this time can be defined as co-occurring attribute values appearing in the same document. When a document is automatically generated from an untagged document, it can be defined that attribute values included in the same sentence or paragraph in the untagged document co-occur. Further, the co-occurrence relationship is not expressed by a binary value of [0, 1] depending on the proximity in the document, but a large value may be given between attribute values that appear closer together like the co-occurrence. Good.

  In this way, the document search becomes possible after the processing up to S109 is completed. However, when the document in the document DB 106 is updated, added, or deleted, the attribute value normalization, the index 107, the first statistical DB 110, The second statistics DB 111 and the like are updated.

[Search process]
Next, a search process performed by the search device 1 will be described.

  For example, when the keyword “nuclear power” is input by the user, the keyword input unit 21 transmits the keyword “nuclear power” to the communication unit 101 of the search device 1.

  FIG. 9 is a flowchart of processing performed by the search device 1 that has received the keyword.

  First, the communication unit 101 gives the transmitted keyword “nuclear power” to the request processing unit 102, and the request processing unit 102 gives the keyword to the document search unit 109. The document search unit 109 searches for the document ID associated with the keyword “nuclear power” in the index 107 and returns them to the request processing unit 102 (S201: document search).

  The request processing unit 102 gives the document ID to the label candidate selection unit 113 (S203).

  The label candidate selection unit 113 generates first search result statistical information for each attribute name of the setting file 103 based on the first statistical information DB 110 and the searched document ID, and temporarily stores the first search result statistical information (S205).

  As shown in FIG. 10, one attribute name is assigned to one first search result statistical information.

  One first search result statistical information is a document ID of a document including the attribute value for each attribute value included in the first statistical information, and a document ID included in the searched document ID. It is a correspondence.

  Next, the label candidate selection unit 113 generates and temporarily stores second search result statistical information for each attribute name based on the first search result statistical information (S207).

  As shown in FIG. 11, one attribute name is assigned to one second search result statistical information.

  One second search result statistical information is obtained by associating each attribute value of one first search result statistical information with the number of document IDs associated with the attribute value as the number of appearances.

  Next, the label candidate selection unit 113 generates third statistical information for each second statistical information based on the second search result statistical information to which the same attribute name as the second statistical information is assigned (S209).

  As shown in FIG. 12, one piece of third statistical information is statistical information including one or more rows included in one piece of second statistical information, and the attribute value of the row corresponds to the second search result statistical information. It is also included in the line.

  Next, the label suitability calculation unit 114 generates and temporarily stores label suitability information for each second search result statistical information based on the second search result statistical information, the third statistical information, and the retrieved document ID. (S211).

  As shown in FIG. 13, one attribute name is assigned to one label suitability information.

  One label suitability information is obtained by associating a label suitability with each attribute value included in one second search result statistical information.

  The label suitability is calculated as follows, for example.

The number of applications corresponding to one attribute value in the second search result statistical information is set as h, the number of applications corresponding to the attribute value in the third statistical information is set as d, and the number of retrieved document IDs is set as | H | The label suitability is calculated by equation (1).

  In the equation (1), h / d represents the completeness of the attribute value in the retrieved document, and | H | / h represents the rarity of the attribute value in the retrieved document.

  Also, in equation (1), h / | H | is substituted for h in the first term, and d / | D | (| D | is the number of documents including the attribute value) instead of d in the first term. Also good.

  Next, the label determination unit 115 temporarily stores the label information obtained by subtracting the combination of the attribute value and the label compatibility from the label compatibility information (S213). The label information is generated and stored for each label suitability information. The attribute value in the label information is a label because it is a label of the document.

  As shown in FIG. 14, the label information is obtained by associating the label suitability with the label. By selecting a label from the label suitability higher in the label suitability information, the label information in the label information The number of sets of label suitability is smaller than the number of sets of attribute values and label suitability in the label suitability information.

  FIG. 15 is a flowchart of label selection performed by the label determination unit 115.

  The label determining unit 115 selects the number of labels written in the setting file 103 from the one with the highest label matching degree (S301). Next, it is determined whether or not to additionally select a label whose label suitability is the next point (S303).

Specifically, C (n) is the lowest selected label fitness level, C (n + 1) is the label fitness level one above, and C (n-1) is the label fitness level of the next point. When 2) is established, the label of the next point is additionally selected (S305), and the process returns to S303.

  However, e is a threshold value written in the setting file 103 or the like.

  In other words, the inclination of the value is evaluated, and a determination is made by applying the concept of using the boundary where the inclination exceeds a certain threshold.

  By this method, it is possible to prevent leakage from the selection of the label even though the label matching degree is close. That is, it can be made not to select only when there is a difference in the label suitability.

  In S301, a label may be selected by comparison with a threshold value of the label suitability written in the setting file 103 or the like.

  Next, the label determination unit 115 generates attribute name fitness information based on the label information and temporarily stores it (S215).

  As shown in FIG. 16, the attribute name suitability information indicates the attribute name suitability for each attribute name.

  For example, the attribute name suitability for the attribute name “genre” is calculated as follows.

  First, the number dl of documents including any label in the “genre” label information is obtained from the first search result statistical information of “genre”. At this time, one document including a plurality of labels is calculated as 1.

And comprehensiveness S1 is calculated | required by Formula (3).

  Here, dr is the number of retrieved document IDs.

  The greater the S1, the greater the extent to which the search result is covered by the label.

Next, the degree of overlap and the clarity of classification S <b> 2 are obtained by Expression (4).

Here, dr is the number of retrieved document IDs, dl _i is the number of documents including the i-th label l _i in the “genre” label information, and the second search result statistical information of “genre” It is obtained from.

  The larger S2 is, the greater the degree to which the search result is clearly classified by the label.

Next, the uniformity S3 of the classification is obtained from Equation (5). Here, S3 is obtained by calculating an average entropy of a cluster to be described later.

Here, dr is the number of retrieved document IDs, and dl _i is the number of documents including the i-th label l _i in the label information of “genre”. dl _i can be obtained from the second search result statistical information.

  The larger this S3, the greater the degree to which the search results are uniformly classified by the label.

Next, the attribute name suitability S is obtained by Expression (6).

  Here, α, β, and γ are parameters written in the setting file 103.

  Next, the request processing unit 102 reads the second search result statistical information, the label information, and the attribute name fitness information, and provides the label information to the cluster information generation unit 116.

  The cluster information generation unit 116 generates cluster information for each label included in the label information and temporarily stores it (S217).

  As shown in FIG. 17, the cluster information corresponds to each label included in the label information, the document ID of the document including the label and also included in the retrieved document ID, and the title of the document. It is attached.

  Next, the request processing unit 102 reads out all the second search result statistical information, label information, attribute name fitness information, and cluster information and gives them to the communication unit 101, and the communication unit 101 transmits them to the browser 2. (S219).

  FIG. 18 is a flowchart of processing performed by the browser 2 to which these pieces of information are transmitted.

  As shown in FIG. 19, the document display control unit 22 of the browser 2 displays document IDs and titles included in all cluster information (S401), and further displays labels included in the label information (S403). At this time, since the number of displayed labels is reduced depending on the degree of fitness, the user can easily instruct the label.

  Then, for the convenience of the user, for example, labels are displayed together for each attribute name. Further, the label of the attribute name having a high attribute name matching degree in the attribute name matching degree information is displayed so as to be easier to see. In addition, for labels included in one label information, an associated label having a high degree of label matching is displayed so as to be easier to see. In addition, the number of document IDs associated with the second search result statistical information is displayed in association with the label.

  Then, when one label is instructed by the user (S405), the document display control unit 22 deletes the displayed document ID and title, and is included in the cluster information including the label as shown in FIG. The document ID and title are displayed (S407).

 Then, when the document ID is instructed by the user (S409), the document display control unit 22 transmits the document ID to the communication unit 101 of the search device 1 (S411). In practice, the document ID can be designated by clicking the position of the document ID and the title.

  As shown in FIG. 20, in the present embodiment, after the label instruction, the number of document IDs and titles is smaller than before the instruction, so the user can easily instruct.

  The communication unit 101 of the search device 1 gives the transmitted document ID to the request processing unit 102. The request processing unit 102 gives the given document ID to the document search unit 109. The document search unit 109 reads the document with the given document ID and returns it to the request processing unit 102.

  The request processing unit 102 gives the returned document to the communication unit 101, and the communication unit 101 transmits it to the browser 2.

  The document display control unit 22 of the browser 2 displays the transmitted document.

  As described above, according to the search device 1 of the first embodiment, the document search unit 109, which is a document search unit, searches the document DB 106, which is a document storage unit that stores documents, and searches the document. The label suitability calculation unit 114 that constitutes the label selection unit calculates the suitability when the included attribute value is used as a document label, and the label determination unit 115 that constitutes the label selection unit calculates from the higher suitability The number of attribute values smaller than the number of the attribute values is selected as a label, label information indicating the selected label is stored, the request processing unit 102 constituting the document display control unit reads out the label information and the label When a label is displayed by transmitting information to the browser 2 and the label is designated by the user, the sentence including this label and also included in the retrieved document Is retrieved from the document DB 106 by the document search unit 109 constituting the document display control means, and transmitted to the browser 2 for display by the request processing unit 102 constituting the document display control means, so that it is not necessary to prepare a label in advance. In addition, the number of labels can be reduced, and as a result, a setting change according to the update of the document by the administrator and a complicated operation by the user become unnecessary.

  Further, since control is performed so that selection and display of labels is performed for each attribute name, it is possible to perform display for each attribute name by reducing the number of labels.

  In addition, by storing the statistical information, it is not necessary to generate the statistical information at the time of searching, so that quick processing can be performed at the time of searching.

  Moreover, since the normalization part 107 normalizes an attribute value, the normalized label can be displayed.

[Second Embodiment]
Next, a second embodiment to which the present invention is applied will be described. Here, the same reference numerals and the same step numbers are assigned to the same components and processes as those in the first embodiment.

  FIG. 21 is a block diagram illustrating an apparatus configuration according to the second embodiment.

  The search device 10 generates a document vector that includes the number of attribute values (hereinafter referred to as vector elements) included in one document stored in the document DB 106 for each attribute value in addition to the components of the search device 1 A vector generation unit 117 and a document vector DB 118 that stores the generated document vectors are provided. The vector element may be binary information corresponding to the presence or absence of an attribute value.

  In addition, the search device 10 calculates a similarity between the cluster vector generation unit 119 that generates a cluster vector including vector elements for each attribute value included in one cluster information, and the content of the document that is not included in the cluster. And a cluster expansion unit 121 that increases the document ID of the cluster information based on the calculated similarity.

[Pre-search process]
FIG. 22 is a flowchart illustrating processing performed by the search device 10 before search.

  When the processing of S109 ends, the document vector generation unit 117 generates a document vector for each document stored in the document DB 106, and stores this in the document vector DB 118 (S111).

  As shown in FIG. 23, one document vector includes the number of attribute values (referred to as vector elements) included in one document for each attribute value.

[Search process]
FIG. 24 is a flowchart of processing performed by the search device 10 to which a keyword has been transmitted.

  When the processing of S217 is completed, the cluster vector generation unit 119 generates a cluster vector for each cluster information in response to a request from the request processing unit 102 (S2181).

  As shown in FIG. 25, one cluster vector is composed of vector elements for each attribute value included in one cluster.

The cluster vector can be obtained by Expression (7).

  That is, a cluster vector is generated by obtaining a vector sum of documents included in the cluster.

  Then, S2183 and 2185 below are performed for all the combinations of clusters and documents not included in the clusters.

  First, the similarity calculation unit 120 calculates the similarity between the clusters and the contents of the document not included in the clusters (S2183). The similarity can be, for example, a cosine measure between the cluster vector obtained in S2181 and the document vector of a document not included in the cluster.

  Then, the cluster expansion unit 121 adds a document ID and a title to the cluster information based on the calculated similarity (S2185). For example, if the similarity calculated in S2183 exceeds the threshold value written in the setting file 103, the corresponding document ID and title are added to the cluster information.

  After this, the processes after S219 in the first embodiment are performed.

  As described above, according to the search device 10 of the second embodiment, the cluster information generation unit 116 as cluster information generation means includes a cluster that includes a label and indicates a document included in the searched document. If the similarity calculation unit 120 that generates information and constitutes the cluster information changing unit calculates the similarity between the document indicated by the cluster information and the document not indicated by the cluster information, and the similarity is high, the cluster information Since the cluster expansion unit 121 constituting the changing unit changes the cluster information so that the latter document is displayed, setting change according to the update of the document by the administrator and a complicated operation by the user become unnecessary. Since the number of documents displayed can be increased according to the instruction of the label displayed less, the possibility that a desired document can be displayed is increased.

  Next, the document generation unit 105 will be described. Here, a general document that is not limited to a news article will be described as an example.

  26 is a block diagram showing a configuration example of a specific expression extraction rule generation system applied to the document generation unit 105 and a specific expression extraction device provided with the specific expression extraction rule generation system. FIG. 27 shows a specific expression extraction rule generation system in FIG. 2 is a block diagram illustrating a hardware configuration example of a specific expression extraction device. FIG. The document generation unit 105 includes blocks excluding the new document A11 corresponding to the untagged document in FIG. 26, and generates a tagged document using the specific expression included in the specific expression list A13 as an attribute value. This eliminates the need to give attribute values to untagged documents.

  In FIG. 27, 1021 is a display device such as a CRT or LCD, 1022 is an input device such as a keyboard or mouse, 1023 is an external storage device such as an HDD (Hard Disk Drive), and 1024 is a CPU (Central Processing Unit) 1024a. And an information processing device 1025 having a main memory 1024b and the like and performing computer processing by a storage program method, 1025 is a CD-ROM (Compact Disc-Read Only Memory) or DVD (Digital Video Disc / Digital Versatile Disc) on which a program and data are recorded ) And the like, 1026 is a drive device for reading the program and data recorded on the optical disc 1025, and 1027 is a communication device such as a LAN (Local Area Network) card or a modem.

  The program and data stored in the optical disk 1025 are installed in the external storage device 1023 by the information processing device 1024 via the drive device 1026, and then read from the external storage device 1023 into the main memory 1024b and processed by the CPU 1024a. In the apparatus 1024, a specific expression extraction rule generation system shown in FIG. 26 and a specific expression extraction apparatus including the system are configured.

  26, the training document A1, the correct answer list A2, the specific expression extraction rule group A5, the improved specific expression extraction rule group A5a, the training record A7, and the new document A11 corresponding to the untagged document. , And the extracted list of specific expressions A13 are stored in the external storage device 1023 or the main memory 1024b in FIG. 27, and the morphological analysis / part-of-speech character type assigning unit A3, the restriction generating unit A4, and the training Each of rule application unit A6, rule evaluation unit A8, rule deletion unit A9, rule refinement unit A10, and enforcement rule application unit A12 is configured in information processing apparatus 1024 based on the program stored in CD-ROM 1025 in FIG. Is done.

  The morphological analysis / part-of-speech character type assigning unit A3, the regulation generating unit A4, the training rule applying unit A6, the rule evaluating unit A8, the rule deleting unit A9, and the rule refining unit A10 each constitute a specific expression extraction rule generating system. ing.

  The morphological analysis / part-of-speech character type assigning unit A3 divides the training document A1 into words, and adds information on the part-of-speech name and constituent character type to each word.

  The rule generation unit A4 matches the word string obtained by the processing of the morphological analysis / part-of-speech character type adding unit A3 with the data of the specific expression to be extracted given in the correct answer list A2, and extracts the word string constituting each specific expression, A rule is generated by generalizing this.

  The result is recorded in the external storage device 1023 in FIG. 27 as a specific expression extraction rule group A5.

  The training rule application unit A6 applies the specific expression extraction rule group A5 obtained from the generation result of the rule generation unit A4 to the training document A1. The result is recorded as training record A7 in the external storage device 1023 in FIG.

  The rule evaluation unit A8 evaluates each rule based on the training record A7. The rule deletion unit A9 deletes a rule with a poor result based on the evaluation result of the rule evaluation unit A8.

The rule refining unit A10 refines the rules so that the grades are improved.

  The implementation rule application unit A12 applies the specific expression extraction rule group A5 improved in this way (the improved specific expression extraction rule group A5a) to the actual new document A11 to obtain the specific expression list A13.

  Each of the training rule application unit A6 and the implementation rule application unit A12 applies a group of rules to a document to extract a specific expression, and the processing content is almost the same. It is also possible to serve as. However, the implementation rule application unit A12 does not need to leave the training record A7, but differs in that it needs to select a final candidate.

  First, the operation of the implementation rule application unit A12, that is, the specific expression extraction apparatus using the specific expression extraction rule group A5 generated and improved by the specific expression extraction rule generation system of this example, and the improved specific expression extraction rule group A5a The operation will be described.

  The implementation rule application unit A12 applies the improved specific expression extraction rule group A5a to the new document A11 from which a specific expression is to be extracted, extracts the specific expressions included in the document, and outputs the specific expression list A13. To do.

  For example, if there is a new document A11 “In the Taro Tanaka Prize Selection Committee,” the candidate names of “Tanaka”, “Taro”, and “Taro Tanaka” are listed as specific expressions in this document. A candidate for the artifact name “Taro Tanaka Award” and a candidate for the organization name “Taro Tanaka Award Selection Committee” are considered, but in general, only the longest “Taro Tanaka Award Selection Committee” Is often extracted and output as a specific expression. In this case, other candidates (specific expressions) such as “Tanaka” and “Taro” that overlap this should not be output.

  Such a wrinkle interpolation relationship can be reduced to a competitive relationship caused by overlap and a suppression relationship based on the priority of each candidate. In other words, the “Taro Tanaka Award Selection Committee” competes with other candidates such as “Tanaka”, but the long “Taro Tanaka Award Selection Committee” has higher priority and suppresses other short candidates. You can think that you are.

  In this example, the implementation rule application unit A12 first applies all the rules to the document based on this concept, so that a set of candidates for all the specific expressions (“Tanaka”, “Taro”, “Tanaka” Taro ”,“ Taro Tanaka Award ”,“ Taro Tanaka Award Selection Committee ”, etc.). Next, among these candidates, the longest of the same specific expressions ("Tanaka" in each of the above candidates) appears first ("Taro Tanaka Prize Selection Committee" in each of the above candidates) Is output.

  When one candidate is output in this way, all other candidates competing with this candidate (“Tanaka”, “Tanaka Taro”, “Tanaka Taro Award”) are deleted from the candidate set. By repeating this operation until the candidate set becomes empty, a list A13 of specific expressions is obtained.

However, if there is a plurality of candidates having the same length, it is difficult to make a determination only by determining the selection from each competing candidate by paying attention only to the length.

  For example, “White House” may be considered as a place name or an organization name, and therefore, a candidate having the same character string “White House” as a place name and a candidate as an organization name can be considered.

  Therefore, a priority order for extraction is set between the two candidates.

  For example, considering the words before and after that, if it is "in a park near the White House ...", the place name is likely; if it is "according to the White House ...", the name of the organization Probability is high. Also, for example, in consideration of the frequency of appearance, if “white house” appears as a place name once in the training document A1, and if it appears as an organization name 20 times, the organization The person who judges it as a name is more likely to answer correctly.

  In this example, each rule in the improved specific expression extraction rule group A5a is given a priority in consideration of these conditions.

  The implementation rule application unit A12 calculates the priority of each candidate by combining such priority and the length of the above-described specific expression. There are various variations of this priority setting, but as mentioned above, the one with the highest priority among the ones with the earliest start position and the slowest end position is selected. Is clear. In other words, the priority definition of candidates is based on the following definition.

  (2) If the start position of candidate A is earlier (smaller as a number) than the start position of candidate B, candidate A has priority.

  (2) If the start position of candidate A and the start position of candidate B are the same, a candidate with a late end position (larger number) is given priority.

  (2) If the starting position and the ending position of both candidates are exactly the same, a candidate having a high priority u given in advance by a rule is given priority.

  In the specific expression extraction rule generation system of this example, a specific expression extraction rule group A5 and an improved specific expression extraction rule group A5a that facilitate processing by the implementation rule application unit A12 are generated. In the following, the operation of each part constituting the specific expression extraction rule generation system related to the rule generation processing taking such priority relationships into account will be described.

  First, the morphological analysis / part of speech character type assigning unit A3 divides the document into word strings. Typically, it has a morphological analysis function, and a given document such as a training document A1 or a new document A11 is divided into words, and the part of speech name and the type of characters constituting the word (constituent character type information) Create a data structure with a list and create a list.

  For example, if there is a sentence “President Nakano is ...”, “Tokyo” is a proper noun, “Iron” is a common noun, “no” is a particle, “Nakano” is a proper noun, The result is that “President” is a common noun, and “Ha” is a particle.

“Tokyo” is composed of a plurality of kanji characters, and “no” is hiragana. Therefore, the morphological analysis / part-of-speech character type assigning unit A3 outputs, for this sentence, a list having the following data structure, for example. [(Tokyo, multiple kanji, proper noun), (steel, multiple kanji, common noun), (no, hiragana, particle), ...]
On the other hand, the correct answer list A2 lists what kind of specific expressions are included in which position in the training document A1. The correct answer list A2 prepared in advance corresponding to the document A1 includes, for example, the following data.

  In this list, the first line is a unique expression whose type is “organization name” whose “position from the 0th character to the 3rd character” of this document is “Tokyo Steel,” and the next line is “5 characters. This indicates that the “position of the sixth character from the eye” is a unique expression of “person name” of “Nakano” as its type. Thus, in the correct answer list A2 of this example, the position of the specific expression is abbreviated as a pair of numbers indicating the start position and end position of each specific expression.

  The rule generation unit A4 matches such a correct list A2 with the word string output from the morphological analysis / part-of-speech character type assigning unit A3, converts the specific expression into a variable, etc. Generate an extraction rule. anytag (3) <-<@ (organization name, 21), word (_, multiple kanji, proper noun), word (steel, multiple kanji, common noun),> @ (organization name). This rule (rule) is a rule to which the number “21” is assigned, and there is an arbitrary (variable) kanji proper noun (“word (_, plural kanji, proper noun)”), followed by If the word is a common noun of multiple kanji called “steel” (“word (steel, multiple kanji, common noun)”), the two words can be considered as candidates for the proper expression of “organization name” It is a rule.

  Generation of such a rule (rule) can be expressed more generally as follows. First, it is assumed that the unique expression is made up of N + 1 words [(w0, c0, p0),..., (Wi, ci, pi), ..., (wN, cN, pN)]. Here, w i is a word (such as “steel” or “Nakano”), c i is a constituent character type (such as “multiple kanji” or “number”), and p i is a part of speech name (such as “proper noun” or “common noun”). .

  Actually, some words before and after are also important clues for judging whether they are proper expressions, so it is common to consider them, but here they are simplified and included in the proper expressions Think only of words.

  Next, a rule is generated from such a word string by using an existing generalization technique such as minimum generalization. However, in this example, it is easily generated as follows.

That is, a specific word string [(w0, c0, p0),..., (Wi, ci, pi),..., (WN, cN, pN] constituting a specific expression included in the training document A1. )] Is applied to the list [(w0 ′, c0 ′, p0 ′),..., (Wi ′, ci ′, pi ′),. wN ′, cN ′, pN ′)], and the following rule is created. anytag (u) <-<@ (t + df, k), word (w0 ', c0', p0 '), ..., (wi', ci ', pi'), ..., word (wN ' , CN ′, pN ′),> @ (t−dt).
Here, “t” represents the type of specific expression (for example, “organization name”).

  “+ Df” represents the number of characters to which the start position of this specific expression is shifted to the right, and is a non-negative integer less than the number of characters of the first word. Further, “−dt” represents how many characters the end position of the specific expression is shifted to the left, and is a non-negative integer less than the number of characters of the last word.

  For example, there is a training document A1 “in Atsugi City ...”, and according to the correct answer list A2, “Atsugi City” is a place name, but the morphological analysis / part-of-speech character type assigning part A3 In the morphological analysis, when words are divided into “Atsugi”, “city”, and “de”, the word string constituting the proper expression is [(Atsugi, multiple kanji, proper noun), (city, multiple Kanji, common noun)], and the last character (“inside”) is extra. Therefore, “dt = 1” is set in order to shift the end position to the left by one character. Since the start position is not shifted, “df = 0”. In the above rule (rule), “k” is a number given to this rule, and “u” is the priority of this rule. The data (wi ', ci', pi ') including each variable has the following empirical rules for the data (wi, ci, pi) corresponding to the specific specific expression included in the training document A1, It is obtained by examining in order from the top and applying the first fit.

  (2) When “i” is “0” or “N” and includes a boundary of proper expression (df> 0 or dt> 0), these are not converted into variables. As the rules (rules) “df” and “dt”, the values for the original specific expressions are used as they are.

  ■ In the case of numbers, “wi” is converted into a variable.

  ■ In the case of proper nouns, variable “wi”.

  ■ If it is the last word in the list or a functional word such as a symbol, single kanji, suffix, prefix, or particle, it is not converted into a variable.

  ■ Otherwise, “wi” is made variable.

  By applying the above processing to each specific expression, the specific expression extraction rule group A5 can be automatically generated.

  Further, as the priority (u) of each rule, for example, the “total number of times” in which the unique expression that is the basis of the rule appears in the correct answer list is adopted. As a result, a rule with a small number of correct answers (in the above example, “White House” as a place name) may suppress a rule with a large number of correct answers (“White House” as an organization name) without a valid reason. can avoid.

  Each rule (specific expression extraction rule group A5) thus obtained by the rule generation unit A4 is applied to the word string of the training document A1 in the training rule application unit A6, thereby obtaining the training record A7. That is, in the training rule application unit A6, the positions where each rule matches are sequentially examined from the beginning to the end of the training document A1, and if they match, it is repeatedly added as a candidate to the training record A7.

  Specifically, in the training record A7, the rule number (k), the matched position, and the uniqueness are set so that the competition output and the suppression relationship between the candidates can be compared later and final output can be performed. Data such as expression type (t) is recorded.

  Such processing by the training rule application unit A6 is performed on all the rules of the specific expression extraction rule group A5 to create a training record A7.

If bottom-up syntax analysis is used, it is possible to efficiently obtain application results of a plurality of rules at once.

  The rule evaluation unit A8 reads the training record A7 created in this way, and scores the results of each rule. Various criteria can be used for scoring, but it is easy to use evaluation based on the number and ratio of incorrect answers. However, strictly speaking, the number of incorrect answers for each rule depends on which rule is used in combination with each other. Therefore, at this point in time, which rule is to be adopted is not yet determined, an accurate number cannot be obtained. Therefore, the records of each rule (R) are considered classified as follows.

(◯) Candidate obtained by matching with the specific expression that is the basis of the rule R, that is, one that inevitably becomes a correct answer if not suppressed by another candidate (correct candidate specific expression).

(Δ) Another competing specific expression is registered in the correct answer list A2 and is suppressed thereto. In other words, since the output is suppressed when the proper expression is correct, the rule group with high accuracy has a high possibility of not lowering the grade (intermediate candidate specific expression).

(X) Other than that, that is, since there is no correct answer specific expression to be suppressed, a rule group with high accuracy is likely to output a wrong candidate and lower the grade (incorrect candidate candidate specific expression).

  The rule evaluation unit A8 counts the number of “○”, “△”, and “×” for each rule, and substitutes the number of “×” for the number of incorrect answers and the number of “○” for the number of correct answers. Adopt as. If all “Δ” are considered to be incorrect, it is better to avoid a rule that extracts a short unique expression such as “Tanaka”. Therefore, the rule evaluation unit A8 counts the number of incorrect answers by the following method.

  That is, the rule evaluation unit A8 reads the training record A7 in order from the front, the rule R is applied at the position L of the training document A1, and the type of the unique expression given by the rule R (the distinction between place names, personal names, etc.) ) Is T, and the pair of type T and position L is not included in the correct answer list A2 as a correct answer. If the candidate based on the rule R is superior in priority to the candidate corresponding to the correct answer, the number of incorrect answers for the rule R is increased by one. This is repeated until the end of training record A7 is reached.

  When the rule evaluation unit A8 counts the number of “O”, “Δ”, and “×” in each rule, the rule deletion unit A9 and the rule refinement unit A10 refer to this result to the specific expression extraction rule group A5. Make corrections.

  The rule deletion unit A9 deletes, for example, a rule in which the number of “x” exceeds the number of “◯” among the parent rules of the specific expression extraction rule group A5. The rule refining unit A10 adds restriction information on the preceding and following words to a rule having a score of “X” or more than half of the number of “O” in the rules of the specific expression extraction rule group A5, Make the rules perform better.

  For example, when including two words before and after the specific expression, in each of the specific expressions extracted by the above rules and classified as “◯” or “×”, [(w−2, c− 2, p-2), (w-1, c-1, p-1), (w0, c0, p0), ..., (wN + 1, cN + 1, pN + 1), (wN + 2, cN + 2, pN + 2),] are considered for each. Therefore, for each unique expression, (w-2, c-2, p-2, w-1, c-1, p-1, wN + 1, cN + 1, pN + 1, wN + 2, cN + 2, pN + 2), this is typical if the proper expression classified as “○” is a positive example and the proper expression classified as “×” is a negative example. This is a problem of inductive learning, and the existing machine learning method can be used as it is.

  For example, by using learning based on a decision tree, it is possible to determine which property value of which word among several words before and after and which should be made variable. As a specific example, “10” specific expressions classified as “×” are extracted, and among them, “wX” is specified as the previous word (w−1) in “8” specific expressions. For example, a constraint condition {w−1 ′ ≠ wX} is added to the original rule as follows to constrain a specific expression having “wX” as the previous word (w−1) from being extracted. anytag (u) <-word (w-1 ', c-1', p-1 '), <@ (t + df, k), word (w0', c0 ', p0'), ..., ( wi ′, ci ′, pi ′),..., word (wN ′, cN ′, pN ′),> @ (t−dt), {w−1 ′ ≠ wX}. Since the rule obtained in this way is more restrictive than the original rule, it matches only the same part where the original rule matched. Therefore, even if it is not applied to the entire training document A1, if it is applied only to the matched part of the original rule remaining in the training record A7, the result of the new rule can be known.

  As described above, in this example, the rule can be improved almost independently of other rules. As described above, a rule (an improved specific expression extraction rule group A5a after improvement) is generated from the original rule (specific expression extraction rule group A5).

  FIG. 28 is a flowchart illustrating an example of a processing procedure of the specific expression extraction rule generation method.

This example shows processing operations of the morphological analysis / part-of-speech character type assigning unit A3, the rule generating unit A4, the training rule applying unit A6, and the rule evaluating unit A8 in the named entity extraction rule generating system in FIG. In the morphological analysis / part-of-speech character type assigning unit A3, the training document A1 is morphologically analyzed and divided into words (S1301), and information such as part-of-speech and character type is added to each word (S1302).

  Next, the rule generation unit A4 extracts a specific string of the correct answer list A2 and a word string composed of words in the vicinity thereof (S1303), and applies an empirical rule or the like to the correct word string to generate an extraction rule. (S1304), recorded as a specific expression extraction rule group A5. Then, in the training rule application unit A6, the extraction rule generated in this way is applied to the training document A1, and the resulting unique expression is recorded as a candidate (S1305). Further, the rule evaluation unit A8 obtains and classifies the correctness (◯, Δ, ×) of the unique expression extracted by each extraction rule, and scores the appropriateness of each extraction rule based on the correctness (S1306).

  As a result of the scoring, a rule group having a bad result (low degree of appropriateness) that cannot be corrected is deleted in the rule deletion unit A9 (S1307), and a rule group whose degree of appropriateness is increased by correction is rule refining part A10. In step S1308, the modification is made to make a new rule (S1308), which is recorded in the improved specific expression extraction rule group A5a. By repeating the processing from S1305, it is possible to generate a rule group with better results.

  FIG. 29 is a flowchart showing an example of processing operation of the named entity extraction apparatus in FIG. This example shows the processing operation for the new document A11 in the named entity extraction apparatus shown in FIG. 26. First, the morphological analysis / part of speech character type assigning unit A3 divides the new document A11 into words by performing morphological analysis. (S1401), information such as part of speech and character type is added to each word list (S1402).

  Next, the implementation rule application unit A12 applies each extraction rule of the improved specific expression extraction rule group A5a to each word list, and lists each specific expression as a candidate (S1403). On the other hand, the following priority control process is performed (S1404). That is, the highest priority candidate among the candidates is output (S1405), and the candidate competing with the output candidate is deleted (S1406).

  As described above with reference to FIG. 26 to FIG. 29, in the named expression extraction rule generation system and method of this example, first, the training document A1 prepared in advance is divided into words by morphological analysis, and the part-of-speech name. And information such as constituent character types are added to each word, a word string constituting a specific expression is extracted from the words thus obtained, and experience is obtained by referring to a correct answer list A2 prepared in advance corresponding to the training document A1. A large number of proper expression extraction rules are generated by a generalization means such as a rule or minimum generalization.

  Next, each of these extraction rules is independently applied to the training document A1, and a record of where the rule matches in the training document A1 is prepared. What is included in this record is a candidate for the specific expression output by the system for the training document A1. And when combining multiple rules, out of all candidates in the records corresponding to those rules, the final candidate column to be output is determined in consideration of the competition and priority. Select based on clear criteria.

  As a result, if there is a rule having a very high frequency or rate of incorrect answers in the training document A1, it is deleted. However, it can be seen where in the training document the rule is correct and where it is incorrect. Therefore, by comparing the word string before and after the correct answer part with the word string before and after the incorrect answer part and adding restrictions, it can be determined whether a rule that improves the result in the training document can be made. If it improves, add a rule with constraints.

  Thus, according to this example, when a training document including a specific expression and a correct answer list enumerating what kind of specific expression is included in which position in the document are given, the system Since the specific expression extraction rule is generated based on the correct answer, it is not necessary for the human to write down the extraction rule with great effort. Furthermore, the evaluation of each rule output for the training document A1 prepared in advance is obtained, and then each evaluation value when various rules are combined in various ways is easily calculated from the evaluation value of each rule. Can be calculated.

  As a result, it is possible to reduce the processing time required for trial and error when obtaining a combination of rules for obtaining good results. Moreover, since the improvement of such a specific expression extraction rule can be performed almost independently of other rules, it is easy to improve accuracy. In addition, the specific expression extraction apparatus of this example automatically extracts specific expressions from the new document A11 by applying the improved rules generated based on the training document and the correct answer list to the new document A11. In addition, if there is a partial overlap in the extracted plurality of specific expressions, the one with the earlier description start position in the document is preferentially extracted, and if the description start position is the same, the one with the later description end position is extracted. If there are specific expressions with the same expression but different types, extract those with higher priority given in advance to each rule used to extract each specific expression. Therefore, extraction can be limited to only proper proper expressions. In addition, it is not limited to the example demonstrated using FIGS. 26-29, Various implementation can be performed. For example, in this example, when a restriction is added to a rule, the restriction is provided based on the preceding and following words in the training document for candidate specific expressions, but the character type (kanji, katakana,...) Restrictions may be placed on nouns, verbs,. In this example, the optical disk 1025 is used as a recording medium, but an FD may be used as a recording medium. As for program installation, the program may be downloaded and installed via the communication device 1027 via a network.

  A computer program that causes the search apparatus 1 or 10 to execute the label display type document search method described in the first and second embodiments is a computer such as a semiconductor memory, a magnetic disk, an optical disk, a magneto-optical disk, or a magnetic tape. It can be widely distributed by storing it in a readable recording medium or transmitting it via a communication network such as the Internet.

It is a block diagram which shows the apparatus structure of 1st Embodiment. It is a flowchart which shows the process which the search device 1 performs before a search. It is a figure which shows an example of an untagged document. It is a figure which shows an example of a tagged document. It is a figure which shows the detection method of a synonym using a co-occurrence pattern. It is a figure which shows an example of an index. It is a figure which shows an example of 1st statistical information. It is a figure which shows an example of 2nd statistical information. It is a flowchart of the process which the search device 1 which transmitted the keyword performs. It is a figure which shows an example of 1st search result statistical information. It is a figure which shows an example of 2nd search result statistical information. It is a figure which shows an example of 3rd statistical information. It is a figure which shows an example of label compatibility information. It is a figure which shows an example of label information. It is a flowchart of the label selection which the label determination part 115 performs. It is a figure which shows an example of attribute name compatibility information. It is a figure which shows an example of cluster information. It is a flowchart of the process which the browser 2 performs. It is a figure which shows the example of a display before the label instruction | indication by the document display control part. It is a figure which shows the example of a display after the label instruction | indication by the document display control part 22. FIG. It is a block diagram which shows the apparatus structure of 2nd Embodiment. It is a flowchart which shows the process which the search device 10 performs before a search. It is a figure which shows an example of a document vector. It is a flowchart of the process which the search device 10 which transmitted the keyword performs. It is a figure which shows an example of a cluster vector. It is a block diagram which shows the structural example of the specific expression extraction rule production | generation system applied to the document production | generation part 105, and the specific expression extraction apparatus provided with the same. It is a block diagram which shows the hardware structural example of the specific expression extraction rule production | generation system in FIG. 26, and a specific expression extraction apparatus. It is a flowchart which shows the example of a process sequence of the specific expression extraction rule production | generation method. It is a flowchart which shows the processing operation example of the specific expression extraction apparatus in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1,10 ... Search apparatus 2 ... Browser 21 ... Keyword input part 22 ... Document display control part 101 ... Communication part 102 ... Request processing part 103 ... Setting file 104 ... Document generation part 105 ... Document generation part 106 ... Document DB
DESCRIPTION OF SYMBOLS 107 ... Index 107 ... Normalization part 108 ... Index production | generation part 109 ... Document search part 110 ... 1st statistics DB
111 ... 2nd statistics DB
DESCRIPTION OF SYMBOLS 112 ... Statistical processing part 113 ... Label candidate selection part 114 ... Label suitability calculation part 115 ... Label determination part 116 ... Cluster information generation part 117 ... Document vector generation part 118 ... Document vector DB
119: Cluster vector generation unit 120 ... Similarity calculation unit 121 ... Cluster expansion unit

Claims (10)

Document storage including a document including a text composed of a character string, further including a title of the document and document identification information indicating the document, and a plurality of documents in which attribute values that are predetermined character strings are included in the text Means,
A document vector generation unit that generates a document vector corresponding to each document stored in the document storage unit and includes the number of attribute values included in the document for each attribute value;
For each attribute value included in at least one of the plurality of documents stored in the document storage unit, statistical information recording the number of appearances of the attribute value in the plurality of documents is generated, and the storage unit provided in advance Statistical processing means to be stored;
Document retrieval means for retrieving a plurality of documents from the document storage means;
Search result statistical information generating means for generating and storing search result statistical information in which the number of occurrences of the attribute value in the plurality of documents is recorded for each attribute value included in at least one of the plurality of searched documents; ,
For each attribute value included in at least one of the plurality of retrieved documents, the degree to which the attribute value is suitable as a label that is a character string representing a plurality of documents forming part of the plurality of documents is indicated. A fitness calculation means for calculating the fitness of the attribute value using a calculation formula using the number of appearances in the statistical information of the attribute value and the search result statistical information for calculating the fitness; ,
The fitness level as long as the fitness level satisfies a preset condition from the higher fitness level corresponding to each of a plurality of attribute values including attribute values included in at least one of the retrieved plurality of documents. By selecting an attribute value corresponding to, a plurality of attribute values forming a part of the plurality of attribute values are selected, and the plurality of selected attribute values are used as labels, respectively, and label information including the plurality of labels is displayed. Label information generating means to generate;
For each label included in the label information, document identification information and titles indicating documents that include the character string that is the label and that are any of the plurality of retrieved documents are included for the number of the documents. Cluster information generating means for generating cluster information that is cluster information and includes the label;
For each cluster information, a cluster vector that is a vector sum of the document vectors corresponding to each of a plurality of documents including the document identification information included in the cluster information is generated, and the document identification information included in the cluster information A cosine measure between the document vector corresponding to the document stored in the document storage means and the cluster vector is calculated, and the cosine measure exceeds a preset threshold value. If there is, cluster information changing means for including the document identification information and title of the document in the cluster information,
Each label included in the label information is displayed, one of the labels is selected, and a set of document identification information and title included in the cluster information including the label is displayed, respectively. A label display type document retrieval apparatus comprising: a document display control unit that reads out and displays a document including the document identification information and the set of titles from the document storage unit when a set of titles is selected . The document stored in the document storage means is provided with a tag indicating that it is an attribute value with respect to an attribute value included in the document,
2. The label display type document retrieval apparatus according to claim 1, wherein the document display control means displays the labels for each of the tags attached to the labels. Attributes that detect a plurality of attribute values included in the document stored in the document storage means and have the same meaning and different expressions, and make the expression of the attribute values the same. 3. The label display type document retrieval apparatus according to claim 1, further comprising a value normalization unit. The document stored in the document storage means is a tag in which an attribute value is added to an attribute value included in the document,
A document generation unit that adds a tag to the attribute value as a character string that is a character string in a document having no tag and is equal to the specific expression in the list generated in advance, and stores the document in the document storage unit; 4. The label display type document retrieval device according to claim 1, wherein A label display type document search method performed by a label display type document search device,
The label display type document search apparatus includes a document storage unit, a document vector generation unit, a statistical processing unit, a document search unit, a search result statistical information generation unit, a fitness calculation unit, a label information generation unit, a cluster information generation unit, and cluster information. A change means and a document display control means,
The document storage means includes
A document including a text composed of a character string, further including a title of the document and document identification information indicating the document, and a plurality of documents including attribute values that are predetermined character strings included in the text,
The label display type document search method includes:
A document vector generation step in which the document vector generation unit generates a document vector corresponding to each document stored in the document storage unit and including the number of attribute values included in the document for each attribute value. When,
The statistical processing unit generates statistical information that records the number of appearances of the attribute value in the plurality of documents for each attribute value included in at least one of the plurality of documents stored in the document storage unit; A statistical processing step to be stored in a storage means provided in advance;
A document search step in which the document search means searches for a plurality of documents from the document storage means;
The search result statistical information generating means generates search result statistical information in which the number of appearances of the attribute value in the plurality of documents is recorded for each attribute value included in at least one of the plurality of searched documents. A search result statistical information generation step to be stored;
For each attribute value included in at least one of the searched plurality of documents, the fitness level calculating unit is a label that is a character string representing a plurality of documents that form part of the plurality of documents. The calculation value using the number of appearances in the statistical information and the search result statistical information of the attribute value to calculate the fitness indicating the degree of suitability as A fitness calculation step to calculate;
The label information generation means presets the fitness level from the higher fitness level corresponding to each of a plurality of attribute values consisting of attribute values included in at least one of the searched plurality of documents. As long as the condition is satisfied, by selecting an attribute value corresponding to the fitness level, a plurality of attribute values forming a part of the plurality of attribute values are selected, and the plurality of selected attribute values are used as labels, respectively. A label information generation step for generating label information including the label of
Document identification information and title indicating that the cluster information generating means is a document that includes a character string that is the label for each label included in the label information and is one of the plurality of searched documents. A cluster information generation step of generating cluster information including the label and the cluster information including the number of the documents,
The cluster information changing unit generates, for each cluster information, a cluster vector that is a vector sum of the document vectors corresponding to each of a plurality of documents including document identification information included in the cluster information, and the cluster information The cosine scale between the document vector corresponding to the document stored in the document storage means and the cluster vector is calculated, and the cosine scale is set in advance. A cluster information changing step for including the document identification information and title of the document in the cluster information,
The document display control means displays each label included in the label information, and one set of the label is selected, and a set of document identification information and a title included in cluster information including the label is displayed. A document display control step of reading out and displaying a document including the document identification information and title set from the document storage means when one set of the document identification information and title is selected;
Labeling type document retrieval method, which comprises a. The document stored in the document storage means is provided with a tag indicating that it is an attribute value with respect to an attribute value included in the document,
6. The label display type document search method according to claim 5, wherein the document display control means displays the labels for each of the tags attached to the labels. The label display type document search device further includes attribute value normalization means,
The attribute value normalization means detects a plurality of attribute values included in the document stored in the document storage means and having the same meaning and different expressions, and the plurality of attribute values 7. The label display type document search method according to claim 5 or 6 , further comprising an attribute value normalizing step for making each expression the same . The label display type document search device further includes a document generation unit,
The document stored in the document storage means is a tag in which an attribute value is added to an attribute value included in the document,
The document generation means assigns a tag to the attribute value as a character string that is a character string in a document with no tag and is equal to the specific expression in the list generated in advance, and stores the document in the document storage means The label display type document retrieval method according to claim 5 , further comprising a document generation step . A computer program for causing a computer to execute the label display type document retrieval method according to claim 5 . A computer-readable recording medium in which the computer program according to claim 9 is stored.

Images