JP2002041573A - Information retrieval system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the burden on user required for information retrieval and to automatically update an information to be retrieved. SOLUTION: A question and an answer, which are made correspond to each other, are stored in a document storage part 11. A cluster sorting part 16 performs cluster sorting of the answers on the basis of the feature vectors of respective answers in the document storage part 11. When the question of a free sentence is inputted from the user, a database retrieving and updating part 33 retrieves similar questions and presents the correspondent answers to the user or an expert by collecting these answers for every cluster. Further, the database retrieving and updating part 33 automatically updates the document storage part 11 on the basis of the relevant selected answer when the user or the expert selects the most appropriate answer, or on the basis of answer newly inputted by the expert when there is no appropriate answer. The answer of the free sentence inputted by the expert is presented to the user as it is.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information retrieval system for enabling a user to easily find information desired from a large amount of information.

[0002]

2. Description of the Related Art In recent years, with the spread of the Internet, W
HTML (Hyper Text Mar) on WW (World Wide Web)
General users can access a large amount of information, for example, by posting various homepages described in kup Language). FAQ (Freq
A list of frequently asked questions and their answers, referred to as a collection of uently Asked Questions), is disclosed, and users can get answers to the questions. Such information is convenient for the user because it can be viewed immediately if the location of the desired information is known. On the contrary, finding the desired information from a large amount of information is a serious task.

For this reason, a search technique has been known in which a keyword is cut out from a document and used as a feature amount of the document, an inner product between the feature amounts is calculated to obtain a similarity between the documents, and a similar document is searched for a query. ing.

[0004]

However, since information on the Internet or a collection of FAQs stored on a case basis is provided independently by many people, duplication of information can be avoided. And there are many documents with similar contents. Therefore, in the prior art,
In many cases, a large number of documents having similar contents are searched as documents similar to a question sentence, so that the user has to work to find desired information from a large amount of search results. If you limit the search results to a certain number, there is a problem that you do not have the information you want.

[0005] Further, even if the user succeeds in finding the desired information from the search results, the information is not reflected in the FAQ collection. Therefore, even when another user performs a search under the same conditions, the same trouble of finding the same is necessary. Met. In order to further enhance the FAQ collection while avoiding duplication of information, it was necessary to check whether or not similar information already existed, which was a burden on the information provider.

[0006] It is an object of the present invention to provide an information retrieval system which reduces the burden on a user for information retrieval.

Another object of the present invention is to provide an information retrieval system that can easily update information to be retrieved.

[0008]

In order to achieve the above object, the present invention calculates a feature vector of a document, classifies the documents into clusters based on the feature vectors, and collectively displays the search results of the documents for each cluster. It is decided to do. This makes it easy for the user to grasp the search result as a set of similar documents.

Further, according to the present invention, when a question from a user is input, a similar question is searched, and a corresponding answer is presented to the user or expert, and the user or expert is determined to be most appropriate. When the determined answer is selected, the document database is automatically updated based on the selected answer. If there is no appropriate answer, the expert updates the document database based on the newly entered answer. Thereby, when the same question is input from the next time, an appropriate answer can be made.

[0010]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, two embodiments of the present invention will be described with reference to the drawings.

<< First Embodiment >> FIG. 1 shows a first embodiment of the present invention.
1 shows a configuration of an information search system according to an embodiment. Figure 1
The information retrieval system includes a document storage unit 11, a cluster storage unit 12, a cluster label storage unit 13, a document label storage unit 14, a feature vector extraction unit 15, a cluster classification unit 16, a cluster label creation unit 17 A document label creating unit 18, a database searching unit 19, an interface unit 20, a user input unit 21, and a user display unit 22.
And is realized by, for example, a document server and a user terminal connected to each other via the Internet. The document storage unit 11 stores a plurality of documents. The feature vector extraction unit 15 extracts a feature vector from a document stored in the document storage unit 11. The cluster classification unit 16 classifies documents stored in the document storage unit 11 into clusters based on the feature vectors obtained by the feature vector extraction unit 15. The cluster storage unit 12 stores clusters of documents classified by the cluster classification unit 16. The cluster label creating unit 17 creates, for each cluster created by the cluster classifying unit 16, a cluster label representing the content of the cluster. The cluster label is a word label composed of words or a sentence label composed of sentences. The cluster label storage unit 13 stores the cluster label created by the cluster label creation unit 17. Document label creation unit 18
Creates, for each document that is an element of the cluster created by the cluster classification unit 16, a document label representing the content of the document. The document label storage unit 14 stores the document label created by the document label creation unit 18. User input section 2
1 receives a search condition given by a user. The search condition may be anything that can be a document search condition, such as a document keyword or a document ID. The interface unit 20 manages input and output with the user. The database search unit 19 searches the document storage unit 11 for documents that satisfy the search condition. The user display unit 22 presents a search result to the user.

FIG. 2 shows an example of a document stored in the document storage unit 11 in FIG. The document storage unit 11 stores given n (n ≧ 2) documents to be searched. Each document is composed of a unique document ID and a text body. Let the i-th document be Di (1 ≦ i ≦
n).

FIG. 3 shows a display example of a search result on the user display unit 22 in FIG. According to FIG. 3, the search results of documents for a certain search condition are displayed collectively for each cluster. Specifically, the cluster ID, the document ID of the document included in the cluster, and the body of the document are displayed in a table format for each cluster. By displaying the cluster of, all search results can be displayed. This makes it easy for the user to grasp the search result as a set of similar documents. Moreover, in the displayed cluster, a cluster label indicating the content of the cluster is displayed, and a sentence specified in the document label is displayed with an underline. Therefore, the user can easily grasp the contents of the cluster. Although the cluster ID and the document ID are also displayed as the search result, they need not be displayed.

Hereinafter, the details of the first embodiment will be described separately for a document registration operation and a document search operation. The document registration operation is an operation when a document is registered in the document storage unit 11 for the first time or when a document is added / changed / deleted thereafter. The document search operation is an operation when a registered document is searched and browsed.

<Operation at Document Registration> FIG. 4 shows a processing procedure of the feature vector extracting unit 15 in FIG. First, the feature vector extraction unit 15 sequentially retrieves all documents Di stored in the document storage unit 11 and extracts a feature vector Vi of each document Di. The feature vector is a vector having a set of a word Tj representing the feature of the document and its weight Wij as an element, and the number of elements differs depending on the document. here,
j is a unique number that identifies a word. In FIG. 4, in step S101, the counter i of the document is set to i = 1.
Set. In step S102, the document Di is extracted from the document storage unit 11, and the appearing word Tj is extracted from the body by a generally known method such as morphological analysis, syntax analysis, or unnecessary word removal, and the word Tj in the document Di is extracted. Is counted. End determination step S10
In 3, if the processing of step S102 is completed for all documents, that is, if i = n, the process proceeds to S105. If not, the process proceeds to S104. In step S104, the counter i is incremented by 1, and the process proceeds to step S102. In step S105, as a degree of importance of the word Tj with respect to all the documents, a degree indicating a small number of documents in which the word Tj appears, that is, an IDF (inverse document frequency) value is calculated by the equation (1).

[0016]

(Equation 1)

Here, Mj represents the number of documents in which the word Tj appears. In step S106, i = 1 is set in the counter i of the document. In step S107, the TFIDF value obtained by multiplying the IDF value by the TF (term frequency) value representing the appearance ratio of the word Tj in the document Di as the weight Wij that characterizes the document Di by the word Tj is given by the following equation (2). It is calculated by:

[0018]

(Equation 2)

In the end determination step S108, when the processing in step S107 has been completed for all documents, that is, when i = n, the processing ends. Otherwise, S10
Go to 9. In step S109, the counter i is incremented by 1, and the process proceeds to step S107.

FIG. 5 shows the extracted document feature vector Vi.
Here is an example. Note that the calculation of the feature vector
Although the DF value has been used, another method such as simply setting the number of appearances of a word may be used.

FIG. 6 shows a processing procedure of the cluster classification unit 16 in FIG. The cluster classification unit 16 classifies all documents into m clusters using the feature vectors extracted by the feature vector extraction unit 15 (1 <m <n). here,
The k-th cluster is Ck (1 ≦ k ≦ m). As a procedure of cluster classification, hierarchical clustering in which clusters are sequentially classified in a tree diagram is used. FIG.
In step S111, an initial calculation of the inter-cluster distance is performed. Here, each of the initial clusters is 1
Clusters Ci having only one document Di as an element
Set. Each cluster Ck, Cl (1 ≦ k, l ≦ n)
As the distance Lkl, the similarity ratio of Expression (3) representing the distance between the feature vectors of each document is adopted.

[0022]

(Equation 3)

In step S112, i = 1 is set in a counter i for the number of times of clustering. Step S113
Then, among the combinations of all clusters, the clusters Ck, Cl (k <l) with the smallest intercluster distance Lkl
Search for a set of. In step S114, the cluster C
k and Cl are integrated into a cluster Cg. That is, C
Let g = Ck∪Cl, Cl = φ (φ represents an empty set). With the integration of the clusters, the inter-cluster distance between the cluster Cg and another cluster Ch (1 ≦ h ≦ n) is calculated by the equation (4) using the Ward method.

[0024]

(Equation 4)

Here, Nk is the number of elements of the cluster Ck. In the end determination step S115, when the number of clustering is n-1, that is, when all the initial clusters are 1
If they are integrated into one cluster, the process proceeds to step S117. If not, the process proceeds to S116. Step S
At 116, the counter i is incremented by 1, and the process proceeds to step S112. In step S117, the number of clusters is determined.
In the cluster classification process from step S111 to step S115, the number of clusters decreases by one for each clustering count. In step S117, the cluster classification process is reviewed, and an appropriate number of times of clustering is determined. Here, it is assumed that the number of times of clustering that maximizes the number of clusters having two or more elements is an appropriate number of times of clustering. In step S118, the elements included in each cluster at the time of performing the cluster classification up to the number of times of clustering determined in step S117 are written to the cluster storage unit 12.

FIG. 7 shows an example of a cluster written to the cluster storage unit 12. Each cluster has a cluster ID
And the document ID of the document included in the cluster. For example, cluster 1, 1,190,432,6
It contains four documents, number 44. This is these four
This indicates that the feature vectors of one document are similar to those of another document. In the above example, hierarchical clustering is used as a method of cluster classification.
Non-hierarchical clustering may be used. Although the similarity ratio of Expression (3) is used as the initial inter-cluster distance, another distance such as a Euclidean square distance may be used. Although the Ward method of Expression (4) is used as a method of calculating the distance between clusters at the time of cluster integration, another method such as the longest distance method may be used. As a method of determining the number of clusters, the number of clusterings is set such that the number of clusters having two or more elements is maximized.

FIG. 8 shows the cluster label creating unit 1 in FIG.
7 shows a word label creation procedure. Step S20
At 1, k = 1 is set to the counter k of the cluster. In step S202, for each word Tj included in the feature vector Vi of every document Di that is an element of the cluster Ck, the word Tj of the document Di that is an element of the cluster Ck
The number of appearing documents in which j appears is counted. Step S2
03, for each word Tj included in all documents Di that are elements of the cluster Ck, the TFIDF value of the word Tj (=
Wij), all documents Di that are elements of cluster Ck
Is calculated. In step S204, all the words Tj included in the feature vectors Vi of all the documents Di, which are the elements of the cluster Ck, are sorted in descending order of the number of appearing documents obtained in step S202. If the number of appearing documents is the same, the documents are sorted in descending order of the sum of the TFIDF values obtained in step S203. In step S205, the top three words sorted in step S204 are selected and written to the cluster label storage unit 13 as cluster word labels. In the end determination step S206, when the processing from step S202 to step S205 has been completed for all clusters, that is, when k = m, the processing ends. Otherwise, the process proceeds to S207. In step S207, the counter k is incremented by 1 and step S202
Proceed to.

FIG. 9 shows an example of a word label written in the cluster label storage 13. For example, it indicates that the cluster 1 has a word label of “candy”, “snacks”, and “cheese”. Although the method of creating the word labels is based on the number of documents in which the words appear, other methods such as sorting only the TFIDF values may be used. Further, the number of words in the word label is set to three, but may be other than three.

FIG. 10 shows a sentence label creating procedure in the cluster label creating section 17 in FIG. Step S30
At 1, k = 1 is set to the counter k of the cluster. In step S302, for each word Tj included in the feature vector Vi of every document Di that is an element of the cluster Ck, the word T
The number of appearing documents in which j appears is counted. Step S3
In step 03, for each sentence constituting all the documents Di which are elements of the cluster Ck, the sum of the number of appearing documents counted in step S302 of the word Tj included in the sentence is calculated. Here, a sentence refers to each character string obtained by separating a document with a period such as “.”. In step S304, the sentences constituting all the documents Di that are elements of the cluster Ck are sorted in descending order of the total number of appearing documents obtained in step S303. In step S305, step S305
The top sentence sorted in 304 is selected and written to the cluster label storage unit 13 as a sentence label of the cluster.
If there are a plurality of top-level sentences, a sentence with the smallest number of characters is selected from among them. In the end determination step S306, when the processing from step S302 to step S305 is completed for all clusters, that is, when k = m, the processing ends. If not, the process proceeds to S307. In step S307, the counter k is incremented by 1 and step S302
Proceed to.

FIG. 11 shows an example of a sentence label written in the cluster label storage unit 13. For example, it is indicated that the cluster 1 has a sentence label of "a substance with a lot of water (jelly, pudding, yogurt) ...". Although the sentence label is created based on the total number of documents in which the words appear, other methods such as sorting based on the total TFIDF values may be used. Further, when there are a plurality of sentences having the highest number of appearing documents, the sentence with the smallest number of characters is selected, but another method such as selecting the sentence with the start position of the sentence being the foremost may be used.

FIG. 12 shows the document label creating section 18 in FIG.
The following shows the processing procedure. In step S401, i = 1 is set in a document counter i. In step S402, a total of TFIDF values (= Wij) of all words Tj included in the sentence is calculated for each sentence constituting the document Di. In end determination step S403, step S403 is performed for all documents.
When the process of 402 is completed, that is, when i = n, the process proceeds to S405. If not, the process proceeds to S404. In step S404, the counter i is incremented by 1, and the process proceeds to step S402. In step S405, k = 1 is set to the cluster counter k. In step S406, the sentences constituting all the documents Di which are the elements of the cluster Ck are sorted in descending order of the total obtained in step S402. In step S407, the top sentence sorted in step S406 is selected as the document label of the document Di. However, when the selected sentence is the same as the sentence label of the cluster created by the cluster label creating unit 17, the second highest sentence sorted in step S406 is selected as the document label of the document Di. Step S
In 408, the document label of the document Di selected in step S407 is written to the document label storage unit 14. In the end determination step S409, the step S
When the process from 406 to step S408 is completed, that is, when k = m, the process ends. Otherwise, the process proceeds to S410. In step S410, the counter k is incremented by 1, and the process proceeds to step S406.

FIG. 13 shows an example of a document label written in the document label storage unit 14. For example, it indicates that the document 1 included in the cluster 1 has a document label of “there is a crispness and is not left behind,...”.

According to the above operation, at the time of document registration, a feature vector is extracted for each document, and a cluster, a cluster label, and a document label are created and stored in the respective storage units.

<Operation During Document Search> First, the interface unit 20 receives a document search condition through the user input unit 21. The database search unit 19 searches the document storage unit 11 for a document that satisfies the search condition, searches the cluster storage unit 12 for a cluster including the searched document, and returns the document included in the searched cluster to the document again. The search is performed from the storage unit 11 and the result is sent to the interface unit 20 together with the cluster label and the document label. The interface unit 20 presents the search result to the user through the user display unit 22 (FIG. 3).

In this embodiment, a given document is stored in advance. However, the document may be newly introduced or revised later using a storage medium such as an optical disk or a network medium such as the Internet. . Further, the document search may be a full-text search or a fuzzy search other than the search using the keyword or the document ID.

<< Second Embodiment >> FIG. 14 shows the configuration of an information retrieval system according to a second embodiment of the present invention. The information search system in FIG. 14 is a system that returns an appropriate answer based on past case searches to a user's question in a free sentence, such as a document server and a user terminal connected to each other via the Internet. And an expert terminal. The configuration of FIG. 14 is different from the configuration of FIG. 1 in that the feature vector storage unit 31, the similarity calculation unit 32, and the expert input unit 41
And an expert display section 42, and the database search section 19 in FIG. The document storage unit 11 stores a plurality of question documents and a plurality of answer documents associated with each other.
The expert display unit 42 presents a search result to an expert. The expert input unit 41 receives a selection input from a specialist and a response input in a free sentence. The interface unit 20
Manage input and output with users and experts. The feature vector extraction unit 15 has a function of extracting a feature vector from each of the question document and the answer document in the document storage unit 11, a function of extracting a feature vector from a user's free text question input, and an expert free sentence. And a function of extracting a feature vector from an answer input by the The feature vector storage unit 31
The feature vector extraction unit 15 stores the feature vectors extracted from each of the question document and the answer document in the document storage unit 11. The similarity calculation unit 32 has a function of calculating the similarity between the feature vector extracted from the user question input and the feature vector of the question document stored in the feature vector storage unit 31, and the similarity calculation unit 32 And a function of calculating the similarity between the extracted feature vector and the feature vector of the answer document stored in the feature vector storage unit 31. The database search updating unit 33 has a function of updating the document storage unit 11 based on a response from a user or an expert, in addition to a function of searching for a document in the document storage unit 11.

FIGS. 15 and 16 show examples of documents stored in the document storage unit 11 in FIG. FIG. 15 shows a part of a questionnaire in which question documents are collected. This questionnaire includes a unique question ID, a question in text format, and an answer ID corresponding to the question. FIG. 16 shows a part of an answer table in which answer documents are collected. This answer table is unique answer I
D, and answer in text format. Qi for i-th question
And the k-th answer is Ak (1 ≦ i ≦ n and 1 ≦
k ≦ m). Here, the relationship of n ≧ m holds. That is, one answer may correspond to a plurality of questions.

FIG. 17 shows a display example of the search result on the expert display section 42 in FIG. In FIG. 17, in addition to the question from the user, the answer candidates are displayed in a state where they are classified into clusters, together with the sentence labels of the clusters and the document labels in the clusters. In FIG. 17, all search results can be displayed by pressing the "previous page" button or the "next page" button with the mouse to display another page. Thus, the expert can easily select the most appropriate answer by referring to the search result displayed as a group of similar documents. Alternatively, an expert answer in a free text can be input.
Although the sentence label is displayed as the cluster label in the example of FIG. 17, a word label may be displayed together with or instead of this. Also, as a search result, cluster I
D, the document ID is also displayed, but need not be displayed.

FIG. 18 shows a display example of search results on the user display section 22 in FIG. Here, it is assumed that the document with the number 1 is selected as the expert answer.

The details of the second embodiment will be described below with reference to the first embodiment.
In the same manner as in the first embodiment, the operation during document registration and the operation during document search will be described separately.

<Operation at the time of document registration> First, the feature vector extraction unit 15 extracts and extracts the feature vector VQi of the question and the feature vector VAk of the answer from all the documents stored in the document storage unit 11. The feature vector is written to the feature vector storage unit 31. The procedure for extracting the feature vector is the same as in the first embodiment. The difference from the first embodiment is that a feature vector is calculated for each of a question and an answer, and the feature vector is stored in a feature vector storage unit 31.
It is a point to write out.

Next, the cluster classification unit 16 reads the answer feature vector VAk from the feature vector storage unit 31, classifies all answer documents into clusters, and writes the clusters into the cluster storage unit 12. The procedure of cluster classification is the same as in the first embodiment. The difference from the first embodiment is that cluster classification is performed using the answer feature vector VAk. The operations of the cluster label creating unit 17 and the document label creating unit 18 are the same as in the first embodiment.

With the above operation, at the time of document registration, a feature vector is extracted for each of a question and an answer, and a cluster, a cluster label, and a document label are created for the answer and stored in their respective storage units.

<Operation at Document Retrieval> First, the interface unit 20 receives a user question Q in a free sentence through the user input unit 21. The feature vector extraction unit 15
The feature vector VQ of the user question is extracted.

FIG. 19 shows a procedure for extracting a feature vector of a user question in the feature vector extracting unit 15 in FIG. In step S501, the word Tj that appears is extracted from the user question Q, and the number Fi of occurrences of the word Tj in the document is extracted.
Count j. The method of extracting words is the same as in the first embodiment. In step S502, the IDF of the word Tj
Calculate the value. If the word Tj exists in any of the documents in the document storage unit 11, the IDF value has already been calculated at the time of document registration, and is used in step S502. If the word Tj does not exist, the IDF value (IDFj) of the word Tj is calculated by Expression (5).

[0046]

(Equation 5)

In step S503, the weight WQj (TFIDF value) of the word Tj in the user question Q is calculated. The method of calculating the TFIDF value is the same as in the first embodiment. FIG. 20 shows an example of the feature vector VQ extracted from the user question Q.

Next, the similarity calculation unit 32 extracts the feature vectors VQi of all questions from the feature vector storage unit 31 and calculates the similarity between these feature vectors VQi and the feature vector VQ of the user question.

FIG. 21 shows a processing procedure of the similarity calculating section 32 in FIG. In step S511, i = 1 is set in the counter i of the document. In step S512, the feature vector VQi and the feature vector VQ of the question from the user
Is calculated as the inner product of the vectors by the equation (6).

[0050]

(Equation 6)

In the end determination step S513, if the processing of step S512 has been completed for all questions, that is, if i = n, the flow proceeds to S515. Otherwise, the process proceeds to S514. In step S514, the counter i is incremented by 1, and the process proceeds to step S512. In step S515, all the question documents are sorted in descending order of the similarity Ei obtained in step S512.

Next, the database retrieval / updating unit 33 stores the predetermined number of question documents having the higher similarity Ei calculated by the similarity calculator 32 and the corresponding answer documents in the document storage unit 1.
1, the cluster including the searched answer document is searched from the cluster storage unit 12, the answer document included in the searched cluster is searched again from the document storage unit 11, and the result is referred to as the cluster label and The document is sent to the interface unit 20 together with the document label. Although the inner product of the vectors is used as the method of calculating the similarity of the feature vectors, another method such as using the similarity ratio of the vectors may be used.

Next, the interface unit 20 presents the answer part of the search result to the expert through the expert display unit 42 (FIG. 17), and refers to the display of the expert display unit 42 through the expert input unit 41. Accepts expert's choice of answer or input of free text answer. Further, the interface unit 2
0 indicates an expert answer to the user through the user display unit 22 (FIG. 18). Therefore, only useful information is presented to the user.

FIG. 22 is a flowchart showing the processing procedure of the database search / update unit 33 in FIG. In step S601, an answer case search display is performed. Specifically, the interface unit 20 receives the user question Q in a free sentence, and presents the search result to the expert through the expert display unit 42 (FIG. 17). In step S602,
Judge the search results. The expert looks at the display of FIG. 17 and determines whether or not there is an answer that seems appropriate to the user question Q. If there is an answer deemed appropriate, the process proceeds to S603. If there is no answer deemed appropriate, the process proceeds to S606. Step S60
In 3, the expert selects the document ID of the answer deemed most appropriate for the user question Q. The interface unit 20 receives an input of the selected document ID through the expert input unit 41. Further, the document ID is transferred to the database search / update unit 33 for step S605 to be described later. In step S604, the interface unit 20 presents the document of the document ID selected by the expert to the user through the user display unit 22 as a response (FIG. 18).

In step S605, a question adding process is performed. If the similarity of the question having the highest similarity to the user question Q is one or more of the one or more questions corresponding to the answer of the received document ID, the database search updating unit 33 determines that the similarity is equal to or less than a predetermined value. Assuming that an appropriate automatic answer has not been made, a new unique question ID,
A line consisting of the user question Q and the selected document ID is added. Next, the process proceeds to step S612. Step S612
Then, the feature vector extraction unit 15 extracts the feature vectors VQi and VA from all the questions Qi and answers Ak stored in the document storage
k is extracted, and the extracted feature vector is written to the feature vector storage unit 31.

If there is no appropriate answer in step S602, the expert inputs an appropriate answer A to the user question Q in a free sentence in step S606. The interface unit 20 accepts a free sentence answer through the expert input unit 41. In step S607, the interface unit 20 presents the answer A input by the expert to the user. In step S608, the feature vector extraction unit 15
Extracts the feature vector VA of the answer A input by the expert. The procedure for extracting the feature vector is the same as the procedure for extracting the feature vector VQ of the user question Q described with reference to FIG. In step S609, the similarity calculation unit 32 stores the feature vectors V of all the answers from the feature vector storage unit 31.
Ak is taken out, and the similarity Ek of the answer A input by the expert with the feature vector VA is calculated. The procedure for calculating the similarity is the same as the procedure for calculating the similarity of the user question Q described with reference to FIG. In step S610, the similarity calculating unit 32 determines that the largest one of the similarities Ek obtained in step S609 is equal to or greater than a predetermined value, and the document storage unit 11
It is determined that there is an answer similar to the answer A input by the expert, and the document ID of the similar answer Ak is passed to the database search and update unit 33, and the process proceeds to step S605. Otherwise, the process proceeds to step S611. In step S611, a question answer addition process is performed. The database search / update unit 33 adds a row including the new unique document ID and the answer A input by the expert to the answer table in FIG. In addition, a row including a new unique question ID, a user question Q, and a document ID assigned to the added answer is added to the questionnaire in FIG. And step S
Proceed to 612. The processing in step S612 is as described above.

On the other hand, when there is no expert who can select or input an answer, the interface unit 20 presents the same search result as that of FIG. 17 to the user through the user display unit 22. The user looks at the display in FIG.
The interface unit 20 receives the input of the selected document ID through the user input unit 21 by selecting the document ID of the answer that seems to be the most appropriate. When the similarity of the question having the highest similarity to the user question Q is one or more of the questions corresponding to the answer of the input document ID, the database search updating unit 33 determines that Assuming that no appropriate automatic answer was given,
A row including a new unique question ID, a user question Q, and a selected document ID is added to the questionnaire of No. 5 (similar to step S605). Then, the feature vector extraction unit 1
5 extracts the feature vectors VQi and VAk from all the questions Qi and answers Ak stored in the document storage unit 11 and stores the extracted feature vectors in the feature vector storage unit 31 in the same manner as in the document registration. (Similar to step S612).

As described above, according to the second embodiment,
Since the document storage unit 11 is automatically updated in response to a response from a user or an expert, it is possible to provide an information search system that can provide an appropriate answer when a similar question is input from the next time. it can.

[0059]

As described above, according to the present invention, a feature vector of a document is calculated, documents are classified into clusters based on the feature vector, and the search results of the documents are displayed collectively for each cluster. Therefore, the user can easily grasp the search result as a set of similar documents. Therefore, it is possible to provide an information search system that reduces the burden on the user for information search.

Further, according to the present invention, when a question from a user is input, a similar question is searched and a corresponding answer is presented to the user or expert, and the user or expert is most appropriate. The document database is automatically updated based on the selected answer when the answer is judged to be present, or based on the newly entered answer when there is no appropriate answer. Therefore, it is possible to provide an information search system capable of easily updating information to be searched.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an information search system according to a first embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a document stored in a document storage unit in FIG. 1;

FIG. 3 is a diagram showing a display example of a search result on a user display unit in FIG. 1;

FIG. 4 is a flowchart illustrating a processing procedure of a feature vector extracting unit in FIG. 1;

FIG. 5 is a diagram illustrating an example of an extracted document feature vector.

FIG. 6 is a flowchart illustrating a processing procedure of a cluster classification unit in FIG. 1;

FIG. 7 is a diagram illustrating an example of a cluster classification result.

FIG. 8 is a flowchart showing a word label creation procedure in a cluster label creation unit in FIG. 1;

FIG. 9 is a diagram illustrating an example of a created word label.

FIG. 10 is a flowchart showing a sentence label creating procedure in a cluster label creating unit in FIG. 1;

FIG. 11 is a diagram illustrating an example of a created sentence label.

FIG. 12 is a flowchart illustrating a processing procedure of a document label creating unit in FIG. 1;

FIG. 13 is a diagram illustrating an example of a created document label.

FIG. 14 is a block diagram illustrating a configuration of an information search system according to a second embodiment of the present invention.

FIG. 15 is a diagram illustrating an example of a questionnaire portion of a document stored in the document storage unit in FIG. 14;

FIG. 16 is a diagram illustrating an example of a part of an answer table in a document stored in a document storage unit in FIG. 14;

FIG. 17 is a diagram showing a display example of a search result on an expert display unit in FIG. 14;

FIG. 18 is a diagram showing a display example of a search result in a user display unit in FIG.

FIG. 19 is a flowchart showing a procedure for extracting a feature vector of a user question in a feature vector extracting unit in FIG. 14;

FIG. 20 is a diagram illustrating an example of a feature vector extracted from a user question.

FIG. 21 is a flowchart illustrating a processing procedure of a similarity calculation unit in FIG. 14;

FIG. 22 is a flowchart mainly showing a processing procedure of a database search / update unit in FIG. 14;

[Explanation of symbols]

 Reference Signs List 11 document storage unit 12 cluster storage unit 13 cluster label storage unit 14 document label storage unit 15 feature vector extraction unit 16 cluster classification unit 17 cluster label creation unit 18 document label creation unit 19 database search unit 20 interface unit 21 user input unit 22 User display unit 31 Feature vector storage unit 32 Similarity calculation unit 33 Database search and update unit 41 Expert input unit 42 Expert display unit

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G06F 17/30 350 G06F 17/30 350C (72) Inventor Hiroshi Kutsumi 1006 Kadoma Kadoma, Kadoma City, Osaka Matsushita Electric Industrial Co., Ltd. (72) Inventor Jun Ozawa 1006 Kadoma Kadoma, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Inventor Susumu Susumu Maruno 1006 Kadoma Kadoma, Osaka Pref. 1006 Matsushita Electric Industrial Co., Ltd. (Reference) 5B075 KK07 ND03 NK32 NR02 NR12 NR20 PP24 PQ02 PQ41 PQ46 PR06 QM08

Claims (18)

[Claims] 1. An information retrieval system for retrieving information desired by a user from a plurality of documents, comprising: a document storage means for storing the plurality of documents; and a document storage means for storing the plurality of documents. A feature amount extraction unit for extracting a feature amount of each of the plurality of documents; and a plurality of documents stored in the document storage unit, each cluster of which is a single document or approximate to each other, based on the extracted feature amounts. So that it consists of multiple documents with
Cluster classification means for classifying into a plurality of clusters; and a document search means for searching a document satisfying a search condition given by the user from a plurality of documents stored in the document storage means, Interface means for presenting the searched document together with other documents in the cluster when the cluster to which the searched document belongs is composed of a plurality of documents, as a search result, Information retrieval system. 2. The information retrieval system according to claim 1, wherein the feature amount extracting unit includes one or more words appearing in the document from each of the plurality of documents stored in the document storage unit; An information retrieval system characterized by extracting a feature vector having a set of a word and a weight that characterizes the document as an element. 3. The information retrieval system according to claim 1, wherein said cluster classification means is configured to employ clustering that maximizes the number of clusters composed of a plurality of documents. . 4. The information retrieval system according to claim 1, further comprising: cluster label creating means for creating a plurality of cluster labels each representing the contents of a corresponding cluster among the plurality of clusters, wherein said interface means Is an information retrieval system configured to present, together with the search result, a cluster label representing the contents of a cluster to which the searched document belongs, among the plurality of created cluster labels. 5. The information retrieval system according to claim 4, wherein, for each of the plurality of clusters, one or a plurality of words characterizing the cluster from among all the documents belonging to the cluster. Is selected as the cluster label. 6. The information retrieval system according to claim 4, wherein the cluster label creating unit is configured to, for each of the plurality of clusters, select one sentence characterizing the cluster from all documents belonging to the cluster. An information retrieval system, which is configured to select a cluster label. 7. The information retrieval system according to claim 4, wherein a plurality of document labels are generated for each of the plurality of documents stored in the document storage unit, the plurality of document labels representing the contents of the corresponding document. The interface unit is configured to present a document label representing the contents of each document in a cluster to which the searched document belongs, among the created document labels, together with the search result. An information retrieval system characterized by the following. 8. The information retrieval system according to claim 7, wherein the document label creating unit is configured to, for each of the plurality of documents stored in the document storage unit, select the document from all the sentences in the document. An information retrieval system characterized in that one sentence characterizing is selected as the document label. 9. The information search system according to claim 1, wherein the plurality of documents include a plurality of question documents and a plurality of answer documents which are associated with each other, and the search condition is a user question in a free sentence. The feature amount extracting means comprises: a plurality of answer documents stored in the document storage means such that the plurality of answer documents stored in the document storage means are classified into a plurality of clusters by the cluster classification means. The information retrieval system extracts a feature amount of each of the plurality of question documents stored in the document storage means, and calculates a similarity between the plurality of question documents and the document related to the user question. Further comprising a degree calculation means, wherein the document search means is based on the calculated similarity degree,
Search for a question document having a high degree of similarity from among the plurality of question documents stored in the document storage means, and associate with the searched question document from among the plurality of answer documents stored in the document storage means When the retrieved answer document is searched, the interface means, together with the other answer documents in the cluster, if the retrieved answer document belongs to a cluster including a plurality of answer documents, An information search system characterized in that it is configured to be presented as the search result. 10. The information search system according to claim 9, wherein said interface means is configured to present said search result to said user. 11. The information search system according to claim 10, wherein the interface unit is configured to receive a selection of the answer document by the user from the presented search results. Searching for a question document associated with the selected answer document from among the plurality of question documents stored in the document storage unit, and comparing the searched question document with the document related to the user question. When the similarity is lower than a predetermined value, a document updating unit for newly storing the document related to the user question in the document storage unit in association with the selected answer document is further provided. An information retrieval system characterized by the following. 12. The information search system according to claim 9, wherein the interface unit presents the search result to an expert together with a document related to the user question, and selects the specialized search result from the presented search result. An information retrieval system, wherein an answer document selected by a house is presented to the user. 13. The information search system according to claim 12, wherein a question document associated with the selected answer document is searched from a plurality of question documents stored in the document storage unit. If the similarity between the question document and the document related to the user question is lower than a predetermined value, the document related to the user question is associated with the selected answer document and stored in the document storage unit. An information retrieval system, further comprising a document updating unit for newly storing the document. 14. The information search system according to claim 9, wherein the interface unit presents the search result to an expert together with a document relating to the user question, and refers to the presented search result to execute the search. An information retrieval system, characterized in that an answer document input by an expert in a free sentence is presented to the user. 15. The information retrieval system according to claim 14, wherein a similarity between each of the plurality of answer documents stored in the document storage unit and the input answer document is lower than a predetermined value. The information retrieval system according to claim 1, further comprising a document updating unit for associating the document relating to the user question and the input answer document with each other and newly storing the document in the document storage unit. 16. An information retrieval system for retrieving information required by a user from a plurality of documents, the document storage for storing a plurality of question documents and a plurality of answer documents associated with each other. Means, when a user question in a free sentence is given by the user, for each of a plurality of question documents stored in the document storage means, a similarity between the document related to the user question and A similarity calculating means for calculating, based on the calculated similarity, searching a plurality of question documents having a high similarity from a plurality of question documents stored in the document storage means, and storing the document A document search unit for searching an answer document associated with each of the plurality of searched question documents from a plurality of answer documents stored in the means; and the search together with a document related to the user question. The plurality of answer documents presented to the expert as search results, and the expert referring to the answer document selected by the expert from the presented search results or the presented search result. And an interface means for presenting to the user a response document input by the user in a free sentence. 17. The information retrieval system according to claim 16, wherein a question document associated with the selected answer document is searched from a plurality of question documents stored in the document storage unit. When the similarity between the question document and the document related to the user question is lower than a predetermined value, the document related to the user question is associated with the selected answer document and stored in the document storage unit. An information retrieval system, further comprising a document updating unit for newly storing the document. 18. The information retrieval system according to claim 16, wherein a similarity between each of the plurality of answer documents stored in the document storage unit and the input answer document is lower than a predetermined value. The information retrieval system according to claim 1, further comprising a document updating unit for associating the document relating to the user question and the input answer document with each other and newly storing the document in the document storage unit.