JP2006309377A - Document retrieval device, document retrieval method, its program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a document retrieval device capable of automatically creating a related term dictionary, and capable of further improving accuracy of document retrieval by detecting a related term of a word included in a retrieval sentence more accurately than conventional. <P>SOLUTION: Vector retrieval of a retrieval target document is carried out by a retrieval sentence, and a retrieval sentence in which a word is replaced by a related term. A word included in a selected subject in a retrieval result, and a word included in the retrieval sentence are used to automatically generate the related term dictionary database. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a document search apparatus, a document search method, a program thereof, and a recording medium that output a search target document related to an input search sentence.

In recent years, interest in document search systems has increased due to the spread of technology for accessing a Web server connected to the Internet and displaying Web pages on a terminal, and digitization of a large number of documents. In a document search system, a user can obtain a desired document by inputting a search sentence or a search term. By the way, in a document search system, when a user searches with a search sentence including a word that is not included in a document, problems such as a search failure or an inappropriate search result occur. In order to solve such a problem, some document search systems have a related word dictionary to deal with this problem. In such a document search system, when a word not included in the search target document appears in a search sentence entered by a user, a related word of the word is acquired from the related word dictionary, and the word is replaced with the related word. , Add or perform processing to supplement the search sentence. Here, since the task of manually adding a related word to this related word dictionary is laborious, a method of adding the related word semi-automatically or automatically is desired. As a method for automatically creating a related word dictionary, Patent Document 1 is disclosed.
JP 11-31168 A

  Here, in the above-described synonym calculation device and program in Patent Document 1, a search history is acquired, a relevance degree is obtained based on a time interval and a frequency of words included in the history, and words having a high relevance degree are related words. Suggest a way to register with. If this method is incorporated into a document search system, a related word dictionary can be automatically created to supplement the search sentence. However, with this method, words that are close in time interval are simply associated with each other, so the words included in the search sentence when the search fails or when the desired search result cannot be obtained are also associated with each other. As a result, there is a problem that the search result is not improved.

  Therefore, the present invention is a document search apparatus that can automatically create a database of related word dictionaries, and by detecting related words of words included in a search sentence more accurately than before, further accuracy of document search can be improved. An object of the present invention is to provide a document search apparatus and a document search method, a program thereof, and a recording medium.

  The present invention has been made to solve the above-described problem, and is a document search apparatus that outputs a search target document related to an input search sentence, and is included in the search sentence as already learned information. A relationship that stores information on the correspondence between related words and related words related to the related words, and stores the association level indicating the strength of the relationship between the related words and the related words. A word storage unit and a search word replacement processing unit for replacing the related word with the related word stored in the related word storage unit among the words included in the search sentence; Among the search sentences after the replacement, for the replaced word, a search document vector of each word of the search sentence after the replacement is generated using a value obtained by multiplying the weight of the word by the relevance, Included in search text after replacement A target document vector is generated based on a weight value of each word in the search target document, and a predetermined number of search targets having a small angle formed according to an angle formed by the search document vector and the target document vector A search target document extraction processing unit that extracts a document from a plurality of search target documents stored in the search target document storage unit, a subject of the extracted search document, the search document vector and the target document vector A document search apparatus comprising: a search result screen generation processing unit that generates search result screen data that can be sequentially displayed according to an angle formed by; and a search result screen output processing unit that outputs the search result screen data. It is.

  According to the present invention, replacement with related words learned to search for a search target document with higher accuracy is performed. Search processing is performed by the search sentence after the replacement. Thereby, it is possible to perform a search process with higher accuracy than in the past.

  Further, the present invention is based on at least the display instruction of the search target document received by selection of the subject, whether any subject is selected from the subject of the search target document displayed on the output search result screen. A subject selection determination processing unit to be detected, a same session determination processing unit that determines whether or not each of the search sentences when a plurality of the search sentences are input is the same session, and a subject in the search result screen The related word detection processing unit that detects the word in the search sentence as the related word, which prompted the output of the search result screen when is selected, and previously selected the subject of the search result screen in the same session as the search A related word detection processing unit that identifies a search that has not been performed and detects a word in the search sentence as a related word, and a correspondence relationship between the detected related word and the related word includes the relationship A related word learning processing unit that associates the correspondence between the related word and the related word and a predetermined degree of association with each other and registers the related word in the related word dictionary DB when the word storage unit does not store the related word in advance. It is characterized by providing.

  According to the present invention, when a user inputs different search sentences in a series of search processes managed by information of the same session, they can be considered as related search sentences. Words included in a plurality of search sentences can be automatically registered as related word pairs (combinations of related words and related words). As a result, the work of the administrator for registering the related word dictionary database can be reduced. In addition, by increasing the number of registered related word pairs, a search with related words other than the words included in the search text input by the user is automatically performed, so that the search accuracy can be improved as compared with the related art.

  Further, the present invention is based on at least the display instruction of the search target document received by selection of the subject, whether any subject is selected from the subject of the search target document displayed on the output search result screen. The subject selection selection means to detect, the same session determination processing unit for determining whether or not each of the search sentences when a plurality of the search sentences are input is the same session, and the input in the same session Among a plurality of search sentences, a search sentence that has been input last time and that is output in response to the input of the search sentence is not selected, and a search sentence that is not selected is specified, and a word in the search sentence is determined. A related word detection processing unit for detecting as a related word, and a search sentence input this time among a plurality of search sentences input in the same session, and input of the search sentence Of the words included in the subject selected on the search result screen that is output in response, the words that are not displayed in the subject of the search result screen that is output in response to the input of the previously input search sentence Is detected as a related word, and a correspondence relationship between the detected related word and the related word is not stored in association with the related word storage unit in advance. A related word learning processing unit for registering a correspondence relationship between a word and a related word and a predetermined degree of association in association with each other.

  According to the present invention, in a series of search processes managed by the same session ID, when the user inputs a different search sentence, the search sentence word before prompting the output of the search result screen not selected, Since the words of the selected subject in the search result screen output based on the next search sentence can be considered to be related, the words can be automatically registered as related word pairs. . As a result, the work of the administrator for registering the related word dictionary database can be reduced. In addition, by increasing the number of registered related word pairs, a search with related words other than the words included in the search text input by the user is automatically performed, so that the search accuracy can be improved as compared with the related art.

  In the present invention, when the correspondence relationship between the detected related word and the related word is already stored in association with the related word storage unit, the degree of association obtained based on a predetermined calculation formula And a related degree increase processing unit for adding to the related degree and the related degree stored in association with the related word. As a result, it is possible to automatically correct the related degree corresponding to the related word and the related word in the related word pair having strong relation.

  Further, the present invention is based on at least the display instruction of the search target document received by selection of the subject, whether any subject is selected from the subject of the search target document displayed on the output search result screen. If no subject is selected from the subject selection determination processing unit to be detected and the subject of the search target document displayed on the output search result screen, the search used for the output of the search result screen Among the correspondences between the words in the sentence and the words included in the non-selected subject, the correspondences that are already associated and stored in the related word storage unit are identified, and the combinations of the words in the correspondences And a relevance reduction processing unit that subtracts the relevance recorded for the image using a relevance obtained based on a predetermined calculation formula. As a result, it is possible to automatically correct the related degree corresponding to the related word and the related word in the weakly related word pair.

  Further, the present invention is a document search method in a document search apparatus for outputting a search target document related to an input search sentence, wherein the related word storage unit includes the word included in the search sentence as already learned information. Is stored in association with the related word and the related word related to the related word, and the degree of association indicating the strength of the relationship between the related word and the related word is stored in association with each other and searched. The word replacement processing unit replaces, with respect to each word included in the search sentence, a word stored in association with the related word in the related word storage unit, and performs search target document extraction processing. Part of the search sentence after the replacement, using a value obtained by multiplying the weight of the word by the relevance for the word replaced, to generate a search document vector of each word of the search sentence after the replacement, In addition, after the replacement A target document vector is generated based on a weight value of each word included in the search sentence in the search target document, and the predetermined angle is small according to an angle formed by the search document vector and the target document vector. A plurality of search target documents are extracted from a plurality of search target documents stored in the search target document storage unit, and a search result screen generation processing unit determines the subject of the extracted search document as the search document vector A search result screen data to be sequentially displayed according to an angle formed by the target document vector and a search result screen output processing unit outputs the search result screen data.

  In addition, the present invention stores, as already learned information, information on the correspondence relationship between a related word and a related word related to the related word as a word included in the search sentence, and also relates to the related word. A program for causing a computer of a document search apparatus to output a search target document related to the search sentence, comprising a related word storage unit that stores a degree of association indicating the strength of a relation with a word, Among the words included in the search sentence, for the word that is stored in association with the related word in the related word storage unit, the process of replacing the related word with the related word and the replacement of the search sentence after the replacement For each word, a value obtained by multiplying the weight of the word by the relevance is used to generate a search document vector for each word in the replacement search sentence, and each word included in the replacement search sentence of A target document vector is generated based on the weight value in the search target document, and a predetermined number of search target documents having a small angle formed according to the angle formed by the search document vector and the target document vector are searched. A process of extracting from among a plurality of search target documents stored in the target document storage unit, and a subject of the extracted search document are sequentially displayed according to an angle formed by the search document vector and the target document vector. A program for causing a computer to execute processing for generating search result screen data and processing for outputting the search result screen data.

  In addition, the present invention stores, as already learned information, information on the correspondence relationship between a related word and a related word related to the related word as a word included in the search sentence, and also relates to the related word. A storage medium for storing a program to be executed by a computer of a document search apparatus that outputs a search target document related to the search sentence, including a related word storage unit that stores a degree of association indicating the strength of the relation with the word Of the words included in the search sentence, for words stored in association with the related word in the related word storage unit, processing for replacing the related word and the search after the replacement Among the sentences, for the replaced word, a search document vector of each word of the search sentence after the replacement is generated using a value obtained by multiplying the weight of the word by the relevance, and the search sentence after the replacement A target document vector is generated based on a weight value of each word included in the search target document, and a predetermined number of searches with a small angle formed according to an angle formed by the search document vector and the target document vector The process of extracting the target document from the plurality of search target documents stored in the search target document storage unit, and the subject of the extracted search document at an angle formed by the search document vector and the target document vector A recording medium for storing a program for causing a computer to execute processing for generating search result screen data to be sequentially displayed and processing for outputting the search result screen data.

  Hereinafter, a document search system (document search apparatus) according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a document search system according to the embodiment. In this figure, reference numeral 1 denotes a document search server. Reference numeral 2 denotes a Web server. Reference numeral 3 denotes a terminal such as a PC (Personal Computer). The document search server 1 is connected to the Web server 2 via a communication network, and the terminal 3 is connected to the Web server via the Internet. In the present embodiment, when the terminal 3 accesses the Web server 2 and notifies the search text information, the Web server 2 transfers the search text information to the document search server 3. Then, the document search server 3 outputs information on the search target document related to the search sentence from the search target documents to the Web server. In addition, the Web server 2 performs processing for transmitting data of a Web page for displaying information of a search target document related to the search sentence to the terminal 3. In this process, the document search server 1 performs a process of extracting information on a search target document with high search accuracy, that is, a search target document whose content is closer to the search sentence than in the past, by a process described later.

FIG. 2 is a diagram showing functional blocks of the Web server and the document search server.
As shown in this figure, the Web server 2 includes a session ID generation unit 21 and a web page processing unit 22. The session ID is identification information for identifying a series of document searches. With this session ID, the user can identify a search sentence received in a series of processes for retrieving a document having a certain content. The session ID generation unit 21 generates this session ID by processing to be described later. Further, the web page processing unit 22 performs processing for receiving information on a search sentence and transmitting web page data of the search result.

  In the document search server 1, reference numeral 101 denotes a control unit that controls each processing unit. Reference numeral 102 denotes a search vector and a document vector generation unit that generates a target document vector. Reference numeral 103 denotes a document search unit that performs a process of extracting a predetermined number of search target documents related to a search sentence from a plurality of search target documents. Reference numeral 104 denotes a related word learning unit that performs processing for registering related words of words included in a search sentence in a database. Reference numeral 105 denotes a morpheme analysis dictionary DB (database) that stores words and the like for performing morpheme analysis. Reference numeral 106 denotes a related word dictionary DB that stores a word (related word) included in a search sentence, a word related to the word (related word), and a degree of association of the combination of the words in association with each other. Reference numeral 107 denotes a search target document DB that stores a plurality of search target documents. Reference numeral 108 denotes a search history DB that stores a history of information when a search process is performed using the search sentence each time information on the search sentence is received.

FIG. 3 is a diagram illustrating a configuration of data stored in the search target document DB.
As shown in this figure, the search target document DB 107 stores a document ID, subject information, body information, document vector, and word information included in the subject in association with each other. Here, the document vector stored in the search target document DB 107 is represented by the weight value of each word included in the subject and the text.

FIG. 4 is a diagram illustrating a configuration of data stored in the search history DB.
As shown in this figure, the search history DB 108 stores a session ID, a search ID, a search sentence word, a search result, a viewed document, and a related word pair in association with each other. Here, the search ID is information for identifying a process each time information of a search sentence is received. The search sentence word is a word included in the search sentence. The search result is a document ID of a search target document searched by the document search server 1 based on a search sentence. The browse document is the document ID of the search target document browsed by the user who uses the terminal 3 in the search results. The related word pair stores a combination of a word (related word) included in the search sentence and a word (related word) stored in the related word dictionary DB 106 in association with the word in the document search server 1. Information.

FIG. 5 is a diagram showing a configuration of data stored in the related word dictionary DB.
As shown in this figure, the related word dictionary DB 106 stores a word (related word) included in a search sentence, a word related to the word (related word), and a degree of relevance indicating the strength of the relation in the combination of these words a (0 ≦ a ≦ 1) is stored in association with each other. Note that the document search server 1 determines whether to increase each time a search sentence is input for a combination of a related word and a related word stored in the related word dictionary DB 106 by a process described later, Perform a reduction process. Further, the relevance degree a is increased or subtracted. As a result, the related word dictionary DB 106 is automatically generated to reduce the labor of the administrator, and a process for outputting a search result with high accuracy based on the information stored in the related word dictionary DB 106 is performed.

FIG. 6 is a diagram showing a processing flow of the document search server.
Next, the processing flow of the document search server will be described with reference to FIG.
First, the Web server 2 that has received access from the Web browser function of the terminal 3 outputs a search text input screen for document search to the terminal 3. When a user using the terminal 3 inputs a search sentence on this search sentence input screen and gives a search instruction, the terminal 3 transmits information on the search sentence to the Web server 2. Then, the Web server 2 receives the search text, the session ID generation unit 21 generates a session ID, and transmits the search text and the session ID to the document search server 1. At this time, it is assumed that the document search server 1 has already received the previous search statement indicating the same session ID, and the search history information about this has been recorded in the search history DB 108. As for the session ID, for example, a search sentence received at an interval not exceeding a predetermined time interval may be managed by the same session ID, or an instruction such as “change search” is received on the search sentence input screen, and the information May be managed by the same session ID until the document retrieval server 1 receives the message. Until the document search server 1 determines that a series of processes for searching for the same content is completed, the search sentence received from the terminal 3 by the process described later is stored in the search history DB 108 in association with the same session ID.

  When the document search server 1 accepts a search sentence, the document vector generation unit 102 decomposes the search sentence into words (step S101). Then, the document vector generation unit 102 sets a word in the search sentence as a related word, replaces the word of the related word with a related word recorded in the related word dictionary DB 106 (step S102), and after the replacement A search sentence vector of the search sentence is generated (step S103). At this time, processing is performed to change the weight of each word included in the search sentence vector of the search sentence after replacement using the relevance. Then, the document search unit 103 calculates an angle formed by the search sentence vector of the search sentence after replacement and the target document vector for each combination of the search sentence and the search target document.

  Next, as a vector search process, the document vector generation unit 102 specifies a target document vector having a small angle with the search sentence vector of the replacement search sentence, and determines a predetermined number of search target documents as search results. (Step S104). Here, the search processing may be performed using the target document vectors of all the search target documents included in the search target document 107 and the search sentence vectors of the search sentences after replacement, or included in the search target document DB 107. Search processing using the target document vector of a plurality of search target documents that always include any word included in the replacement search sentence and the search sentence vector of the replacement search sentence. Also good. The process of determining the search target document that is the search result will be described in detail later.

  Next, when the search result is specified, the document search unit 103 generates a search ID unique to the search sentence. Then, the document ID of the search target document determined as the search result and the subject information are read from the search target document DB 107, and the read information is received from the session ID generation unit 21 and the generated search. The ID is registered in the search history DB 108 in association with the ID. Further, the document search unit 103 extracts a word by performing a morphological analysis on the search sentence, and registers each word in association with the search history DB 108. Further, the document search unit 103 registers, in the search history DB 108, a combination of related words used by the document vector generation unit 102 to generate search sentence vectors and related words used to extract the related words as related word pairs. To do. With the above processing, the document search server 1 completes the search processing.

  Then, the control unit 101 transmits the document ID of the search target document determined as the search result and the subject information to the Web server 2. Next, the web page processing unit 22 of the web server 2 generates search result screen data that displays a list of document IDs and subject information of the search target documents determined as the search results, and the search result screen data is generated. Transmit to terminal 3.

  In the terminal 3, if a subject representing the content of the document that the user desires to browse is displayed in the subject displayed on the search result screen, an instruction to display the search target document with the subject is input. . Also, if the subject displayed on the search result screen does not display a subject representing the content of the document that the user wants to view, the display instruction for the search target document with that subject is not input. Here, when an instruction to display a search target document for a certain subject is input to the terminal 3, the terminal 3 transmits information on a search target document display request including information on the subject and document ID to the Web server 2. The Web server 2 transmits information on the search target document display request to the document search server 1. The document search server 1 determines whether or not the user has browsed any search target document in the search result depending on whether or not the search target document display request information is received (step S105). If it is determined that the document has been browsed, the document search unit 103 reads the text information and the like recorded in the search target document DB 107 in association with the document ID, and transmits the information to the Web server 2 via the control unit 101. The web page processing unit of the web server 2 generates text display screen data and transmits the data to the terminal 3.

  When transmitting the text display screen data to the terminal 3, the related word learning unit 104 performs an automatic update process of the related word dictionary DB 106. This automatic update processing includes related word pair candidate creation processing (step S106), related word pair candidate related word dictionary DB 106 registered determination processing (step S107), unregistered related word pair candidate related word dictionary DB 106. Registration processing (step S108), relevance increase processing of registered related word pairs (step S109), and the like are performed. In addition, when the subject displayed on the search result screen is not selected (that is, when it is determined that any of the search target documents in the search result has not been browsed), the related word learning unit 104 performs a process described later. In the search sentence after replacement, the relevance reduction process is performed for the related word pair including the word before replacement and the word after replacement (step S110). Further, it is determined again whether or not the information of the search sentence is accepted (step S111). When the search sentence is accepted, the search process is started.

Next, a detailed processing flow after step S105 of the document search server 1 when it is determined that any search target document on the search result screen has been browsed in the process of step S105 will be described.
First, after transmitting the text display screen to the terminal 3, the related word learning unit 104 reads the word of the search sentence that prompted the transmission of the text display screen. Further, it is a search ID of another search sentence recorded in the search history DB 108 in association with the same session ID as the session ID that accepted the search sentence, and information is recorded in the recording area of the browse document. A word of a search sentence having no search ID (that is, an ID of a search sentence in which all search target documents have not been browsed on the search result screen) is read. And the related word pair candidate which matched these read words is produced (process of the said step S106). At this time, the word of the search sentence that prompted the transmission of the text display screen is the related word, and the word of the search sentence in which all the search target documents are not browsed on the search result screen is the related word. The sum of the relevance levels corresponding to the word is 0.5. The value of the total degree of relevance can be changed as appropriate.

  The processing in step S106 will be described in more detail with reference to FIG. 4. First, the search ID of the search sentence that prompted the transmission of the text display screen is “R103”, and the search ID of another search sentence in the same session is used. If the search IDs of the previous search sentences that were not browsed on the search result screen are “R102” and “R101”, first, the search ID “R103” is obtained by the combination of the search IDs “R103” and “R102”. It can be seen that the search sentences of "compression tool" and "restoration" include the words (related words), and the search ID "R102" includes the word "cab" (related words). . Therefore, the related word learning unit 104 creates a relationship “cab” → “compression tool” and a relationship “cab” → “restoration” as related word pair candidates. At this time, since there are two words included in the current search sentence, each of the related word pair candidates is expressed by “total sum of relevance ÷ number of words included in current search sentence” = “0.5 ÷ 2.” Calculate relevance. FIG. 7 is a first diagram showing extracted related word and related word pair candidates of related words and their degrees of association.

  In FIG. 4, the related word learning unit 104 includes “compression tool” and “restoration” words (related words) for the search sentence with the search ID “R103”, and “search for the search ID“ R101 ”. Confirm that the word “zip” and the word “development” (related words) are included. Then, four relationships of “zip” → “compression tool”, “zip” → “restoration”, “decompression” → “compression tool”, “decompression” → “restoration” are extracted as related word pairs. At this time, since there are two search sentence words (related words) with the search ID “R101” with respect to the related word “zip”, “zip” → “compression tool” and “zip” The degree of relevance of “→ restoration” is “sum of relevance / number of words included in current search sentence” = 0.25. Similarly, each degree of association of “decompression” → “compression tool” and “decompression” → “restore” is 0.25. FIG. 8 is a second diagram showing extracted related word and related word pair candidates of related words and their degrees of association.

  According to the related word pair candidate creation as shown in FIGS. 7 and 8 above, when a user inputs a different search sentence in a series of search processes managed by the same session ID, the search result is browsed. Search terms that are not browsed and search terms that are not viewed can be considered as related search terms. You can register automatically. As a result, the administrator's work for registering the related word dictionary DB 106 can be reduced. In addition, by increasing the number of registered related word pairs, a search with related words other than the words included in the search text input by the user is automatically performed, so that the search accuracy can be improved as compared with the related art.

Moreover, the related word learning part 104 performs the process which produces a related word pair candidate by the following process besides the above-mentioned process.
The related word learning unit 104 specifies a search ID of a search sentence in which all search target documents are not browsed on the search result screen among a plurality of search IDs associated with the same session ID, and the search sentence of the search ID The document ID of the retrieval target document as a result of retrieval based on the retrieval history DB 108 is read from the retrieval history DB 108. Further, a word (subject word) included in the subject recorded in the search target document DB 107 in association with the document ID is read. And the word which is not contained in the read word is acquired from the words contained in the subject browsed from the search result. Then, the acquired word (related word) and the word (related word) included in the search sentence in which the entire search target document is not browsed are extracted as related word pairs. Hereinafter, this process will be described in more detail.

First,
1. Words included in a search sentence in which all search target documents have not been browsed (a group of these words is assumed to be a search sentence word A) are defined as “zip” and “development”.
Also,
2. “Drive”, “Folder”, “Error”, “Word” included in the subject displayed on the search result screen where all the search target documents have not been browsed are assumed to be a subject word A. Computer ".
Also,
3. The subject displayed on the search result screen is selected and the text is transmitted, and the word included in the subject displayed on the search result screen <a group of these words as subject word B> is expressed as “compressed”, “ “Folder”, “Right click”.

  In such a case, it can be paraphrased as the following processing. That is, “the user inputs the search sentence word A, and as a result, the subject word A is output, but even if the subject word A is seen, it is determined that the desired search target document is not obtained as a result. Further, the search sentence word B is input as the next search, the search result screen is output, and when the subject of the desired search target document is on the screen, the subject word B included in the subject is selected. It can be paraphrased as a process such as “

  At this time, the word group of the subject word B <“compression”, “folder”, “right click”> to the word group of the subject word A <“drive”, “folder”, “error”, “computer”> Each word <“compressed”, “right click”> excluding the included words can be considered as a word related to the word group <“zip”, “expand”> of the search sentence word A. Because the word of each subject displayed on the search result screen that was not selected is ineligible as a word representing the desired search target document, the ineligible word is displayed on the selected search result screen. This is because it can be said that each word of the search sentence that prompts the output of the search result screen that has not been selected is related except for the displayed subject words. Therefore, in a series of search processes managed by the same session ID, when the user inputs a different search sentence, the search sentence word <"zip", " "Expanded"> (related word) and a word that is not displayed in the search result screen output by the previous search sentence among the words of the selected subject in the search result screen output based on the next search sentence < Four combinations of “compressed”, “right-click”> (related word) of related words and related words are extracted as related word pair candidates.

  According to the above related word pair extraction, when a user inputs a different search sentence in a series of search processes managed by the same session ID, the search before prompting the output of the search result screen not selected. Sentence words and words obtained by subtracting the subject words displayed in the search result screen not selected from the selected subject words in the search result screen output based on the next search sentence Can be considered as a related search sentence, the words can be automatically registered as related word pairs. As a result, the administrator's work for registering the related word dictionary DB 106 can be reduced. In addition, by increasing the number of registered related word pairs, a search with related words other than the words included in the search text input by the user is automatically performed, so that the search accuracy can be improved as compared with the related art.

  Next, the related word learning unit 104 determines whether or not the related word pair candidate extracted by the above-described process is registered in the related word dictionary DB 106 (the process of step S107). If it is already registered in the related word dictionary DB 106, a process of increasing the degree of association associated with the related word pair and recorded in the related word dictionary DB 106 is performed (the process of step S108). If not registered in the related word dictionary DB 106, the extracted related word pair candidates are registered in the related word dictionary DB 106 (the process in step S109). In this process of increasing relevance, it is assumed that related word pairs are “cab” → “restoration” and “cab” → “expansion”. In addition, the degree of association for the related word pair is already registered in the related word dictionary DB 106, the degree of association of the related word pair “cab” → “restoration” is 0.5, and the related word “cab” → “expand”. Assume that the relevance of a pair is 0.6.

  First, the related word learning unit 104 includes a related word (restoration and expansion) of a related word pair candidate created by the above-described processing in the subject of the retrieved search target document determined in the process of step S105. The document ID of the document is read from the search history DB 108. Then, the document vector recorded in the search target document DB 107 in association with the document ID is read. Then, a vector value corresponding to a related word (restoration, expansion) of a related word pair candidate is extracted from the document vector. Here, it is assumed that the vector value for restoration is W1 and the vector value for expansion is W2. As described above, this vector value is represented by the weight value in the word document. This weight value is calculated by calculating a conventional TF (Term Frequency: the number of occurrences of one word in one document) value, IDF (inverse document frequency: the occurrence frequency of words in all documents), or the like. This is a value calculated and recorded in advance. Then, the maximum increase value with respect to the related word pair candidate is set to 0.1, and the value obtained by proportionally distributing the value by the vector value of the “related word” of the related word pair candidate is expressed as “related word” → “related word”. In the related word pair with “restore” as the related word,

0.1 × W1 / (W1 + W2) (1)

  Also, in the related word pair with “expanded” as the related word,

0.1 × W2 / (W1 + W2) (2)

  To calculate the increase in relevance. For example, when the calculation result of Expression (1) is 0.01, the relevance of the related word pair “cab” → “restoration” is 0.51, and the calculation result of Expression (2) is 0. In the case of 0.09, the relevance of the related word pair “cab” → “expand” is increased to 0.69. Although the maximum increase value for the related word pair is 0.1, this value is not limited to this and can be set as appropriate.

  Next, processing when a subject is not selected on the above-described search result screen will be described. When the information of the next search sentence is transmitted from the terminal 3 and received, when the instruction information such as “search change” is received from the terminal 3, or when a predetermined time has elapsed since the output of the search result screen When it is detected that the time has elapsed, it is determined that the subject displayed on the search result screen has not been selected (the process of step S105). Then, the document ID of the search target document recorded in the search history DB 108 in association with the search ID of the search sentence that prompted the output of the search result screen is read from the search result column. Then, a subject word recorded in the search target document DB 107 in association with the document ID is read. Then, using the read subject word as a related word and the word of the search sentence as a related word, the relevance of the related word pair of the related word → the related word is subtracted. The reduction value is desirably 0.1 or less, for example. The degree of association a can be increased or decreased in the range of 0 ≦ a ≦ 1. The reason for this will be described later.

Next, details of the search for the search target document will be described.
In the process of step S104, for example, “zip” and “development” are included as words included in the original search sentence, and “zip” → “compression tool” and “zip” → “restoration” are related word pairs. Assume that the related word dictionary DB 106 is registered. At this time, words included in the search sentence in which the related word is replaced with the related word are “compression tool”, “restoration”, and “development”. A search sentence vector is generated by the same method as the conventional vector search method using words included in the search sentence after replacement. Now, the word included in the search sentence by the replacement,
1. "Compression tool"
2. "Restore"
3. "Deployment"
The related word pair registered in the related word dictionary DB 106 is zip (related word) → compression tool (related word): 0.2 (relevance)
zip (related word) → restoration (related word): 0.3 (relevance)
Therefore, if the weight value of the word included in the search sentence after the replacement in the search target document is (compression tool, decompression, expansion) = (W1, W2, W3), the search sentence vector is ( W1 × 0.2, W2 × 0.3, W3). In other words, when generating a search sentence vector of a search sentence after replacement, the relevance level associated with the relevance pair having the word as a related word is added to the weight value of the word included in the search sentence. Multiply to calculate the vector value for that word. Further, the search target document vector (target document vector) is determined by the weight component corresponding to the word included in the search sentence of the search target document holding any of the words included in the search sentence. That is, the words included in the search sentence are “compression tool”, “restoration”, and “decompression”, and the target document vector of the search target document including all the words is “compression tool”, “ It is represented by the weight of each word of “restoration” and “development”. The weight value is obtained by acquiring a value corresponding to each word from among a plurality of values included in the document vector of the search target document DB 107.

  Then, the document search unit 103 calculates an angle formed between the search sentence vector of the search sentence after replacement and the target document vector for each combination of the search sentence and the plurality of search target documents. Then, the document vector generation unit 102 determines a predetermined number of search target documents corresponding to the target document vector having a small angle with the search sentence vector as a search result. The search target document is searched by the above processing. Here, the plurality of search target documents for calculating the target document vector may be all search target documents included in the search target document 107, or a search sentence among the search target documents included in the search target document 107. It may be a plurality of search target documents that always include the words included in. Then, the document search unit 103 registers the document ID of the search target document determined as the search result in the search history DB 108 associated with the search ID indicating the search process.

  According to this processing, the degree of relevance indicating a value in the range of 0 ≦ relevance a ≦ 1 is multiplied by the value of the word weight included in the search sentence after replacement with the related word, and the search sentence vector is calculated by the value. Is generated. As a result, the angle formed with the target document vector of the search target document is smaller than the search sentence vector of the search sentence that is not replaced with the related word. Therefore, the angle formed between the search sentence vector of the search sentence after replacement with the related word and the target document vector of the search target document rather than the angle formed between the search sentence vector of the original search sentence and the target document vector of the search target document. Since it is expected that is smaller, it is possible to output a search result screen in which the result of the original search sentence is the upper search result. As a result, it is possible to display on the terminal 3 giving priority to the search result using the original search sentence.

  Next, the control unit 101 transmits the document ID of the search target document determined as the search result and the subject information to the Web server 2. Next, the web page processing unit 22 of the web server 2 generates a search result screen that displays a list of document IDs of subject documents determined as search results and subject information, and stores the data of the search result screen as a terminal. 3 to send.

  Each of the above servers and terminals has a computer system inside. The process described above is stored in a computer-readable recording medium in the form of a program, and the above process is performed by the computer reading and executing this program. Here, the computer-readable recording medium means a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, or the like. Alternatively, the computer program may be distributed to the computer via a communication line, and the computer that has received the distribution may execute the program.

  The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

It is a block diagram which shows the structure of a document search system. It is a figure which shows the functional block of a Web server and a document search server. It is a figure which shows the structure of the data which search object document DB memorize | stores. It is a figure which shows the structure of the data which search history DB memorize | stores. It is a figure which shows the structure of the data which related word dictionary DB memorize | stores. It is a figure which shows the processing flow of a document search server. It is a 1st figure which shows the related word pair of a related word and a related word, and its relevance degree. It is a 2nd figure which shows the related word pair of a to-be-related word and a related word, and its relevance degree.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Document search server, 2 ... Web server, 3 ... Terminal, 101 ... Control part, 102 ... Document vector generation part, 103 ... Document search part, 104 ... Related Word learning unit, 105 ... morphological analysis dictionary DB, 106 ... related word dictionary DB, 107 ... search target document DB, 108 ... search history DB, 21 ... session ID generation unit, 22 ..Web page processing section

Claims (8)

A document search device that outputs a search target document related to an input search sentence,
As information that has already been learned, information on the correspondence relationship between the related word and the related word related to the related word is stored as the word included in the search sentence, and the relationship between the related word and the related word is stored. A related word storage unit for storing the degree of association indicating strength in association with each other;
Among words included in the search sentence, for words stored in association with the related word in the related word storage unit, a search word replacement processing unit to replace the related word,
Among the search sentences after the replacement, for the replaced word, a search document vector of each word of the search sentence after the replacement is generated using a value obtained by multiplying the weight of the word by the relevance, A target document vector is generated based on a weight value in the search target document of each word included in the search sentence after replacement, and the formed angle is determined according to an angle formed by the search document vector and the target document vector. A search target document extraction processing unit for extracting a small predetermined number of search target documents from a plurality of search target documents stored in the search target document storage unit;
A search result screen generation processing unit that generates search result screen data capable of sequentially displaying the subject of the extracted search document in accordance with an angle formed by the search document vector and the target document vector;
A search result screen output processing unit for outputting the search result screen data;
A document search apparatus comprising: Subject selection selection for detecting whether any subject is selected from the subject of the search target document to be displayed on the output search result screen, based on at least the display target document display instruction received by selection of the subject A processing unit;
The same session determination processing unit that determines whether each of the search sentences when the search sentences are input over a plurality of times is the same session,
A related word detection processing unit that detects a word in a search sentence as a related word, which prompted the output of the search result screen when a subject is selected on the search result screen;
In the same session as the search, a search that has not previously selected the subject of the search result screen is specified, and a related word detection processing unit that detects a word in the search sentence as a related word;
When the correspondence relationship between the detected related word and the related word is not stored in the related word storage unit in advance, the correspondence relationship between the related word and the related word and a predetermined degree of association are obtained. A related word learning processing unit to be registered in the related word dictionary DB in association with each other;
The document search apparatus according to claim 1, further comprising: Subject selection selection for detecting whether any subject is selected from the subject of the search target document to be displayed on the output search result screen, based on at least the display target document display instruction received by selection of the subject Means,
The same session determination processing unit that determines whether each of the search sentences when the search sentences are input over a plurality of times is the same session,
Among the plurality of search sentences input in the same session, a search sentence that has been input last time and that has been selected in response to the input of the search sentence is identified. A related word detection processing unit for detecting a word in the search sentence as a related word;
Among the plurality of search sentences input in the same session, among the words included in the subject selected in the search result screen that is the search sentence input this time and output according to the input of the search sentence, A related word detection processing unit that detects a word that is not displayed in the subject of the search result screen that is output according to the input of the previously input search sentence, as a related word;
When the correspondence relationship between the detected related word and the related word is not stored in the related word storage unit in advance, the correspondence relationship between the related word and the related word and a predetermined degree of association are obtained. A related word learning processing unit registered in association with each other;
The document search apparatus according to claim 1, further comprising: When the correspondence relationship between the detected related word and the related word is already stored in association with the related word storage unit, the degree of relevance obtained based on a predetermined calculation formula is calculated. A degree-of-association increasing processing unit for adding to the degree of association stored in association with the word and the related word;
The document search apparatus according to claim 2, further comprising: Subject selection selection for detecting whether any subject is selected from the subject of the search target document to be displayed on the output search result screen, based on at least the display target document display instruction received by selection of the subject A processing unit;
When no subject is selected from the subject of the search target document displayed on the output search result screen, the word in the search sentence used for the output of the search result screen and the selection are not performed. Among the correspondences of the words included in the subject name, the correspondences that are already associated and stored in the related word storage unit are specified, and the degree of association recorded for each word combination of the correspondences, A relevance reduction processing unit that subtracts using the relevance obtained based on a predetermined calculation formula;
The document search apparatus according to claim 1, further comprising: A document search method in a document search apparatus for outputting a search target document related to an input search sentence,
The related word storage unit stores, as already learned information, information on the correspondence relationship between the related word and the related word related to the related word for the word included in the search sentence, and the related word Stores the association level indicating the strength of the association with the related word,
The search word replacement processing unit replaces each word included in the search sentence with respect to a word that is stored in association with the related word in the related word storage unit.
The search target document extraction processing unit uses a value obtained by multiplying the weight of the word by the relevance for the replaced word in the search sentence after the replacement, and the search document of each word of the search sentence after the replacement A vector is generated, and a target document vector is generated based on a weight value in the search target document of each word included in the search sentence after the replacement, and an angle formed by the search document vector and the target document vector In response to the above, a predetermined number of search target documents having a small angle formed are extracted from a plurality of search target documents stored in the search target document storage unit,
The search result screen generation processing unit generates search result screen data for sequentially displaying the subject of the extracted search document according to the angle formed by the search document vector and the target document vector,
The search result screen output processing unit outputs the search result screen data.
A document search method characterized by the above. As information that has already been learned, information on the correspondence relationship between the related word and the related word related to the related word is stored as the word included in the search sentence, and the relationship between the related word and the related word is stored. A related word storage unit that stores association levels indicating strength in association with each other;
A program to be executed by a computer of a document search device that outputs a search target document related to the search sentence,
Among the words included in the search sentence, for the word stored in association with the related word in the related word storage unit, processing to replace the related word,
Among the search sentences after the replacement, for the replaced word, a search document vector of each word of the search sentence after the replacement is generated using a value obtained by multiplying the weight of the word by the relevance, A target document vector is generated based on a weight value in the search target document of each word included in the search sentence after replacement, and the formed angle is determined according to an angle formed by the search document vector and the target document vector. A process of extracting a small predetermined number of search target documents from a plurality of search target documents stored in the search target document storage unit;
A process of generating search result screen data for sequentially displaying the subject of the extracted search document according to an angle formed by the search document vector and the target document vector;
Processing to output the search result screen data;
A program that causes a computer to execute. As information that has already been learned, information on the correspondence relationship between the related word and the related word related to the related word is stored as the word included in the search sentence, and the relationship between the related word and the related word is stored. A related word storage unit that stores association levels indicating strength in association with each other;
A recording medium for storing a program to be executed by a computer of a document search apparatus that outputs a search target document related to the search sentence,
Among the words included in the search sentence, for the word stored in association with the related word in the related word storage unit, processing to replace the related word,
Among the search sentences after the replacement, for the replaced word, a search document vector of each word of the search sentence after the replacement is generated using a value obtained by multiplying the weight of the word by the relevance, A target document vector is generated based on a weight value in the search target document of each word included in the search sentence after replacement, and the formed angle is determined according to an angle formed by the search document vector and the target document vector. A process of extracting a small predetermined number of search target documents from a plurality of search target documents stored in the search target document storage unit;
A process of generating search result screen data for sequentially displaying the subject of the extracted search document according to an angle formed by the search document vector and the target document vector;
Processing to output the search result screen data;
Medium for storing a program for causing a computer to execute the program.

Images