JP2002324077A - Apparatus and method for document retrieval - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that it is difficult for a conventional apparatus for document retrieval to select a searching word to efficiently narrow a search. SOLUTION: An apparatus for document retrieval comprises a document retrieval part 2 to search documents, a document feature extracting part 4 to output a document vector group 5, a topic classifying part 6 to prepare a topic by classifying the group 5, a narrowing effect presumption part 10 to calculate narrowing effect indicator 11,a generating part 8 for choice of retrieving word for provision to select the retrieval word with high indicator 11 and output it as a candidate of retrieval word 12 for provision, a classifying result providing part 14 to provide the candidate 12 and indicator 11 topic-by topic and a document feature setting part 20 to change the group 5.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a document search apparatus capable of efficiently executing a document search by presenting a narrowing-down effect for each topic which is a unit larger than a document and facilitating selection of an additional search word. And a document search method.

[0002]

2. Description of the Related Art H that can be browsed using the Internet
When searching a TML document or an electronic document recorded and managed in a large-scale text database, the amount of document information obtained as a search result becomes extremely large, and it takes much time for a user to find a desired document. And labor are needed. For this purpose, the document information obtained as a search result is classified according to the content, and a search word candidate to be added for a narrow search is presented to the user so that the user can efficiently search for a target document. There is an increasing demand for a technology that assists a user in finding a desired document.

[0003] With respect to a technology that allows a user to efficiently search for a target document, Japanese Patent Application Laid-Open No. 9-231238, entitled "Method and Apparatus for Displaying Text Search Results" (hereinafter referred to as Document 1), and Japanese Patent Application Laid-Open No. 11-213000. There is a technique disclosed in Japanese Patent Laid-Open Publication No. 2000-214, "Interactive information search method and apparatus and storage medium storing an interactive information search program" (hereinafter referred to as Document 2).

The technique disclosed in Document 1 classifies search results into subjects by fuzzy clustering, and presents the subject-categorized categories together with a set of search words to a user. The technique disclosed in Document 2 classifies search results by clustering, presents the classified categories to a user together with a set of search words, and further classifies the user-specified categories into sub-categories to provide an interactive method. It enables search.

[0005] With regard to the technology for supporting a refined search,
Hayakawa et al., “Development of an interface for narrowing search results in WWW search service” (IPSJ, Human Interface Workshop (HI76-5), pp.2)
5, 1998, hereinafter referred to as Document 3).

[0006] The technique disclosed in Document 3 visualizes information on how much a search word can narrow down the number of search results in order to easily grasp the relationship between a search word and a set of search result documents. Provides an interface to present to the user. In addition, the results of narrowing down search terms are visualized using a document word matrix, and a matrix having words and documents in rows and columns, respectively, is directly presented to the user in a table form, and the weight of words for the documents is given. Is expressed by the brightness of the cells of the matrix so that the effect of narrowing down the additional search words can be listed.

[0007] Further, regarding the prior art described in Document 3, Japanese Patent Application Laid-Open No. H11-85764, entitled "Method and apparatus for statistically estimating the number of search results and a storage medium storing a program for statistically estimating the number of search results," , Reference 4). The technique disclosed in Document 4 is a method for statistically estimating the number of search results to know how the number of search results changes when a search word for further narrowing the search results is added in Document 3. Is disclosed.

FIG. 10 is a block diagram showing a conventional document retrieval apparatus, and is a block diagram of an estimating apparatus as an example of the embodiment disclosed in Reference 4. In FIG. 10, reference numeral 101 denotes an estimating apparatus for searching a document by full-text search; 102, a search term for searching the document; 103, a text database in which documents are recorded and managed; 104, output from the text database 103; Total search results for documents
5 is a document sample set randomly extracted from the search results and output from the text database 103;
Reference numeral 106 denotes a document word matrix generation unit that receives a document sample set 105 and counts the number of appearances of each word for each document.
Reference numeral 07 denotes a document word matrix generated by the document word matrix generation unit 106. Note that the document sample set 105 is not limited to 50, and can be appropriately set according to the situation.

[0009] In FIG. 10, reference numeral 108 denotes a search word candidate presentation unit that calculates word importance for the document word matrix 107, 109 denotes a search word candidate output from the search word candidate presentation unit 108, and 110 denotes a search word A search term selection unit 111 for outputting the candidate 109 to a monitor or the like;
09 is a selection signal input when selecting 09, 112 is an additional search word output from the search word selection unit 110 based on the selection signal 111, 113 is an additional search word based on the document word matrix 107 and the additional search word 112 The number estimation unit 112 for calculating the appearance rate of 112 is the estimated number obtained by multiplying the total number of search results 104 output from the number estimation unit 113 by the appearance rate. FIG. 11 is an explanatory diagram showing an example of a document word matrix 107 in a conventional document search device.

Next, the operation will be described. Search term 102
Is entered into the text database 103, the search term 1
02 is searched for documents recorded and managed in the text database 103 based on the 02. Text database 1
03, a document word matrix generation unit 106 extracts 50 items at random from the search results and sets the sample as a document sample set 105;
And the total number of search results 104 is
13 is output. The document word matrix generation unit 106 generates the document word matrix 107 shown in FIG. 11 by counting the number of times each word appears for each document based on the document sample set 105, The number is output to the number estimation unit 113. In the document word matrix 107, the row and the column indicate the document identifier and the search word list, respectively, and the value of the cell in the table indicates the number of times the search word of the corresponding column appears in the document of the corresponding row.

[0011] The retrieval word candidate presentation unit 108 calculates the word importance indicating how important an arbitrary word is in an arbitrary document, excluding words such as particles and pronouns from the document word matrix 107. The word importance is an index that is high for words appearing intensively in a specific document, and is low for words appearing in many documents. In addition, the search word candidate presentation unit 108 outputs a word having a high word importance as the search word candidate 109 to the search word selection unit 110 based on the calculated word importance.

[0012] The search word selection unit 110 provides a search word candidate 109.
Is output to a monitor or the like, and the user is caused to select an arbitrary search word from the search word candidates 109 and input the selected search word as a selection signal 111. Further, the search term selection unit 110 outputs the additional search term 112 to the number estimation unit 113 based on the selection signal 111 input by the user.

The total number of search results 10
4. The document word matrix 107 and the additional search word 112 are input. The case number estimation unit 113 counts the number of rows in which the columns of the additional search words 112 in the document word matrix 107 are not “0”, and divides this by the total number of rows to obtain the appearance rate of the additional search words 112. Further, the number of cases estimation unit 113
The estimated number 114 obtained by multiplying the appearance rate of Twelve by the total number of search results 104 is output.

[0014]

Since the conventional document retrieval apparatus is configured as described above, with respect to the effect of narrowing down search terms, the conventional techniques disclosed in References 1 and 2 disclose documents of search results. Classifying and presenting the search terms in the classified category to the user. When selecting the presented search terms for use in the refined search, information of the refinement effect corresponding to the presented search terms is used. However, there is no problem, and it is difficult to select a search word for performing a narrowed search efficiently.

[0015] Further, in the conventional document search apparatus, regarding the visualization of search results, in the prior art disclosed in Reference 3, the effect of narrowing down search terms is presented to the user in units of documents, but the search results are several thousand. When the size becomes large, it is extremely difficult to display the effect of narrowing down the entire search result using the document word matrix, and there is a problem that the listability is lacking.

Further, the conventional document search apparatus outputs a set of document samples from the search results in the prior art disclosed in Document 4 with respect to visualization of the search results, and handles each document based on the set of document samples. Since the number of refinements for the entire search result is estimated from the appearance rate of the additional search terms, when the estimated number of refinements is presented to the user, each additional search term is presented. There has been a problem that it is extremely difficult to display the effects and the list is lacking.

Further, in the conventional document search apparatus, with respect to narrowing down search to a target document, in the related art disclosed in Document 2, a category designated for narrowing down to a target document is classified into subcategories in detail. However, not all the target documents exist in the specified category, and the documents desired by the user also exist in other categories.
Documents obtained by the refined search are limited to only documents that exist in the designated category, and there is a problem that it is difficult to perform the refined search again because a search for a target document is omitted.

In the conventional document search apparatus, with respect to narrowing down search to a target document, according to the conventional techniques disclosed in References 3 and 4, an AND search in which a search term is added to an initial search result is performed. The search target of the narrow search is limited to documents of the search result for the immediately preceding search word, and it is impossible to search for the target documents scattered in the entire document space, and it is difficult to perform the narrow search again. There was a problem that was.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problem, and an index representing a narrowing effect is given to a search word presented for a narrow search, thereby facilitating selection of an additional search word. Accordingly, it is an object of the present invention to obtain a document search device and a document search method that can efficiently perform a narrow search of a user.

Further, the present invention classifies documents of a search result, sets a classified document set as a topic, and presents a user with an effect of narrowing down a search word for each topic which is a unit larger than the document, thereby performing a search. It is an object of the present invention to provide a document search apparatus and a document search method that enhance the list of narrowing-down effects for the entire result document.

Further, according to the present invention, the first search result is re-classified by using information of a topic or a search word specified by a user as feedback information, thereby enabling a target document dispersed among topics to be converted into one topic. To collect
It is an object of the present invention to obtain a document search device and a document search method that efficiently narrow down search results.

[0022]

According to the present invention, there is provided a document search apparatus for searching and outputting a document meeting a search condition from a set of documents, based on a frequency of occurrence of a word described in the document. A document feature extracting means for calculating a statistical weight of words and outputting a set of document vectors obtained from the weights; and creating a topic by classifying the set of document vectors according to the similarity between the document vectors. Topic classification means for outputting document information belonging to the topic and search words with importance of the topic, and narrowing effect estimation for calculating and outputting a narrowing effect index of the search word belonging to the topic with reference to the document information and the search words with importance. Means and a narrowing effect index are assigned to a search word, and a search word having a high narrowing effect index is selected as a suggested search word candidate, and the proposed search word candidate and the suggested search word candidate are supported. And a presentation search word candidate generating means for outputting document information to be presented, associating the presentation search word candidate with a topic, presenting it together with a narrowing-down effect index, and inputting either or both of instruction information and selection information. It includes a classification result presenting means for prompting, and a document feature setting means for changing and outputting a document vector included in the document vector set based on the selection information.

In the document search apparatus according to the present invention, the narrowing effect estimating means calculates and outputs a narrowing effect index for at least one or a plurality of topics.

[0024] In the document search apparatus according to the present invention, the classification result presenting means presents the topic and the proposed search term candidates belonging to the topic in a matrix format, and presents a narrowing effect index to each element of the matrix. It was done.

The document retrieval apparatus according to the present invention extracts a word characterizing the document from the document output by the document retrieval means as a topic classification item, refers to a text related to the topic classification item, and assigns a weight to the topic classification item. A topic classification item acquiring unit that learns a vector and outputs a topic classification item and a weight vector to the topic classification unit. The topic classification unit classifies a document vector set based on the topic classification item and the weight vector. It is intended to create a topic.

[0026] In the document search apparatus according to the present invention, the topic classification item acquiring means extracts the topic classification item from the document output from the document retrieval means based on the appearance frequency of the word described in the document. It was done.

[0027] In the document search apparatus according to the present invention, the topic classification item acquiring means extracts a topic classification item from a document output from the document retrieval means by referring to a tag described in the document. It is.

The document retrieval apparatus according to the present invention includes a designated document feature extraction unit that calculates a document vector from a document specified via a classification result presentation unit and outputs the document vector to a document feature setting unit. The document vector set is changed based on the document vector output from the designated document feature extraction means and the document vector set output from the document feature extraction means.

[0029] A document search device according to the present invention defines first words that are related to a predetermined word and records the words as related words, and stores the related words corresponding to the specified search words in the first word.
Related word setting means for extracting from the recording means and outputting the same to the document feature setting means, based on the related words and the selection information input from the classification result presentation means,
The document vector set is changed.

[0030] The document search device according to the present invention includes a second recording unit for recording the knowledge of creating a search request sentence, and a search term output by the presented search word candidate generation unit with reference to the second recording unit. And a search request creating means for creating a search request sentence corresponding to the search request and outputting the search request sentence to the document search means, wherein the suggested search word candidate generating means selects a search word to be output to the search request creating means based on the narrowing effect index. It is like that.

[0031] In the document search apparatus according to the present invention, the suggested search word candidate generation means selects a plurality of search words and outputs the selected search words to the search request creation means, and the search request creation means performs a logical operation on the plurality of search words. This is to create a search request sentence.

According to the document search method of the present invention, a document search step for searching and outputting a document meeting a search condition from a set of documents, and a statistical search of words based on the frequency of occurrence of words described in the document. A document feature extraction step of calculating weights and outputting a set of document vectors obtained from the weights; and creating a topic by classifying the set of document vectors according to the similarity between the document vectors, and generating document information and topic information belonging to the topic. A topic classification step of outputting a search word with importance; a narrowing effect estimation step of calculating and outputting a narrowing effect index of a search word belonging to a topic by referring to the document information and the search word with importance; and a narrowing effect index Is added to the search term, and a search term having a high narrowing effect index is selected as a suggested search term candidate, which corresponds to the suggested search term candidate and the suggested search term candidate. A suggested search word candidate generating step of outputting document information to be presented, associating the suggested search word candidate with a topic, presenting the same together with a narrowing effect index, and inputting either or both of instruction information and selection information. Based on the classification result presentation step to be prompted and the selection information,
A document feature setting step of changing and outputting a document vector included in the document vector set.

In the document search method according to the present invention, the narrowing effect estimation step calculates and outputs a narrowing effect index for at least one or a plurality of topics.

In the document search method according to the present invention, the classification result presenting step presents the topic and the presented search word candidates belonging to the topic in a matrix format, and presents a narrowing effect index to each element of the matrix. It was done.

In the document search method according to the present invention, a word characterizing the document is extracted as a topic classification item from the document output by the document search step, and a text related to the topic classification item is referred to, and a weight for the topic classification item is determined. A topic classification item acquiring step of learning a vector and outputting a topic classification item and a weight vector, wherein the topic classification step creates a topic by classifying a set of document vectors based on the topic classification item and the weight vector It is something to do.

In the document search method according to the present invention, the topic classification item obtaining step extracts the topic classification item from the document output from the document search step based on the appearance frequency of the word described in the document. It was done.

[0037] In the document search method according to the present invention, the topic classification item obtaining step extracts a topic classification item from a document output from the document search step by referring to a tag described in the document. It is.

The document search method according to the present invention includes a designated document feature extraction step of calculating and outputting a document vector from a document designated via the classification result presentation step, and the designated document feature extraction step includes the designated document feature extraction step. The document vector set is changed based on the document vector output from the step and the document vector set output from the document feature extraction step.

In the document search method according to the present invention, a first recording step of defining a word related to a predetermined word and recording the word as a related word, and extracting and outputting a related word corresponding to the specified search word A related word setting step, wherein the document feature setting step changes the document vector set based on the related word and the selection information input from the classification result presentation step.

The document search method according to the present invention creates and outputs a search request sentence corresponding to the search word output by the second recording step of recording the creation knowledge of the search request sentence and the presented search word candidate generation step. A search request creation step, wherein the presented search word candidate generation step selects a search word to be output based on the narrowing effect index.

In the document search method according to the present invention, the suggested search word candidate generating step selects and outputs a plurality of search words, and the search request creating step creates a search request sentence from a logical operation on the plurality of search words. It is something to do.

[0042]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a configuration of a document search device according to Embodiment 1 of the present invention. In FIG. 1, reference numeral 1 denotes a set of documents to be searched, for example, an HTML document that can be browsed using the Internet, an e-mail that can be transmitted and received using the Internet, a large-scale text recorded on a recording device or a recording medium. It consists of digitized text such as digitized documents recorded and managed in a database. A document search unit (document search means) 2 for searching a document from the document set 1 according to a predetermined condition;
Reference numeral 3 denotes a search result document set which is a result of the document search unit 2 searching from the document set 1 according to predetermined conditions. (Document feature extraction means). Note that the document vector is a representation of the weight of each word for each document in the form of a vector.

In FIG. 1, reference numeral 5 denotes a document vector set output based on the document vector created by the document feature extraction unit 4, and reference numeral 6 denotes a calculation based on the document vector set 5 output by the document feature extraction unit 4. A topic classifying unit (topic classifying means) for creating a topic by classifying the document vector set 5 into a plurality of sets according to the similarity between the document vectors, and 7 outputs together with the document information of each topic classified by the topic classifying unit 6 A set of search word candidates with importance, 8 is a presentation search word candidate generation unit (presentation search word candidate generation means) for selecting a presentation search word candidate based on a predetermined criterion from the search word set with importance output from the topic classification unit 6. ). The predetermined criterion for selecting the presented search word candidate is, for example, to select a certain upper number in order of importance.

Further, in FIG. 1, reference numeral 9 denotes a set of search terms with importance of each topic output together with document information of each topic from the presented search word candidate generation unit 8; A narrowing effect estimating unit (a narrowing effect estimating means) for estimating a narrowing effect of a search word belonging to a specified topic; A suggested search word candidate that is selected by the suggested search word candidate generating unit 8 based on the narrowing effect index 11 and is output together with the document information belonging to the corresponding topic and the narrowing effect index 11;
Reference numeral 13 denotes a user who operates the document search apparatus, and reference numeral 14 denotes a narrowed-down effect index 11 by associating the presented search word candidate 12 with each topic.
And a classification result presentation unit (classification result presentation means) for presenting to the user 13, a classification result 15 visualized by the classification result presentation unit 14 for presentation to the user 13, and a classification result 16 for the user 13.
Is instruction information to be transmitted to the classification result presentation unit 14.

Further, in FIG.
Is a search condition input to the document search unit 2, and 18 is a user 13
Is a topic specified by the user 13 input to the narrowing-down effect estimation unit 10, 19 is selection information of the user 13 output from the classification result presentation unit 14 when the user 13 instructs re-classification of search results, and 20 is a user 13 selection information 19
A document feature setting unit (document feature setting unit) 21 that changes the weight of the document vector or the search word based on the document vector 21 is a changed document vector output from the document feature setting unit 20 to the topic classification unit 6.

FIG. 2 is a flowchart illustrating the operation of the document search device according to the first embodiment of the present invention. FIG. 3 is an explanatory diagram showing an example of a document vector according to Embodiment 1 of the present invention. FIG. 4 is an explanatory diagram showing an example of the search term-topic correspondence table according to Embodiment 1 of the present invention.

Next, the operation will be described. First, in step ST1, the user 13 inputs a search condition 17 to the document search unit 2. The search condition 17 is, for example, a search term or a logical expression between a plurality of search terms. Next, step S
At T2, the document search unit 2 searches the input search condition 1
7 is searched for the document of document set 1 based on the search condition 17
A search result document set 3 that conforms to is output (document search step). The document set 1 to be searched is recorded and managed in, for example, an HTML document that can be browsed using the Internet, an e-mail that can be transmitted and received using the Internet, and a large-scale text database recorded on a recording device or a recording medium. It is a digitized text such as a digitized document. Further, the document search unit 2 only needs to be searchable under the search condition 17, and may use, for example, a full-text search engine commonly used on the Internet. Further, the document search unit 2 outputs various information related to the document of the search result, for example, information such as a document ID number for specifying the document, a location of the document file, and a title of the document as the document information together with the search result document set 3. I do.

Next, in step ST3, the document feature extraction unit 4 obtains a document vector for each document in the search result document set 3. As shown in FIG. 3, the document vector is calculated for each document D based on the frequency of occurrence of each word for each document.
, DW,..., KWn are calculated for each of the words KW1, KW2,..., KWi,. In the form of a vector. There are various methods for calculating the statistical weights, such as TF / IDF and chi-square statistical values, which may be appropriately selected and used according to the purpose. The document feature extraction unit 4 outputs a document vector set 5 obtained by calculating a statistical weight (document feature extraction step).

Next, in step ST4, the topic classifying unit 6 calculates the similarity between the document vectors, and classifies the document vector set 5 into a plurality of sets according to the similarity to create a topic (topic classifying step). ). Methods of classifying documents are roughly classified into two types: a document classification in which a classification category is provided by giving a top-down classification category, and a clustering in which documents similar to bottom-up are put together.

In the document classification, the classification destination category is set in advance, and the statistical weight for the classification category is learned from the sample documents belonging to the category as the classification category vector. The similarity between the document vector and the category vector is calculated. As the similarity, for example, an inner product value of a vector is used. Based on the similarity obtained in this way, the document is classified into the category having the highest similarity. On the other hand, in the clustering, the similarity between all the document vectors existing in the input document vector set 5 is calculated, and the document vectors having a high similarity are collectively combined into one cluster, and the similarity for the cluster is calculated and collected. The document is classified by repeating the above processing.

The topic classification unit 6 only needs to have a function of classifying document vectors, and is not limited to the above-described document classification and clustering, but may adopt another classification method (for example, principal component analysis). The topic classification unit 6
The classified set is set as a topic, and a word having high importance calculated for each topic is set as a search word. For example, for words in all documents belonging to a certain topic, words set as unnecessary and separately set words are deleted, the frequency of appearance of each word is counted, and the importance is determined. . further,
The topic classification unit 6 outputs document information belonging to each topic and a search word set 7 with importance of each topic.

Next, in step ST5, the user 1
If the search for the user 3 is not the first time, the process proceeds to step ST6. If the search for the user 13 is not the first time, the process proceeds to step ST7. In step ST6, the suggested search word candidate generating unit 8 selects the suggested search word candidates 12 from the set of search words with importance of each topic 7 based on a predetermined criterion, and proceeds to step ST8. The predetermined criterion for selecting the suggested search word candidate 12 is, for example, a certain number of higher ranks in order of importance. on the other hand,
In step ST7, the narrowing effect index 11 calculated by the narrowing effect estimating unit 10 is assigned to each search word candidate, and a search word candidate with a high narrowing effect index 11 is selected to be a presented search word candidate 12. The presented search word candidate generation unit 8 outputs the presented search word candidate 12 thus selected.
Is output to the classification result presentation unit 14 together with the document information belonging to the corresponding topic (presentation search word candidate generation step).

Next, in step ST8, the classification result presentation section 14 classifies the presentation search word candidates 12 into visualizations together with the narrowing-down effect index 11 in association with each topic.
5 is presented to the user 13 (classification result presentation step). The visualization method is, for example, as shown in FIG.
A search word-topic correspondence table is used. The search term-topic correspondence table is
The presented search word candidate 12 and the corresponding topic are expressed in the form of a matrix, and a narrowing effect index 11 is presented to the user 13 as visualized information for each element of the matrix. Further, the classification result presenting unit 14 uses the document information belonging to the corresponding topic to generate the topics T1, T2,.
Is specified by the user 13, a list of documents belonging to the specified topic and various types of document information can be referred to.

Next, in step ST9, the user 1
3 designates a topic by referring to the search word-topic correspondence table presented in the classification result presentation section 14 and proceeds to step ST1.
If the user 13 does not specify a topic, the process proceeds to step ST12. In step ST10, the user 13 narrows the topic 18 to be specified by the narrowing-down effect estimation unit 10
To enter. Next, in step ST11, the document information belonging to each topic and the search word set 9 with importance are obtained with reference to the presented search word candidate generation unit 8, and the search words belonging to the topic 18 specified by the user 13 are narrowed down. Effect index 11
Is calculated (a narrowing effect estimation step). As a method of calculating the narrowing-down effect index 11, for example, for a search word belonging to the topic 18 specified by the user 13, the statistical weight with respect to each topic is calculated again. Alternatively, the number of documents including a search word belonging to the topic 18 specified by the user 13 is set as the narrowing effect index 11. The narrowing-down effect index 11 calculated in this manner is output to the suggested search word candidate generating unit 8, and further output from the suggested search word candidate generating unit 8 to the classification result presenting unit 14. Is visualized like a search word-topic correspondence table as shown in FIG. 3, and the narrowing-down effect index 11 is presented to the user 13 in correspondence with the presented search word candidates 12 belonging to the topic 18 specified by the user 13 and each topic. When the process in step ST11 ends, the process returns to step ST7.

If a plurality of topics can be specified for the topic 18 to be input to the narrowing-down effect estimating unit 10, the selection range of the topics is widened, so that the contents of the document desired by the user 13 can be more accurately determined. Can be specified.
For example, if a certain number of higher rankings are selected in descending order of importance according to search words belonging to a plurality of designated topics, and a narrowing effect index 11 is presented to the user 13 based on a correspondence between the selected presented search word candidate 12 and each topic. Good.

Further, the narrowing effect index 11 calculated by the narrowing effect estimating unit 10 is calculated not by a search word but by a topic unit, so that the tendency of the narrowing effect index 11 in a wider range is determined for the entire search result. Be able to grasp. For example, a topic weight vector is created based on the frequency of word appearance in a document belonging to a specified topic,
The similarity with each topic may be obtained as the inner product of the vectors, and the similarity between the designated topic and each topic may be presented to the user 13 in the form of a matrix.

On the other hand, in step ST12, the user 13 refers to the search word-topic correspondence table presented by the classification result presentation section 14 and sends the topic 1 to the narrowing-down effect estimation section 10.
If re-classification is instructed without inputting 8, step S
Proceeding to T13, this process ends if no re-classification is instructed.

In step ST13, the user 13 inputs, to the classification result presentation unit 14, instruction information 16 for designating a topic or a search word whose contents are close to the document to be searched, and gives a re-classification instruction. The classification result presentation unit 14 outputs information such as a document ID number of a document belonging to the specified topic and a specified search word to the document feature setting unit 20 as selection information 19 of the user 13. Next, step S
At T14, based on the selection information 19 of the user 13, the document feature setting unit 20 changes the document vector corresponding to the input document ID number and the weight for the specified search word, and changes the changed document vector 21. The topic classification unit 6
(Document feature setting step). Changing the weight
For example, the weight of a document or a search word belonging to a specified topic is changed by adding a predetermined constant value to the weight. When the process in step ST14 ends, the process returns to step ST4.

As described above, according to the first embodiment, the narrowing-down effect estimating unit 10 for calculating the narrowing-down effect index 11 of the search word belonging to the topic 18 specified by the user 13.
And a classification result presenting unit 14 for visualizing the narrowing effect index 11 and presenting the narrowing effect index 11 to the user 13. The presented search word candidate 12 belonging to the topic 18 designated by the user 13 and the narrowing effect index 11 corresponding to each topic are provided. Since it is presented to the user 13, the user 13 can easily select an additional search word, so that an additional search word having a high narrowing effect can be accurately selected, and the effect of improving the efficiency of the narrow search can be obtained.

Further, according to the first embodiment, the topic classifying unit for creating a topic by classifying the document vector set 5 into a plurality of sets according to the similarity between the document vectors calculated based on the document vector set 5 6 and a classification result presentation unit 14 for visualizing the search term narrowing effect index 11 for each topic as a search word-topic correspondence table and presenting it to the user 13, wherein the topic unit is a unit larger than the document unit. Since the narrowing-down effect is presented, the effect of listing the narrowing-down effect for the entire document in the search result is enhanced, and the effect of efficiently grasping the contents of the entire search result is obtained.

Further, according to the first embodiment, the instruction information 16 specified by the user 13 as feedback information.
Document feature setting unit 2 that changes the document vector based on
0, and re-classified for the first search result, so that the target documents dispersed among topics are collected into one topic, so that the efficiency of narrowing down search to the target documents is improved. The effect is obtained.

Embodiment 2 FIG. 5 is a block diagram showing a configuration of the document search device according to the second embodiment of the present invention. In FIG. 5, the same reference numerals as those in FIG. 1 denote the same or corresponding parts, and a description thereof will not be repeated. 31 is a topic classification item for extracting a word characterizing a document from the documents in the search result document set 3 as a topic classification item, referring to a text related to the extracted topic classification item, and learning a weight vector for the topic classification item. An acquisition unit (topic classification item acquisition means). Reference numeral 32 denotes topic classification item information output by the topic classification item acquisition unit 31 to the topic classification unit 6 and including a topic classification item and a weight vector.

Next, the operation will be described. Embodiment 2
, Document set 1, document search unit 2, document feature extraction unit 4, presentation search word candidate generation unit 8, narrowing effect estimation unit 1
The operations of the user 0, the user 13, the classification result presenting unit 14, the document feature setting unit 20, and the like, and the effects produced by them are the same as those in the first embodiment, and thus description thereof will be omitted.

The topic category item acquiring unit 31 extracts words characterizing the document from the documents in the search result document set 3 as topic category items, refers to the text related to the extracted topic category items, and refers to the topic category items. , And outputs the topic classification item together with the weight vector as topic classification item information 32 to the topic classification unit 6 (topic classification item acquisition step).

The topic classification items include, for example, words having a high appearance rate in a document, a title of an HTML document obtained by referring to an HTML tag described in the HTML document, and a domain name included in a URL. The text related to the extracted topic classification item is analyzed, for example, by analyzing the text around the position where the topic classification item exists, and specifying a specific HT such as a paragraph break, a chapter, a bullet point, or a link destination.
The text associated with the topic classification item is extracted by detecting the ML tag and copying the text associated with the detected HTML tag.

The topic classification item acquiring unit 31 learns a weight vector for the topic classification item by using the text related to the extracted topic classification item, and outputs the topic classification item to the topic classification unit 6 together with the weight vector. I do. The topic classification unit 6 calculates the similarity between the topic classification item information 32 input from the topic classification item acquisition unit 31 and each document vector of the document vector set 5 input from the document feature extraction unit 4 by, for example, the inner product of the vectors. By the calculation, the document is classified into the topic classification item having the highest similarity.

As described above, according to the second embodiment, the same effects as those of the first embodiment can be obtained, and the topic classification item is extracted from the documents in the search result document set 3.
A topic classification item acquiring unit 31 that outputs a topic classification item together with a weight vector for the topic classification item;
Since the documents are classified based on the similarity with the respective document vectors, the topic classification items used in the topic classification unit 6 can be automatically obtained, and it is not necessary to set the classification categories in advance. The result is that the work of setting the resulting document item is not required, and the efficiency of the narrowed search of the user 13 is improved.

Embodiment 3 FIG. 6 is a block diagram showing a configuration of a document search device according to Embodiment 3 of the present invention. 6, the same reference numerals as those in FIG. 1 denote the same or corresponding parts, and a description thereof will not be repeated. 41 is a designated document feature extraction unit (designated document feature extraction unit) that calculates a document vector from a document designated by the user 13; 42 is a document vector output by the designated document feature extraction unit 41 to the document feature setting unit 20; This is a designated document output to the designated document feature extraction unit 41 by the user 13.

Next, the operation will be described. Embodiment 3
, The operations of the document set 1, the document search unit 2, the document feature extraction unit 4, the topic classification unit 6, the presentation search word candidate generation unit 8, the narrowing effect estimation unit 10, the classification result presentation unit 14, and the effects produced by them Is the same as in the first embodiment, and a description thereof will be omitted.

The user 13 refers to the classification result presentation unit 14, selects a document whose content is close to the document desired by the user 13, and instructs the designated document feature extraction unit 41 of the document 43 to be designated. The designated document feature extraction unit 41 specifies the designated document 4
3, a document weight 42 is calculated based on the number of appearances of the word included in the document feature setting unit 2, and the document feature setting unit 2
Output to 0. The document feature setting unit 20 calculates the similarity between the document vector 42 of the designated document 43 and the document vector set 5 input from the document feature extraction unit 4, and changes the weight of the document vector in the document vector set 5. (Designated document feature extraction step). For example, a certain number of upper-order document vectors are selected from the document vector set 5 in descending order of similarity, and the similarity is added to the weight of the document vector to be changed. The topic classifying unit 6 classifies the changed document vector set 5.

As described above, according to the third embodiment, the same effects as those of the first embodiment can be obtained, and the designated document feature extracting unit 41 for calculating the document vector from the document designated by the user 13 is provided. , Designated document feature extraction unit 41
Calculates the similarity between the document vector 42 output by the user and the document vector set 5 output by the document feature extraction unit 4 and changes the weight of the document vector. By directly specifying 43, it is possible to change the document vector of the document feature setting unit 20, and it is possible to obtain an effect that the efficiency of the refined search of the user 13 is improved.

Embodiment 4 FIG. 7 is a block diagram showing a configuration of a document search device according to Embodiment 4 of the present invention. 7, the same reference numerals as those in FIG. 1 denote the same or corresponding parts, and a description thereof will not be repeated. 51 is a related word dictionary (first word) in which a word and a related word related to the word are defined.
Is a related word setting unit (related word setting means) for selecting a related word related to the search word from the related word dictionary 51 and outputting the related word when the search word is input, and 53 is a document feature setting. The search word 54 output by the unit 20 and input to the related word setting unit 52 is indicated by 54.
It is a related word that is input to 0.

FIG. 8 is an explanatory diagram showing an example in which words and related words are defined in the fourth embodiment of the present invention. In FIG. 8, the related word has a different notation and a similar word as related words, and corresponds to the word described in each line.

Next, the operation will be described. Embodiment 4
The operations of the document set 1, the document search unit 2, the document feature extraction unit 4, the topic classification unit 6, the presentation search word candidate generation unit 8, the narrowing effect estimation unit 10, the user 13, the classification result presentation unit 14, and the like Are the same as those in the first embodiment, and the description thereof is omitted.

As shown in FIG. 8, a predetermined word and a word related to the word are previously recorded in the related word dictionary 51 (first recording step). Related word setting unit 52
Refers to the related word dictionary 51, extracts a different notation and a similar word corresponding to the input search word 53, sets them as a related word 54, and outputs the related word 54 to the document feature setting unit 20 (a related word setting step). ). The document feature setting unit 20 includes a related word setting unit 52
Is added to the selection information 19 of the user 13 input from the classification result presentation unit 14, and the document vector set 5 input from the document feature extraction unit 4 is changed.

As described above, according to the fourth embodiment, the same effects as those of the first embodiment can be obtained, and the related word setting unit 52 that outputs the related word 54 with reference to the related word dictionary 51 is provided. Related word 54 input from related word setting unit 52
Is added to the selection information 19 of the user 13 input from the classification result presentation unit 14 and the document vector set 5 is changed, so that a document including a different notation or a similar word of the search word 53 can be searched. Therefore, an effect is obtained that search omission is suppressed and the efficiency of search for finding a document desired by the user 13 is improved.

Embodiment 5 FIG. 9 is a block diagram showing a configuration of a document search device according to Embodiment 5 of the present invention. 9, the same reference numerals as those in FIG. 1 denote the same or corresponding parts, and a description thereof will not be repeated. Reference numeral 61 denotes search request creation knowledge (second recording means) which is composed of a magnetic recording device or the like and records search request creation knowledge, and 62 denotes an optimum search request based on the search request search word when a search request search word is input. A search request creation unit (search request creation means) for selecting a sentence from the search request creation knowledge 61 and outputting a search request sentence, 63 is a search request search output by the presented search word candidate generation unit 8 and input to the search request creation unit 62 The word 64 is a search request sentence output by the search request creation unit 62 and input to the document search unit 2.

Next, the operation will be described. Embodiment 5
The operations of the document set 1, the document feature extracting unit 4, the topic classifying unit 6, the narrowing-down effect estimating unit 10, the user 13, the classification result presenting unit 14, the document feature setting unit 20, etc., and the effects of these are implemented. Since it is the same as the first embodiment, the description is omitted.

The suggested search word candidate generating unit 8 refers to the narrowing effect index 11 input from the narrowing effect estimating unit 10, and sets a search word having a high narrowing effect index 11 to the search request creating unit 62 as a search request search word 63. Output. The search request search term 63 may be plural. Search request creation knowledge 6
1, the knowledge for creating the search request sentence 64 instructing the document search unit 2 to execute the search process is defined in advance (second recording step). For example, search request sentence 6
4 is composed of two types, a search command and a search condition.
As the search command, for example, command types such as <search execution>, <acquisition of execution status>, and <database specification> are defined. As the search condition, a description format of parameters such as a search word, a logical operator between the search words, and specification of information obtained as a search result is defined.

The search request creating section 62 refers to the search request creating knowledge 61 and sets a search request search word 63 as a search condition in accordance with the definition of the search request sentence 64 (search request creating step). The search command of the search request sentence 64 is, for example, <search execution>, and the search request sentence 64 is created and output to the document search unit 2. When setting the search condition of the search request sentence 64, for example, an AND
If the narrowing effect index 11 assigned to the search request search word 63 is greater than or equal to a preset threshold, the narrowing effect is considered to be high, and the OR operator is used to search a wider range of documents. Can also be described as

As described above, according to the fifth embodiment, the same effects as those of the first embodiment can be obtained, and a search request creating unit 62 for creating a search request sentence 64 with reference to the search request creation knowledge 61. And a search is performed again based on the search request sentence 64 output by the search request creation unit 62. Therefore, a document desired by the user 13 that is not included in the first search result is automatically searched from the document set 1. As a result, it is possible to obtain the effect that the efficiency of the refined search of the user 13 is improved.

[0082]

As described above, according to the present invention, a document retrieval means for retrieving and outputting a document meeting a retrieval condition from a document set and a word retrieval method based on the appearance frequency of the word described in the document. A document feature extracting means for calculating a statistical weight of the document vector and outputting a set of document vectors obtained from the weight, and generating a topic by classifying the set of document vectors according to the similarity between the document vectors, thereby forming a document belonging to the topic. A topic classification means for outputting information and a search word with the importance of the topic; a narrowing effect estimating means for calculating and outputting a narrowing effect index of the search word belonging to the topic by referring to the document information and the search word with the importance; , Assigning a narrowing effect index to a search word, selecting a search word having a high narrowing effect index as a suggested search word candidate, and presenting the suggested search word candidate and the document information corresponding to the suggested search word candidate A suggested search word candidate generating means for outputting a suggestion search word candidate and a topic in association with a narrowing-down effect index, and prompting the user to input one or both of instruction information and selection information Means, and a document feature setting means for changing and outputting the document vector included in the document vector set based on the selection information. Since the narrowing effect index is provided, it is possible to easily select an additional search word, and it is possible to perform the narrowing search efficiently.

According to the present invention, the narrowing effect estimating means is configured to calculate and output a narrowing effect index for at least one or a plurality of topics. Since the refined search can be executed, it is possible to grasp the tendency of the refined effect index for a wider range of search results, and it is possible to accurately specify a desired document content by the user.

According to the present invention, the classification result presenting means:
Since the topic and the proposed search word candidates belonging to the topic are presented in the form of a matrix, and the narrowing effect index is presented for each element of the matrix, the listability of the narrowing effect for the entire document in the search result is improved, The effect is that the contents of the entire search result can be efficiently grasped.

According to the present invention, a word characterizing the document is extracted as a topic classification item from a document output by the document retrieval means, and a weight vector for the topic classification item is learned by referring to a text related to the topic classification item. And a topic classification item acquiring unit that outputs the topic classification item and the weight vector to the topic classification unit. The topic classification unit creates a topic by classifying a set of document vectors based on the topic classification item and the weight vector. , The topic classification items used by the topic classification means can be automatically obtained, and it is not necessary to set the classification category in advance. This has the effect of improving search efficiency.

According to the present invention, the topic classification item obtaining means is configured to extract the topic classification items from the document output from the document search means based on the frequency of appearance of the word described in the document. This has the effect of automatically and efficiently extracting topic classification items.

According to the present invention, the topic classification item acquiring means is configured to extract the topic classification item from the document output from the document search means by referring to the tag described in the document. The topic classification item can be automatically and efficiently extracted from the document in which is described.

According to the present invention, there is provided a designated document feature extraction means for calculating a document vector from a document designated via the classification result presentation means and outputting the document vector to the document feature setting means. Since the document vector set is changed based on the document vector output from the feature extraction unit and the document vector set output from the document feature extraction unit, a document having contents close to the document desired by the user can be directly input. By specifying, it is possible to change the document vector of the document feature setting means, and it is possible to improve the efficiency of the narrowed search.

According to the present invention, the first recording means for defining a word related to a predetermined word and recording it as a related word,
A related word setting means for extracting a related word corresponding to the specified search word from the first recording means and outputting the related word to the document feature setting means, wherein the document feature setting means inputs the related word and the classification result presentation means The document vector set is configured to be changed based on the selected information, so that it becomes possible to search for a document that includes a different notation or a similar word of the search word. This has the effect of improving the efficiency of the search for finding the desired document.

According to the present invention, the second recording means for recording the creation knowledge of the search request sentence, and the second search means refers to the second storage means and corresponds to the search word output by the presented search word candidate generation means. A search request creating unit that creates a search request sentence and outputs the created search request sentence to the document search unit.
Since a search term to be output to the search request creating means is selected based on the narrowing effect index, a user who is not included in the first search result can automatically search for a desired document from the document set. At the same time, there is an effect that the efficiency of the refined search by the user is improved.

According to the present invention, the presented search word candidate generation means selects a plurality of search words and outputs the selected search words to the search request creation means, and the search request creation means executes a search request sentence from a logical operation on the plurality of search words. Is configured to create
In the D operation, the refined search can be executed more accurately, and in the OR operation, the refined search can be executed in a wider range.

According to the present invention, a document search step for searching and outputting a document meeting a search condition from a set of documents;
A document feature extraction step of calculating a statistical weight of the word based on the frequency of occurrence of the word described in the document and outputting a set of document vectors obtained from the weight; A topic classification step of creating a topic by classifying according to the degree and outputting document information belonging to the topic and a search word with the importance of the topic;
A narrowing effect estimation step of calculating and outputting a narrowing effect index of the search word belonging to the topic by referring to the document information and the search word with importance, and a search word having a high narrowing effect index by assigning the narrowing effect index to the search word Selecting a suggested search word candidate, and a suggested search word candidate generating step of outputting the suggested search word candidate and the document information corresponding to the suggested search word candidate, and associating the suggested search word candidate with the topic, A classification result presenting step of presenting together with the narrowing-down effect index and prompting to input either or both of the instruction information and the selection information; and changing and outputting the document vector included in the document vector set based on the selection information. And a document feature setting step, so that a narrowing effect index is assigned to a suggested search word candidate to be presented for the narrowing search. From additional search term select can facilitate the, an effect that the narrowing search can be efficiently performed.

According to the present invention, since the narrowing effect estimation step is configured to calculate and output a narrowing effect index for at least one or a plurality of topics, the narrowing effect estimation step is performed in a topic unit which is a unit larger than a document unit. Since the refined search can be executed, it is possible to grasp the tendency of the refined effect index for a wider range of search results, and it is possible to accurately specify a desired document content by the user.

According to the present invention, the classification result presenting step is configured to present the topic and the proposed search word candidates belonging to the topic in the form of a matrix, and to present a narrowing effect index to each element of the matrix. In addition, it is possible to enhance the listability of the narrowing-down effect of the search result with respect to the entire document, and it is possible to efficiently grasp the contents of the entire search result.

According to the present invention, a word characterizing the document is extracted as a topic classification item from a document output by the document retrieval step, and a weight vector for the topic classification item is learned by referring to a text related to the topic classification item. And
A topic classification item obtaining step for outputting a topic classification item and a weight vector, wherein the topic classification step is configured to generate a topic by classifying a document vector set based on the topic classification item and the weight vector. So
The topic classification item used in the topic classification step can be obtained automatically, and it is not necessary to set the classification category in advance. Therefore, the work of setting the document item of the search result becomes unnecessary, and the efficiency of the narrowed search is improved. To play.

According to the present invention, the topic classification item obtaining step is configured to extract a topic classification item from a document output from the document search step based on the frequency of occurrence of a word described in the document. This has the effect of automatically and efficiently extracting topic classification items.

According to the present invention, the topic classification item obtaining step is configured to extract the topic classification item from the document output from the document search step by referring to the tag described in the document. The topic classification item can be automatically and efficiently extracted from the document in which is described.

According to the present invention, there is provided a designated document feature extraction step for calculating and outputting a document vector from a document designated via the classification result presentation step, and the document feature setting step includes the step of outputting from the designated document feature extraction step. The document vector set is configured to change the document vector set based on the document vector set output from the document feature extraction step, so that the user directly specifies a document having contents close to the desired document, This makes it possible to change the document vector in the document feature setting step, and to improve the efficiency of the narrowed search.

According to the present invention, a first recording step of defining a word related to a predetermined word and recording it as a related word, and a related word setting for extracting and outputting a related word corresponding to a specified search word And the document feature setting step is configured to change the set of document vectors based on the related words and the selection information input from the classification result presentation step. Can be searched, so that omission of search is suppressed and the efficiency of search for finding a document desired by a user is improved.

According to the present invention, the second recording step of recording the creation knowledge of the search request sentence and the search request creation of creating and outputting the search request sentence corresponding to the search word output by the presented search word candidate generation step And the step of generating a suggested search word candidate is configured to select a search word to be output based on the narrowing effect index, so that a document desired by a user who is not included in the first search result is set in a document set. This makes it possible to automatically perform a search, and to improve the efficiency of a narrowed search by the user.

According to the present invention, the suggested search word candidate generating step selects and outputs a plurality of search words, and the search request creating step creates a search request sentence from a logical operation on the plurality of search words. With this configuration, the AND operation can execute a narrowing search more accurately, and the OR operation has an effect that the narrowing search can be executed in a wider range.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a document search device according to a first embodiment of the present invention.

FIG. 2 is a flowchart illustrating an operation of the document search device according to the first embodiment of the present invention.

FIG. 3 is an explanatory diagram illustrating an example of a document vector according to the first embodiment of the present invention.

FIG. 4 is an explanatory diagram showing an example of a search term-topic correspondence table according to Embodiment 1 of the present invention.

FIG. 5 is a block diagram showing a configuration of a document search device according to a second embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a document search device according to a third embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a document search device according to a fourth embodiment of the present invention.

FIG. 8 is an explanatory diagram showing an example in which words and related words are defined in Embodiment 4 of the present invention.

FIG. 9 is a block diagram showing a configuration of a document search device according to a fifth embodiment of the present invention.

FIG. 10 is a configuration diagram showing a conventional document search device.

FIG. 11 is an explanatory diagram showing an example of a document word matrix in a conventional document search device.

[Explanation of symbols]

1 document set, 2 document search unit (document search means), 3
Search result document set, 4 document feature extraction unit (document feature extraction unit), 5 document vector set, 6 topic classification unit (topic classification unit), 7 search word set with importance, 8 presentation search word candidate generation unit (presentation search Word candidate generating means), 9 search word sets with importance, 10 narrowing effect estimating unit (narrowing effect estimating means), 11 narrowing effect index, 12 presented search word candidates, 13 users, 14 classification result presenting unit (classification result presenting means) ), 15 classification results, 16 instruction information, 1
7 search condition, 18 topics, 19 selection information, 20 document feature setting unit (document feature setting means), 21 document vector, 31 topic classification item acquisition unit (topic classification item acquisition means), 32 topic classification item information, 41 designated document Feature extraction unit (designated document feature extraction means), 42 document vectors,
43 documents, 51 related word dictionary (first recording means), 52
Related word setting section (related word setting means), 53 search words, 54
Related terms, 61 Search request creation knowledge (second recording means), 62 Search request creation section (Search request creation means), 63
Search request search word, 64 search request sentence.

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Katsushi Suzuki 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Mitsubishi Electric Corporation F-term (reference) 5B075 ND03 NK02 NK31 NR05 NR12 NS01 PP13 PP22 PQ32 PQ46 UU06

Claims (20)

[Claims] 1. A document retrieval means for retrieving and outputting a document meeting a retrieval condition from a document set, and based on an appearance frequency of a word described in the document.
A document feature extraction unit that calculates a statistical weight of the word and outputs a document vector set obtained from the weight, and creates a topic by classifying the document vector set according to a similarity between document vectors; A topic classification unit that outputs the document information belonging to the topic and the search word with the importance of the topic, and a narrowing effect index of the search word belonging to the topic with reference to the document information and the search word with the importance. A narrowing effect estimating means for calculating and outputting; and providing the narrowing effect index to the search word, selecting the search word having the high narrowing effect index as a suggested search word candidate, and selecting the suggested search word candidate and the suggested search word. A suggested search word candidate generating means for outputting document information corresponding to the word candidate; and associating the suggested search word candidate with the topic and presenting the suggested search word candidate together with the narrowing effect index A classification result presenting means for prompting the user to input either or both of the instruction information and the selection information; and a document feature setting for changing and outputting a document vector included in the document vector set based on the selection information. Document search device comprising: 2. The document search apparatus according to claim 1, wherein the narrowing effect estimating means calculates and outputs a narrowing effect index for at least one or a plurality of topics. 3. The classification result presenting means presents a topic and proposed search word candidates belonging to the topic in the form of a matrix, and presents a narrowing effect index to each element of the matrix. Document search device as described. 4. A word that characterizes the document is extracted as a topic classification item from the document output by the document search unit, and a weight vector for the topic classification item is learned by referring to a text related to the topic classification item. A topic classification item acquiring unit that outputs the topic classification item and the weight vector to a topic classification unit, wherein the topic classification unit classifies a document vector set based on the topic classification item and the weight vector. 2. The document search apparatus according to claim 1, wherein topics are created. 5. The topic classification item acquisition unit according to claim 4, wherein the topic classification item acquisition unit extracts a topic classification item from the document output by the document search unit based on the appearance frequency of a word described in the document. Document search device. 6. The topic classification item acquiring unit extracts a topic classification item from a document output by the document search unit by referring to a tag described in the document.
Document search device as described. 7. A designated document feature extraction unit that calculates a document vector from a document designated via a classification result presentation unit and outputs the document vector to a document feature setting unit, wherein the document feature setting unit includes the designated document feature extraction unit. 2. The document retrieval apparatus according to claim 1, wherein the document vector set is changed based on the document vector output by the means and the document vector set output by the document feature extraction means. 8. A first recording means for defining a word related to a predetermined word and recording the word as a related word, and extracting the related word corresponding to a specified search word from the first recording means to form a document Related word setting means for outputting to a feature setting means, wherein the document feature setting means changes a document vector set based on the related words and selection information input from the classification result presentation means. The document search device according to claim 1. 9. A second recording means for recording knowledge for creating a search request sentence, and a search request sentence corresponding to a search word output by a presented search word candidate generating means is referred to by referring to the second recording means. Search request creating means for creating and outputting to the document search means, wherein the suggested search word candidate generating means selects the search word to be output to the search request creating means based on a narrowing effect index. The document search device according to claim 1, wherein 10. A proposed search word candidate generating means selects a plurality of search words and outputs the selected search words to a search request creating means, and the search request creating means creates a search request sentence from a logical operation on the plurality of search words. 10. The document search device according to claim 9, wherein the search is performed. 11. A document search step for searching and outputting a document meeting a search condition from a document set, and based on an appearance frequency of a word described in the document.
A document feature extraction step of calculating a statistical weight of the word and outputting a document vector set obtained from the weight; creating a topic by classifying the document vector set according to a similarity between document vectors; A topic classification step of outputting document information belonging to the topic and a search word with the importance of the topic; and a narrowing effect index of the search word belonging to the topic by referring to the document information and the search word with the importance. A narrowing effect estimating step of calculating and outputting; assigning the narrowing effect index to the search word; selecting the search word having the high narrowing effect index as a suggested search word candidate; A suggested search word candidate generating step of outputting document information corresponding to the word candidate; and associating the suggested search word candidate with the topic, And a classification result presentation step of presenting together with the result index and prompting to input either or both of the instruction information and the selection information; and changing and outputting the document vector included in the document vector set based on the selection information. A document feature setting step. 12. The document search method according to claim 11, wherein the narrowing effect estimation step calculates and outputs a narrowing effect index for at least one or a plurality of topics. 13. The classification result presenting step, wherein a topic and a proposed search word candidate belonging to the topic are presented in a matrix format, and a narrowing effect index is presented to each element of the matrix. Document search method described. 14. A word that characterizes the document is extracted as a topic classification item from the document output by the document search step,
A topic classification item acquiring step of learning a weight vector for the topic classification item with reference to a text related to the topic classification item and outputting the topic classification item and the weight vector, wherein the topic classification step includes: 12. The document search method according to claim 11, wherein a topic is created by classifying a set of document vectors based on the topic classification item and the weight vector. 15. The topic classification item obtaining step according to claim 14, wherein the topic classification item is extracted from the document output by the document search step based on the appearance frequency of a word described in the document. Document search method. 16. The document search according to claim 14, wherein the topic category item obtaining step extracts topic category items from the document output by the document search step by referring to tags described in the document. Method. 17. A designated document feature extraction step for calculating and outputting a document vector from a document designated via a classification result presentation step, wherein the document feature setting step includes outputting the document output by the designated document feature extraction step. 12. The document search method according to claim 11, wherein the document vector set is changed based on a vector and a document vector set output by the document feature extraction step. 18. A first recording step of defining a word related to a predetermined word and recording it as a related word, and a related word setting step of extracting and outputting the related word corresponding to a specified search word. 12. The document search method according to claim 11, wherein the document feature setting step changes a document vector set based on the related words and the selection information input from the classification result presentation step. 19. A second method for recording the creation knowledge of a search request sentence.
And a search request creating step of creating and outputting a search request sentence corresponding to the search word output by the suggested search word candidate generating step, wherein the suggested search word candidate generating step is based on a narrowing effect index. 12. The document search method according to claim 11, wherein the search word to be output is selected. 20. A suggested search word candidate generating step for selecting and outputting a plurality of search words, and a search request creating step includes:
20. The document search method according to claim 19, wherein a search request sentence is created from logical operations on the plurality of search words.