JP2012003603A - Information retrieval system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To display an accurate retrieval result by focusing on a relationship between words.SOLUTION: In a retrieval system consisting of one or more computers to retrieve text information, the computers include: means for obtaining a first correlation between words included in a retrieval condition by analyzing the retrieval condition input by a user; means for obtaining a second correlation between words included in text information by analyzing the stored text information; and means for selecting the text information to be output as the one matched with the input retrieval condition on the basis of the similarity between the obtained first correlation and the second correlation.

Description

  The present invention relates to an information search system, and more particularly to a technique for improving search accuracy.

  In a general search system (search engine), content search is realized by performing pre-processing and main processing. Here, the content is information that can be processed by an electronically expressed computer such as an electronic document.

  In the preprocessing, a search index is generated. Specifically, the content to be searched is analyzed, a characteristic character string (characteristic character string) included in the content is acquired, the acquired characteristic character string and the content are mapped, and stored in the storage area as a search index. ing. In this process, a search query is analyzed, a keyword group (search term) used for the search process is extracted, the search index generated in the pre-process is scanned using the extracted search term, and a set of corresponding contents To get. Thereafter, the search results are sorted by calculating the similarity with the corresponding content.

Special table 2007-507801 gazette JP 2005-165632 A JP 2000-200221 A

  However, words often have multiple meanings. In particular, new words and abbreviations tend to have multiple meanings. For this reason, when a search is performed with a certain keyword, there may be a case where a lot of information having an unexpected meaning is arranged at the top of the search result. Thus, the searcher's purpose cannot be achieved with a result different from the searcher's intention.

  On the other hand, the searcher needs to re-execute the search by reviewing the search query, such as adding a search keyword, or browsing the content up to the lower search result. Such re-searching requires the searcher to have search skills and work time, resulting in a decrease in work efficiency and work quality.

  This is a problem that occurs because the search is performed only by the relationship between the content and the characteristic character string (vocabulary), and it is difficult to avoid this problem with current search engines.

  In order to solve such a problem, for example, in Patent Document 1, the personal information of a user is stored as profile information, and the order of appearance in a search result list is determined in consideration of the degree of fitness for user preference and interest. The search method to be adjusted is described. However, the method described in Patent Document 1 cannot cope with the case where the weights of homonyms as viewed from the user are different.

  Patent Document 2 proposes an information search method for recommending content according to the similarity between the content feature vector and the user preference vector. However, the method described in Patent Document 2 cannot cope with the case where the weights of homonyms seen from the user are different.

  There is also a method for solving this problem using a synonym dictionary or ontology (see, for example, Patent Document 3). The method described in Patent Document 3 is a method by extending the search term, and it is possible to reduce search omissions. However, it is difficult to eliminate a document (search noise) that is not intended by the searcher, and rather, it is a factor that increases search noise.

  Accordingly, an object of the present invention is to provide an information search system that displays an accurate search result by paying attention to the relationship between words.

  According to a typical example of the present invention, focusing on the relationship between vocabulary and vocabulary, a search is performed in consideration of not only the relationship between content and vocabulary but also the relationship between vocabulary and vocabulary.

  Specifically, as preprocessing, the content to be searched is analyzed, the vocabulary included in the content is acquired, the relationship between the content and the vocabulary is mapped as a content index, the relationship between the vocabulary and the vocabulary is analyzed, A content vocabulary model indicating the analyzed relevance is generated. And based on a searcher's (user) action history, the vocabulary relevant to a user is extracted and a user index is produced | generated. Then, the relationship between vocabularies contained in the file accessed by the user is also analyzed, and a user vocabulary model indicating the analyzed relationship is generated.

  After that, as a main process, the search term extracted from the search query is used to scan the user vocabulary model to obtain a vocabulary set related to the search term (hereinafter referred to as a user extended search term), and the search term and the related vocabulary set. Is obtained, and a search query matrix in which the obtained degree of association is registered is generated. Further, the user extension term is used to scan the content index generated in the preprocessing, and a set of contents including the user extension term is acquired. Then, the content vocabulary model of each content is scanned, the relevance between vocabularies included in the content is acquired, and a content matrix in which the acquired relevance is registered is generated. Thereafter, the similarity is calculated between the search query matrix and the content matrix, and the search result is sorted based on the calculation result of the similarity.

  According to the embodiment of the present invention, by considering not only the relationship between content and vocabulary but also the relationship between vocabulary and vocabulary, the vocabulary that searchers use everyday even for ambiguous vocabulary. The similarity of content related to is increased. For this reason, the content that the searcher likes can be displayed at the top, the content close to the user's preference can be displayed at the top, and the search accuracy can be improved.

It is a block diagram which shows the whole structure of the information search system of embodiment of this invention. It is a block diagram which shows the structure of the content analysis subsystem 101 of embodiment of this invention. It is a flowchart of the process performed by the content analysis subsystem 101 of embodiment of this invention. It is explanatory drawing of the example of the content index D131 of embodiment of this invention. It is explanatory drawing of the example of the content vocabulary model D132 of embodiment of this invention. It is explanatory drawing of the example of the user action log | history D137 of embodiment of this invention. It is a block diagram which shows the structure of the user analysis subsystem 102 of embodiment of this invention. It is a flowchart of the process performed by the user analysis subsystem 102 of embodiment of this invention. It is explanatory drawing of the example of the user index D135 of embodiment of this invention. It is explanatory drawing of the example of the user vocabulary model D136 of embodiment of this invention. It is a block diagram which shows the structure of the repository management subsystem 103 of embodiment of this invention. It is a flowchart of the process performed by the repository management subsystem 103 of embodiment of this invention. It is a block diagram which shows the structure of the information search server 104 of embodiment of this invention. It is a flowchart of the process performed by the information search server 104 of embodiment of this invention. It is a flowchart which shows the detailed procedure of the search query matrix production | generation process (S412) of embodiment of this invention. It is a flowchart which shows the detailed procedure of the search query matrix production | generation process (S412) of embodiment of this invention. It is explanatory drawing of the example of a production | generation of the search query matrix in step S412 of embodiment of this invention. It is a flowchart which shows the detailed procedure of the similar content extraction process (S413) of embodiment of this invention. It is a flowchart which shows the detailed procedure of the content matrix production | generation procedure (S4134) of embodiment of this invention. It is explanatory drawing of the example of the production | generation of the content matrix in step S413 of embodiment of this invention. It is a flowchart which shows the detailed procedure of the similarity calculation procedure (S4135) of embodiment of this invention. It is a block diagram which shows the structure of the information search client 105 of embodiment of this invention. It is a flowchart of the process performed by the information search client 105 of embodiment of this invention. It is a flowchart of the process performed by the information search client 105 of embodiment of this invention.

<Overall system configuration>
FIG. 1 is a block diagram showing the overall configuration of the information search system according to the embodiment of this invention.

  The information search system according to the embodiment of the present invention includes a content analysis subsystem 101, a user analysis subsystem 102, a repository management subsystem 103, and an information search server 104. Each of these devices is connected by a network (LAN or WAN) 106. An information search client 105 is connected to the information search system via a network 106.

  The content analysis subsystem 101 is a computer that analyzes in advance the content of the search target content before executing the search, and the detailed configuration thereof will be described later with reference to FIG. That is, the content analysis subsystem 101 analyzes and extracts the vocabulary relationships in the content. This vocabulary relationship is necessary for processing in the information search server 104 and is stored in the repository management subsystem 103.

  The user analysis subsystem 102 is a computer that analyzes in advance the profile information of the user who performs the search before executing the search, and the detailed configuration thereof will be described later with reference to FIG. That is, the user analysis subsystem 102 analyzes user profile information by identifying user-specific information (preference, range of interest, etc.) as a vocabulary relationship. This analysis result is necessary for processing in the information search server 104 and is stored in the repository management subsystem 103.

  The repository management subsystem 103 is a database system that stores information analyzed by the content analysis subsystem 101 and the user analysis subsystem 102, and a detailed configuration thereof will be described later with reference to FIG. That is, the repository management subsystem 103 presents information analyzed by the content analysis subsystem 101 and the user analysis subsystem 102 in response to a request from the information search server 104.

  The information search server 104 is a computer that searches for content, and a detailed configuration thereof will be described later with reference to FIG. That is, the information search server 104 extracts a search target from the repository management subsystem 103 in response to a request from the information search client 105, and searches for content similar to the analysis result of the search user from the extracted analysis data. . This search result is presented to the information search client 105.

  The information search client 105 is a computer operated by a user who performs a search, and a detailed configuration thereof will be described later with reference to FIG. That is, the information search client 105 receives a search request from the user, requests a search from the information search server 104, receives the search result from the information search server 104, and presents the search result to the user via the output device.

  Although only one search client is shown in FIG. 1, a plurality (two or more) of search clients may be provided.

  The network 106 is a network that is partially or entirely configured by a LAN or WAN, and connects the content analysis subsystem 101, the user analysis subsystem 102, the repository management subsystem 103, the information search server 104, and the information search client 105. To do.

  Note that the analysis result is transferred from the content analysis subsystem 101 and the user analysis subsystem 102 to the repository management subsystem 103 via the network 106, but a portable storage medium (for example, an optical disk, a non-volatile memory) Etc.) may be used to transfer the analysis results.

  Two or more systems among the content analysis subsystem 101, the user analysis subsystem 102, the repository management subsystem 103, and the information search server 104 can be mounted on a single physical computer. In this case, data can be transferred between the systems via the internal bus.

  Further, the information search client 105 and other subsystems (for example, the information search server 104) can be mounted on one physical computer.

<Content analysis subsystem 101>
FIG. 2 is a block diagram showing a configuration of the content analysis subsystem 101 according to the embodiment of this invention.

  The content analysis subsystem 101 includes a memory 110, a storage device 120, a CPU 130, an output device 140, an input device 150, and a communication interface 160, and is a computer having a general configuration in which these components are connected by a bus 170.

  The memory 110 stores a program executed by the CPU 130. Specifically, the system control program P10 and the content analysis program P11 are stored in the memory 110. In addition, the memory 110 is provided with a work area for temporarily storing data when the CPU 130 executes the program.

  The system control program P10 is a so-called operating system, and controls the entire content analysis subsystem 101.

  The content analysis program P11 is a program for analyzing the content of the search target content.

  The storage device 120 is a non-volatile storage element that can retain stored contents even when the power is shut off, and is configured by, for example, a magnetic disk drive (HDD) or a flash memory (SSD). The storage device 120 stores various programs D100. The various programs D100 include the system control program P10 and the content analysis program P11 described above, and are loaded into the memory 110 when executed by the CPU 130.

  CPU 130 executes a program stored in memory 110.

  The output device 140 is a computer that includes an input device 150 and an interface 160, and each of these components is connected by a bus 170.

  The interface 160 is a network interface (NIC) that transmits and receives data to and from devices connected to the network 106 according to a predetermined protocol.

  The output device 140 is a display that displays the processing result on a screen, a printer that outputs the result to paper, or the like.

  The input device 150 is a keyboard, a mouse, or the like for a user to give an instruction to the content analysis subsystem 101.

  The content analysis subsystem 101 transfers the content analysis result obtained by executing the content analysis program P11 to the repository management subsystem 103 via the network 106. Next, details of this processing will be described.

  FIG. 3 is a flowchart of processing executed by the content analysis subsystem 101 according to the embodiment of this invention.

  First, when the content analysis subsystem 101 receives specification of content to be analyzed, the content analysis subsystem 101 generates a list of all content to be analyzed (S101), and an index D131 of content included in the generated list (FIG. 11). Reference) is acquired from the repository management subsystem 103 (S102). The content index acquired in step S102 includes the analysis result of the content so far, and the configuration thereof will be described later with reference to FIG.

  The content analyzed by the content analysis subsystem 101 is content information D130 stored in the storage device 120 of the repository management subsystem 103. Each content information is text data included in the information content and text data added to the information content. In this case, each content information (each text data) is composed of a plurality of words.

  Further, the content information may be an aggregate including a plurality of text data among text data included in the information content and text data added to the information content. In this case, each text data is constituted by one or more words, and each piece of content information is constituted by a plurality of words by constituting each text data.

  Thereafter, the parameter n for controlling the loop is initialized to 1 (S103).

  Then, the nth content is read from the content information D130 of the repository management subsystem 103 (S104), and if no content identifier is assigned to the read content, a new content identifier is assigned (S105). In step S112, the allocated content identifier is stored in the content information D130 (see FIG. 11) of the repository management subsystem 103.

  Thereafter, it is determined whether or not text information is included in the read content (S106). As a result, when text information is not included in the read content, it is not necessary to analyze the content, and the process proceeds to step S109.

  On the other hand, when text information is included in the read content, vocabulary appearing in the text information is extracted by morphological analysis, and the content index is updated based on the appearance frequency of the extracted vocabulary (S107). In step S112, the updated content index is stored in the content index D131 (see FIG. 11) of the repository management subsystem 103. In this specification, a “vocabulary” is a character string that is a single word or compound word constituting a sentence, and is a unit that is decomposed by morphological analysis described later.

  Next, the content vocabulary model is updated based on the appearance frequency of the combination of the co-occurrence vocabulary, with the vocabulary appearing within the predetermined range before and after the vocabulary extracted in step S107 as the co-occurrence vocabulary (S108). The updated content vocabulary model is stored in the content vocabulary model D132 (see FIG. 11) of the repository management subsystem 103 in step S112.

  Here, a method for acquiring the content vocabulary model will be described.

For example, consider the following sentence:
“My boss is Mr. Kasamatsu. Mr. Kasamatsu is the general manager.”
This sentence is divided into character strings (vocabulary) in units of words or compound words by morphological analysis as follows.

  "I (I) / no (no) / boss (boss) / ha (ha) / Kasamatsu (Kasamatsu) / san (san) / (is) /. (.) / Kasamatsu (Kasamatsu) / san (san) / Ha (ha) / department manager (department manager) / (en) / do (do) / te (te) / is (is) / do (do) /.

  The character string in parentheses indicates the basic vocabulary usage.

  Thereafter, the co-occurrence information acquisition process is executed on the decomposed character string (vocabulary). Specifically, based on the appearance position information of the vocabulary in the text, another vocabulary that is considered relevant in a certain vocabulary is used as a co-occurrence vocabulary, and the co-occurrence vocabulary and its frequency are acquired. For example, if there is a statistically significant increase in another vocabulary that appears within a certain range (a position several words away from immediately before or immediately after) in a certain vocabulary, there is a strong relationship between the two vocabularies. It is thought that there is. From this, n words (n> 0) before and after a certain vocabulary are collected as co-occurrence vocabulary, and the frequency of co-occurrence of the collected vocabulary is statistically processed to increase the frequency of co-occurrence. Vocabulary, that is, highly relevant vocabulary can be acquired. However, when acquiring co-occurrence vocabulary, it is collected and analyzed according to the basic vocabulary usage. This is to prevent a vocabulary having the same meaning from being determined as a different vocabulary when the vocabulary to be utilized such as a verb or an adjective is utilized. If symbols such as punctuation marks are included, they are not collected as a co-occurrence vocabulary.

  Here, a case where n = 4 is set is considered.

  Then, the vocabulary that co-occurs with “Kasamatsu” in the first sentence is up to four “I”, “No”, “boss”, “ha”, and four after the vocabulary subject to processing. It becomes "san", "is" up to. In addition, the vocabulary that co-occurs with “Kasamatsu” in the second sentence is “san”, “is”, “san”, “ha”, “ “General Manager” and “O”.

  Then, the frequency with which lexical pairs co-occur is calculated and used as the relevance. As a method of calculating the degree of association between vocabularies, a t-test method in statistics can be used in addition to a method of calculating and setting the number of co-occurrence between vocabularies. This makes it possible to estimate whether the frequency of the co-occurrence vocabulary is significantly high. Specifically, the following formula may be used.

  Further, as a method for calculating the degree of association between vocabularies, the calculation method based on the appearance position information of the vocabulary has been described, but the degree of association can also be calculated based on the structure (dependency) of the sentence. In this case, the dependency structure can be analyzed using a syntax analyzer, and the analyzed frequency information of the dependency relationship can be used as the relevance level.

  In addition to the above, related vocabulary can also be acquired using the method disclosed in Japanese Patent Laid-Open No. 2003-167894. However, in this case, a procedure for classifying the content into an appropriate category in advance is required.

  Furthermore, the degree of association can be calculated by combining these methods. In the present embodiment, description will be made using a method of calculating the number of co-occurrence between vocabularies based on the information on the appearance positions of vocabularies and setting the degree of association.

  Next, the user action history is updated by registering the parties related to the content in the user action history (S109). For example, information on the creator or updater of content is acquired from content attribute information (for example, file properties or file metadata), and the acquired user ID, action type, and content ID are updated The entry (frequency = 1) is registered in the user action history D137. When an entry having the same user ID, action type, and content ID has already been registered, 1 is added to the frequency. The updated user behavior history is stored in the user behavior history D137 (see FIG. 11) of the repository management subsystem 103 in step S112.

  Thereafter, the parameter n is compared with the number of elements in the list generated in step S101 to determine whether or not the next content to be analyzed exists (S110).

  As a result, when the next (n + 1) th content exists, 1 is added to the parameter n (S111), and the process returns to step S104 to analyze the next content.

  On the other hand, if there is no next (n + 1) th content, the analysis of all the content in the list generated in step S101 has been completed, so the updated content index, content vocabulary model, and user behavior history are stored in the repository. The data is transferred to the management subsystem 103 (S112).

  Upon receiving these data from the content analysis subsystem 101, the repository management subsystem 103 stores the received content identifier in the content information D130, stores the content index in the content index D131, and stores the content vocabulary model in the content vocabulary model D132. And the user action history is stored in the user action history D137.

  FIG. 4 is a diagram illustrating an example of a content index according to the embodiment of this invention.

  The content index updated by the content analysis subsystem 101 in step S107 is stored as a content index D131 in the storage device 120 of the repository management subsystem 103, and includes a vocabulary, a content ID, and a weight.

  The vocabulary is a word or a compound character string included in the content as text information. The content ID is an identifier for uniquely identifying the content, and indicates the content including the vocabulary. The weight indicates how often the vocabulary appears in this content.

  In addition to the method of setting the weight using the number of times that the vocabulary appears in the content, an index normalized by dividing the number of times the vocabulary appears in the content by the total number of vocabulary appearing in the content. The method to use, the method to use the probability of appearance of the vocabulary in the content set including this content, the method to determine according to the role (subject, predicate, object) in the sentence, etc. Various known methods such as a setting method can be used. In the present embodiment, description will be made using a method in which the vocabulary appears in this content.

  FIG. 5 is a diagram illustrating an example of the content vocabulary model according to the embodiment of this invention.

  The content vocabulary model generated by the content analysis subsystem 101 in step S108 is co-occurrence information of the vocabulary included in the content, and is stored as the content vocabulary model D132 in the storage device 120 of the repository management subsystem 103. , Vocabulary 2, content ID and relevance.

  Vocabulary 1 and vocabulary 2 are a set of vocabularies that co-occur within a predetermined range in the content. As described above, the vocabulary that co-occurs within a predetermined range can be determined as, for example, a vocabulary that co-occurs with vocabularies existing within four words before and after a certain vocabulary.

  The content ID is an identifier of content in which the vocabulary 1 and the vocabulary 2 co-occur. The relevance is the frequency at which the set of vocabulary 1 and vocabulary 2 co-occurs. For example, the number of times that the set of vocabulary 1 and vocabulary 2 co-occur in the content can be used. That is, it can be said that a vocabulary set having a large number of times of co-occurrence is highly related.

  FIG. 6 is a diagram illustrating an example of a user behavior history according to the embodiment of this invention.

  The user action history generated by the content analysis subsystem 101 in step S108 is stored as the user action history D137 in the storage device 120 of the repository management subsystem 103, and includes the user ID, action type, content ID, keyword, and frequency. .

  The user ID is an identifier for uniquely identifying a user who uses the information search system, and a name is used in the illustrated example.

  The action type is an identifier that uniquely identifies the type of action that the user has accessed the content. The user action history shown in FIG. 6 records “view” and “search” in addition to “create” and “update” of content, but some types of access (for example, “create”) , “Update” only) may be recorded. Further, other action types (for example, “meta tag assignment”) may be set and the corresponding action may be recorded.

  The content ID is an identifier that uniquely identifies the content accessed by the user. The keyword is a keyword (for example, a search keyword, a meta tag, etc.) that accompanies when the user acts on the content. The frequency is the number of times the user has accessed the content by the action.

  For example, in the user action history shown in FIG. 6, the content “CONT-1” “created” by the user “Takatori” is “updated” twice by the user “Takatori”, and the user “Nozaki” is “3”. It can be seen that the user “Nozaki” has searched “1” times.

  When the action type is an action accompanying content creation such as “creation” or “update”, the content analysis subsystem 101 operates the user action history. When the action type is an action that is not directly associated with content creation, such as “browse” or “search”, the information search server 104 operates the user action history.

  In the present embodiment, as a method of constructing the content index D131 and the content vocabulary model D132, a method of using text information existing in the target content has been described. However, not only the text information in the content but also the content The content index D131 and the content vocabulary model D132 can also be constructed using text information (content attributes, meta tags, etc.) attached to. For example, a content index can be constructed by executing step S107 based on content attributes and meta tags.

  Similarly, for example, it is possible to construct a content vocabulary model by determining that meta tags attached to the same content are related and executing step S108.

  Since these pieces of information are stored in the content vocabulary model D132 of the repository management subsystem 103, necessary information may be appropriately extracted from here and the corresponding process may be executed.

<User analysis subsystem 102>
FIG. 7 is a block diagram showing a configuration of the user analysis subsystem 102 according to the embodiment of this invention.

  The user analysis subsystem 102 has the same configuration as the content analysis subsystem 101 (FIG. 2) described above except that the stored program is different. For this reason, the same components as those of the content analysis subsystem 101 described above are denoted by the same reference numerals, and description thereof is omitted.

  That is, the user analysis subsystem 102 includes a memory 110, a storage device 120, a CPU 130, an output device 140, an input device 150, and a communication interface 160, and these components are computers connected by a bus 170.

  The memory 110 stores a program executed by the CPU 130. Specifically, the system control program P10 and the user analysis program P12 are stored in the memory 110. The user analysis program P12 is a program for analyzing the user's preference for content based on the content accessed by the user.

  The various programs D100 stored in the storage device 120 include a system control program P10 and a user analysis program P12, and are loaded into the memory 110 when executed by the CPU.

  The user analysis subsystem 102 transfers the user analysis result obtained by executing the user analysis program P12 to the repository management subsystem 103 via the network 106. Next, details of this processing will be described.

  FIG. 8 is a flowchart of processing executed by the user analysis subsystem 102 according to the embodiment of this invention.

  The process shown in FIG. 8 periodically analyzes all data, and the analysis result is overwritten and updated.

  First, the user analysis subsystem 102 acquires users to be analyzed, and generates a list of all users to be analyzed (S201). Then, the index D135 (see FIG. 11) of the user included in the generated list is acquired from the repository management subsystem 103 (S202). The user index file acquired in step S202 includes the analysis results of the user so far, and the configuration will be described later with reference to FIG.

  Further, the user action history D137 (see FIG. 11) is acquired from the repository management subsystem 103 (S203).

  Thereafter, the parameter n for controlling the loop is initialized to 1 (S204).

  Then, the nth user information is read from the user information D134 of the repository management subsystem 103 (S205). Then, it is determined whether or not the corresponding user information exists, that is, whether or not the user information included in the list generated in step S201 has been read from the user information D134 of the repository management subsystem 103 ( S206).

  As a result, when the user information exists, that is, when the user information can be read from the user information D134 of the repository management subsystem 103, the process proceeds to step S208. On the other hand, when the user information does not exist, that is, when the user information cannot be read from the user information D134 of the repository management subsystem 103, the user information is not registered in the repository management subsystem 103. Create user information. Therefore, a user identifier is assigned to the user, and the user information of this user is stored in the user information D134 of the repository management subsystem 103 (S207).

  Next, the user index file is updated (S208). As shown in FIG. 9, the user index file indicates the frequency of appearance of vocabulary included in content related to the user (accessed by the user such as creation and browsing). Specifically, the content accessed by the user is specified with reference to the user action history D137 acquired in step S203. Then, the appearance frequency of the vocabulary extracted from the identified content by morphological analysis is totaled over the entire content to obtain a weight. If the corresponding content has already been analyzed and information on the corresponding content is stored in the content index, the weight can be calculated while referring to the content.

  Thereafter, the user vocabulary model information (see FIG. 10) is updated (S209). The user vocabulary model can be updated by the same method as that for updating the content vocabulary model described above. That is, the frequency of appearance of a combination of co-occurrence vocabularies is calculated using the vocabulary appearing within a predetermined range before and after the vocabulary extracted in step S208 as a co-occurrence vocabulary. Then, the appearance frequency of the calculated combination of co-occurrence vocabulary for each content is identified by referring to the user action history D137, the content accessed for each user is identified, and the combination of the co-occurrence vocabulary of the identified content is identified. The appearance frequency is totaled for each user. This refers to the user action history D137, identifies the accessed content for each user, and totals the appearance frequency of the combination of the co-occurrence vocabulary of the identified content for each user. In step S212, the updated user vocabulary model is stored in the user vocabulary model D136 (see FIG. 11) of the repository management subsystem 103. If the corresponding content has already been analyzed and information on the corresponding content is stored in the content vocabulary model, the relevance can be calculated while referring to the content.

  Thereafter, the parameter n is compared with the number of elements in the list generated in step S201 to determine whether or not there is a next user to be analyzed (S210).

  As a result, when the next (n + 1) th user exists, 1 is added to the parameter n (S211), and the process returns to step S205 to analyze the next user.

  On the other hand, if the next (n + 1) th user does not exist, the analysis of all the users in the list acquired in step S201 is completed, and thus the updated user index and user vocabulary model are stored in the repository management subsystem 103. (S212).

  Upon receiving these data from the user analysis subsystem 101, the repository management subsystem 103 stores the received user index in the user index D135 and stores the user vocabulary model in the user vocabulary model D136.

  FIG. 9 is a diagram illustrating an example of a user index according to the embodiment of this invention.

  The user index updated by the user analysis subsystem 102 in step S208 is stored as a user index D135 in the storage device 120 of the repository management subsystem 103, and includes a vocabulary, a user ID, and a weight.

  The vocabulary is a word or a compound character string included in the content as text information. The user ID is an identifier for uniquely identifying the user, and indicates the user who has accessed the content including the vocabulary.

  The weight indicates the frequency that the vocabulary appears in the content accessed by the user. When calculating this weight, the number of appearances may be added by changing the weighting coefficient depending on the type of access. For example, “creation” can be multiplied by a factor of 1 and “update” can be multiplied by a factor of 0.5.

  According to the user index shown in FIG. 9, it can be seen that the vocabulary “BT” appears 10 times in all contents accessed by the user “Takatori”.

  FIG. 10 is a diagram illustrating an example of a user vocabulary model according to the embodiment of this invention.

  The user vocabulary model generated in step S209 by the user analysis subsystem 102 is vocabulary co-occurrence information included in the content accessed by the user, and is stored as the user vocabulary model D136 in the storage device 120 of the repository management subsystem 103. Vocabulary 1, vocabulary 2, user ID and relevance.

  Vocabulary 1 and vocabulary 2 are a set of vocabularies that co-occur within a predetermined range in the content. As described above, the predetermined range can be determined, for example, as a vocabulary that co-occurs with vocabularies existing within four words before and after the vocabulary 1.

  The user ID is an identifier of a user who has accessed content in which vocabulary 1 and vocabulary 2 co-occur. The relevance is the frequency at which the set of vocabulary 1 and vocabulary 2 co-occurs. For example, the number of times the set of vocabulary 1 and vocabulary 2 co-occurs in the content accessed by the user can be used. That is, it can be said that a vocabulary set having a high degree of relevance is a vocabulary set having high relevance of the vocabulary and high preference by the user.

  According to the user vocabulary model shown in FIG. 10, it can be seen that a set of “BT” and “SOA” appears five times as a co-occurrence vocabulary in all contents accessed by the user “Takatori”.

  In the present embodiment, as a method of constructing the user index D135 and the user vocabulary model D136, a method of constructing based on content information that the user has made, accessed, etc. has been described. In addition to information, it can also be constructed using information on content that has been subject to other actions such as search and meta tagging.

  For example, it is considered that a search keyword input when a search is executed and a meta tag given to the content by the user represent the user's preference. Therefore, a user index can be constructed by executing step S208 based on these pieces of information.

  Similarly, for example, it is determined that the vocabulary input by the user at the time of one search or the meta tags attached to the content accessed by the same user are related, and step S209 is executed. A user vocabulary model can also be constructed.

  Since these pieces of information are stored in the user action history D137 of the repository management subsystem 103, necessary information may be appropriately extracted from here and the corresponding process may be executed.

<Repository management subsystem 103>
FIG. 11 is a block diagram showing a configuration of the repository management subsystem 103 according to the embodiment of this invention.

  The repository management subsystem 103 has the same configuration as the content analysis subsystem 101 (FIG. 2) described above except that the stored program is different. For this reason, the same components as those of the content analysis subsystem 101 described above are denoted by the same reference numerals, and description thereof is omitted.

  That is, the repository management subsystem 103 includes a memory 110, a storage device 120, a CPU 130, an output device 140, an input device 150, and a communication interface 160, and these components are computers connected by a bus 170.

  The memory 110 stores a program executed by the CPU 130. Specifically, the system control program P10 and the repository management program P13 are stored in the memory 110. The repository management program P13 is a program for managing information stored in the repository management subsystem 103.

  The storage device 120 stores various programs D100. The various programs D100 include a system control program P10 and a repository management program P13, which are loaded into the memory 110 when executed by the CPU 130.

  The storage device 120 also stores content information D130, content index D131, content vocabulary model D132, content additional information D133, user information D134, user index D135, user vocabulary model D136, and user behavior history D137.

  The content information D130 includes a content identifier and a content entity. The content index D131 indicates the frequency of the vocabulary that appears in the content, as shown in FIG. The content vocabulary model D132 indicates the frequency of co-occurrence vocabulary appearing in the content, as shown in FIG. The content additional information D133 is information added to the content, and includes content attributes and meta tags.

  The user information D134 includes information about the user such as a user identifier, name, affiliation, and authority level. As shown in FIG. 9, the user index D135 indicates the frequency of appearance of the vocabulary included in the content accessed by the user. As shown in FIG. 10, the user vocabulary model D136 indicates the co-occurrence vocabulary included in the content accessed by the user and its frequency. As shown in FIG. 6, the user action history D137 indicates the mode and frequency with which the user has accessed the content.

  The repository management subsystem 103 stores the analysis information transferred through the network 106 in the storage device 120 by executing the repository management program P13, and receives and receives the information inquiry transmitted through the network 106. Information corresponding to the inquiry content is acquired from the magnetic disk device 120 and returned to the requesting information retrieval client 105 via the network 106. Next, details of this processing will be described.

  FIG. 12 is a flowchart of processing executed by the repository management subsystem 103 according to the embodiment of this invention.

  First, when the repository management subsystem 103 receives a request from the information search client 105 (S301), it analyzes the received request (S302).

  If the received request is an information reference request, information corresponding to the requested inquiry content is read from the storage device 120 (S303), and the read information is transmitted to the requesting information retrieval client 105 (S304). Then, it returns to step S301 and waits for reception of another request.

  On the other hand, when the received request is a content information registration request, the information is stored in the storage device 120 according to the requested registration content.

  Specifically, if it is a content information registration request, the received content information is newly registered in the content information D130 (S305). If it is a content index update request, the content index D131 is updated with the received content index (S306). If the request is for updating the content vocabulary model, the content vocabulary model D132 is updated with the received content vocabulary model (S307). If it is a content additional information registration request, the received content additional information is additionally registered in the content additional information D133 (S308). If it is a user action history registration request, the received user action history is additionally registered in the user action history D137 (S309). If it is a user information registration request, registration is performed such that the received information of the newly registered user is added to the user information D134 (S310). If it is a user index registration request, the user index D135 is updated with the received user index (S311). If it is a user vocabulary model update request, the user vocabulary model D136 is updated with the received user vocabulary model (S312). After the processing from step S305 to step S312 ends, the process returns to step S301 and waits for reception of another request.

  On the other hand, when the received request is a subsystem stop command, the operation of the repository management subsystem 103 is terminated.

<Information Search Server 104>
FIG. 13 is a block diagram showing a configuration of the information search server 104 according to the embodiment of this invention.

  The information search server 104 has the same configuration as the content analysis subsystem 101 (FIG. 2) described above except that the stored program is different. For this reason, the same components as those of the content analysis subsystem 101 described above are denoted by the same reference numerals, and description thereof is omitted.

  That is, the information search server 104 includes a memory 110, a storage device 120, a CPU 130, an output device 140, an input device 150, and a communication interface 160, and these components are computers connected by a bus 170.

  The memory 110 stores a program executed by the CPU 130. Specifically, the system control program P10 and the information search control program P14 are stored in the memory 110.

  The information search control program P14 is a program for searching for content based on a search request transmitted from the information search client 105. As subprograms, a content search program P141, a search result acquisition program P142, a content acquisition program P143, a user An action history registration program P144 and a user authentication program P145 are included.

  The content search program P141 is a program for executing a content search, and includes a search condition expression analysis program P1411, a search query matrix generation program P1412, a content matrix generation program P1413, a similarity calculation program P1414, and a content extraction program P1415. .

  The search condition expression analysis program P1411 analyzes the search condition expression input by the user (step S411 in FIG. 14). The search query matrix generation program P1412 generates a search query matrix according to the analysis result of the search condition formula (step S412 in FIG. 14).

  The content matrix generation program P1413 generates a content matrix whose similarity is compared with the search query matrix (step S4134 in FIG. 17). The similarity calculation program P1414 calculates the similarity between the search query matrix and the content matrix (step S4135 in FIG. 17).

  The content extraction program P1415 searches the content index D131, and extracts a list of content that includes the vocabulary of the label of the search query matrix (step S4132 in FIG. 17).

  The search result acquisition program P142 transfers the search result data D140 to the request source according to the inquiry content from the information search client 105.

  The content acquisition program P143 acquires content from the repository management subsystem 103 in accordance with a request from the information search client 105.

  The user action history registration program P144 registers the access history to the content according to the request from the information search client 105 in the user action history D137.

  The user authentication program P145 authenticates a user who operates the information search client 105.

  The storage device 120 stores various programs D100. The various programs D100 include a system control program P10 and an information retrieval control program P14, and are loaded into the memory 110 when executed by the CPU 130.

  The storage device 120 stores search result data D140. The search result data D140 is a search result searched by the content search program P141, and is transferred to the information search client 105 by the search result acquisition program P142.

  The information search server 104 searches the content based on the search request transmitted from the information search client 105 by executing the information search control program P14, and returns the search result to the information search client 105 as the request source. Next, details of this processing will be described.

  FIG. 14 is a flowchart of processing executed by the information search server 104 according to the embodiment of this invention.

  First, upon receiving a request from the information search client 105 (S401), the information search server 104 analyzes the received request (S402).

  If the received request is a user authentication request, the user is authenticated based on the user ID input to the information search client 105, and the authentication result is returned to the requesting information search client 105 (S403).

  On the other hand, if the received request is a search result query, the search result acquisition program P142 is executed, the search result data D140 is read from the storage device 120 according to the query content from the information search client 105, and the search result data D140 is obtained from the request source. The information is transferred to the information search client 105 (S421). The search result data D140 includes a list of searched contents.

  On the other hand, if the received request is a content transfer request, the content content acquisition program P143 is executed, the request from the information search client 105 is transferred to the repository management subsystem 103, and the requested content is transferred from the repository management subsystem 103. Obtain (S431). Then, the user action history registration program P144 is executed, and the cause (for example, browsing) of the content transfer request is additionally registered in the user action history D137 of the repository management subsystem 103 (S432). Thereafter, the requested content is transferred to the requesting information retrieval client 105 (S433).

  On the other hand, if the received request is a content search request, the search condition expression analysis program P1411 is executed to analyze the search condition expression (S411). Specifically, the user ID of the searcher is acquired, and the text input by the searcher is divided into vocabulary using morphological analysis. Then, the appearance frequency of the vocabulary is calculated, and the search condition formula is vectorized. The vectorized search condition formula is stored in the work area of the memory 110 of the information search server 104. Next, the search query matrix generation program P1412 is executed to generate a search query matrix (S412). Details of the search query matrix generation processing will be described later with reference to FIG. In the embodiment of the present invention, the search condition expression is analyzed in step S411, a search query matrix is generated, and the search is executed. Therefore, even if the search condition expression is input as a natural sentence, an appropriate search is performed. Can do.

  Then, based on the search query matrix, the similarity of each content in the content information D130 is calculated (S413), and each content in the content information D130 is sorted in descending order of similarity (S414). Details of the similar content extraction processing will be described later with reference to FIG.

  Thereafter, the user behavior history registration program P144 is executed to additionally register the user behavior history (S415). The sorted search result data set is returned to the requesting information search client 105 (S416).

  On the other hand, when the received request is a server stop command, the operation of the information search server 104 is terminated.

  15A and 15B are flowcharts illustrating a detailed procedure of the information search query matrix generation process (S412) according to the embodiment of this invention.

  First, the search keyword and its weight value obtained by analyzing the search condition formula in step S411 are read from the work area (S41201).

  Then, a square zero matrix of the order of a value obtained by adding 1 to the number of search keywords is created, and the created matrix is initially set as a search query matrix (S41202).

  Then, an abstract node and a search keyword are set in the row label of the created search query matrix, and an abstract node and a search keyword are also set in the column label of the search query matrix (S41203). That is, the same abstract node and search keyword are set for the row label and the column label of the search query matrix. The abstract node is set to the label in the first row and the first column.

  Thereafter, a parameter n for controlling the search keyword to be processed is initialized to 1 (S41204). Note that the maximum value of n is the number of search keywords acquired in step S41201.

  Then, the nth search keyword is read from the work area (S41205).

  The weight value acquired in step S40201 is set to the search query matrix value <1, n> corresponding to the abstract node and the nth search keyword. Similarly, the same weight value is set for the search query matrix value <n, 1> corresponding to the nth search keyword and abstract node (S41206).

  Thereafter, a list of vocabulary related to the nth search keyword is extracted from the user vocabulary model D136 (S41207). Specifically, the vocabulary 1 of the user vocabulary model D136 is searched based on the nth search keyword and the user ID, the vocabulary set of the corresponding vocabulary 2 is acquired, and set to the related vocabulary set. Similarly, the vocabulary 2 of the user vocabulary model D136 is searched based on the nth search keyword and the user ID, the vocabulary set of the corresponding vocabulary 1 is acquired, and set to the related vocabulary set. The acquired related vocabulary is stored in the work area of the memory 110.

  Thereafter, the parameter k for controlling the related vocabulary to be processed is initialized to 1 (S41208). Note that the maximum value of the parameter k is the number of elements of the related vocabulary set acquired in step S41207.

  Thereafter, the kth related vocabulary is read from the work area (S41209). Then, with reference to the row and column labels of the search query matrix, it is determined whether or not the read k-th related vocabulary exists in the search query matrix (S41210). As a result, if the k-th related vocabulary exists in the label of the search query matrix, it is not necessary to add the k-th related vocabulary to the search query matrix, so the process proceeds to step S41213.

  On the other hand, if the k-th related vocabulary does not exist in the search query matrix label, the search query matrix is expanded by 1 row and 1 column, and the expanded row and column elements are set to 0 (S41211). The k-th related vocabulary is set in the row and column labels (S41212).

  Thereafter, it is determined whether or not the (k + 1) th related vocabulary exists (S41213). As a result, when the k + 1-th related vocabulary exists, 1 is added to the parameter k (S41215), and the process returns to step S41209 to process the next related vocabulary.

  On the other hand, if the k + 1-th related vocabulary does not exist, the processing of the related vocabulary related to this search keyword is completed, so it is determined whether or not the n + 1-th search keyword exists (S41214). As a result, when the (n + 1) th search keyword exists, 1 is added to the parameter n (S41216), and the process returns to step S41205 to process the next search keyword.

  On the other hand, if the (n + 1) th search keyword does not exist, the generation of the search query matrix including all the related vocabulary necessary for the search is completed, and the process proceeds to step S41217. The generated search query matrix is stored in the work area of the memory 110.

  In step S41217, after the parameter i indicating the row of the search query matrix is initialized to 2 (S41217), the label (vocabulary) of the i row of the search query matrix is read from the work area (S41218). Thereafter, after initializing the parameter j indicating the column of the search query matrix to 2 (S41219), the label (vocabulary) of the j column of the search query matrix is read (S41220).

  Thereafter, the degree of association between the i-row label of the search query matrix and the label of the j column is obtained from the user vocabulary model D136 (S41221), and the obtained degree of association is set in <i, j> of the search query matrix. However, if a related word cannot be acquired in step S41221, nothing is executed and the process proceeds to the next step (S41222).

  Thereafter, it is determined whether or not the j + 1-th label (vocabulary) exists (S41223). As a result, when the j + 1-th vocabulary exists, 1 is added to j (S41226), and the process returns to step S41120 to process the next column. On the other hand, if the j + 1-th vocabulary does not exist, the processing for the i-th row is completed, so that it moves to the next row, and it is determined whether or not the i + 1-th label (vocabulary) exists (S41224).

  As a result, when the i + 1th vocabulary exists, 1 is added to i (S41227), and the process returns to step S41218 to process the next i + 1 line. On the other hand, if the i + 1th vocabulary does not exist, registration of the degree of association between the vocabularies in the search query matrix is completed, so the generated search query matrix is stored in the work area (S41225), and information search query matrix generation processing ( S412) is ended, and the process proceeds to step S413.

  FIG. 16 is a diagram illustrating an example of generating a search query matrix by analyzing a search condition formula in steps S411 and S412 according to the embodiment of the present invention.

  In step S411, a case is considered in which the searcher inputs the sentence "about an effective system construction method using BT ..." 1701. This sentence 1701 is divided into vocabulary using morphological analysis. Then, the appearance frequency of the vocabulary sentence 1701 is calculated (1702), and the search condition expression is vectorized. This generated vector 1703 is included in the first column of the search query matrix 1704.

  Next, in step S412, a search query matrix 1704 in which the vector 1703 is set to the first row and the first column is generated. In the elements in the second row and the second column of the search query matrix 1704, the frequency with which the vocabulary recorded in the user vocabulary model D136 co-occurs is registered.

  In this embodiment, the search method when the searcher inputs the sentence 1701 as the search term has been described. However, the searcher is requested to input the search term and its weight, and without performing morphological analysis, A search query can be created based on the input search term and the weight for it. When the searcher omits the input of the weight for the search word, a default value (for example, 1) may be set.

  Further, by requesting the searcher to input the content ID of the content related to the content to be searched, it is possible to create a search query matrix based on the input content ID without performing morphological analysis. Specifically, first, if a vocabulary set is acquired from the content index D131 based on the content ID and the top n items of the weight are acquired, the vocabulary set that characterizes the content that the user wants to search can be acquired. A search query matrix can be created by executing the same processing based on the obtained vocabulary set, and the content desired by the searcher can be presented.

  Furthermore, the present invention can be applied to a process for recommending content that matches the user's preference. In this case, if a vocabulary set is obtained from the user index D135 based on the user ID and the top n weights are obtained, a vocabulary set that characterizes the user can be obtained. Then, based on the vocabulary set that characterizes the user, a search query matrix is created using a similar process, so that content that matches the user's preference can be recommended.

  FIG. 17 is a flowchart illustrating a detailed procedure of the similar content extraction process (S413) according to the embodiment of this invention.

  First, the search query matrix generated in step S412 is read from the work area (S4131). After that, the content extraction program P1415 is executed, the content index D131 is searched using the labels of the search query matrix, and the list of contents including the vocabulary of the labels of the search query matrix is extracted (S4132).

  Thereafter, the parameter n for controlling the loop is initialized to 1 (S4133). Then, the content matrix generation program P1413 is executed to generate a content matrix of the extracted nth content (S4134). Details of the content matrix generation processing will be described later with reference to FIG.

  Then, the similarity calculation program P1414 is executed to calculate the similarity between the search query matrix and the generated nth content matrix (S4135). Details of the content matrix generation processing will be described later with reference to FIG.

  Thereafter, the calculated similarity is associated with the content and stored in the work area (S4136).

  Thereafter, it is determined whether or not the (n + 1) th content exists (S4137). As a result, if the (n + 1) th content exists, 1 is added to n (S4138), and the process returns to step S4134 to process the next content. On the other hand, if the (n + 1) th content does not exist, the processing of all the extracted contents is complete, so the similar content extraction process is terminated and the process proceeds to step S414.

  FIG. 18 is a flowchart illustrating a detailed procedure of the content matrix generation procedure (S4134) according to the embodiment of this invention.

  First, a square zero matrix having the same order as the order of the search query matrix is created, and the created matrix is initialized as a content matrix (S413401).

  Then, the row and column labels of the search query matrix are set to the row labels of the created content matrix (S413402). That is, the same abstract nodes and search keywords as the row and column labels of the search query matrix are set in the same order in the row and column labels of the content matrix. The abstract node is set to the label in the first row and the first column.

  Thereafter, the parameter n indicating the vocabulary set in the label of the content matrix is initialized to 2 (S413403), and the label (vocabulary) of n rows in the content matrix is read from the work area (S413404).

  Thereafter, the weights of the labels (vocabulary) of n rows of the content matrix are obtained from the content index D131 based on the labels and content IDs of the n rows, and the obtained weights are <1, n>, <n, 1> is set (S413405).

  Thereafter, it is determined whether or not the (n + 1) th label (vocabulary) exists (S413406). As a result, if the (n + 1) th vocabulary exists, 1 is added to n (S413407), and the process returns to step S413404 to process the next vocabulary. On the other hand, if the (n + 1) th vocabulary does not exist, the abstract node processing is completed, so that the parameter i indicating the row of the search query matrix is initialized to 2 in order to move to the search keyword processing next (S413408).

  Thereafter, the label (vocabulary) of i rows of the content matrix is read from the work area (S413409). Thereafter, after initializing the parameter j indicating the column of the content matrix to 2 (S413410), the label (vocabulary) of the j column of the content matrix is read (S413411).

  Thereafter, the degree of association between the i-line label of the content matrix and the j-column label is obtained from the content vocabulary model D132 based on the i-line label, the j-column label, and the content ID (S413212), and the obtained association The degree is set to <i, j> of the search query matrix. However, if the degree of association cannot be acquired, nothing is done and the process proceeds to the next step (S413413).

  Thereafter, it is determined whether or not the j + 1-th label (vocabulary) exists (S413414). As a result, when the j + 1-th vocabulary exists, 1 is added to j (S413417), and the process returns to step S413411 to process the next column. On the other hand, if the j + 1-th vocabulary does not exist, the processing for the i-th row is completed, so that it moves to the next row, and it is determined whether or not the i + 1-th label (vocabulary) exists (S413415).

  As a result, if the i + 1th vocabulary exists, 1 is added to i (S413418), and the process returns to step S413409 to process the next i + 1 line. On the other hand, if the i + 1-th vocabulary does not exist, registration of the degree of association between the vocabularies in the content matrix is completed, so the generated content matrix is stored in the work area (S413416), and the content matrix extraction process (S4134) is terminated. Then, the process proceeds to step S4145.

  FIG. 19 is a diagram illustrating an example of generating a content matrix in step S413 according to the embodiment of this invention.

  First, in step S4132, the corresponding content is filtered using the label of the search query matrix. This step is the same as step S4132 of FIG. Specifically, the content index D131 is searched using the labels of the search query matrix 1704, and the content list 2000 including the vocabulary of the labels of the search query matrix is extracted.

Next, in step S4134, a content matrix 2001 is generated. The row and column labels of the search query matrix are set in the row and column labels of the content matrix 2001 generated here. In the first row and the first column of the content matrix 2001, the content recorded in the content index D131 and the vocabulary weight are registered. In the elements in the second row and second column of the content matrix 2001, the degree of association between vocabularies recorded in the content vocabulary model D132 is registered.

  Thereafter, in step S4135, the similarity between the search query matrix and the content matrix is calculated (S4135).

  FIG. 20 is a flowchart illustrating a detailed procedure of the similarity calculation procedure (S4135) according to the embodiment of this invention.

  First, the search query matrix generated in step S412 and the content matrix generated in step S4134 are read out from the work area (S4133501).

  Thereafter, the parameter step SUM for calculating the similarity is initialized to 0 (S413502), and the parameter n for controlling the search query matrix and the content matrix column is initialized to 1 (S413503).

  Then, the nth column of the search query matrix is vectorized (QVn), and the vector QVn is stored in the work area (S413504). Furthermore, the nth column of the content matrix is vectorized (CVn), and the vector CVn is stored in the work area (S413505).

  Thereafter, the similarity between the vector QVn and the vector CVn stored in the work area is calculated. For the calculation of the similarity, a technique such as cosine similarity can be used (S413506). The cosine similarity can be obtained by dividing the inner product between the vectors QVn and CVn by the vector norm (the geometric length of the vector). Then, the calculated similarity between vectors is added to the SUM (S413507).

  Thereafter, it is determined whether or not n + 1 columns exist (S413508). As a result, when the (n + 1) th column exists, 1 is added to n (S413510), and the process returns to step S413504 to process the next column. On the other hand, if the n + 1-th column does not exist, the similarity is calculated by dividing SUM by the number of dimensions of the search query matrix (S413509). Thereafter, the similarity calculation procedure (S4135) is terminated, the process proceeds to step S4136, and the calculated similarity is stored in the work area.

  The method for calculating the similarity between the search query matrix and the content matrix using the cosine similarity (the inner product between the vectors) has been described above.

  In this embodiment, the search query matrix creation method and the content matrix creation method have been described with respect to the method of using the analysis results stored in various storage devices as they are. Matrixes and content matrices can also be created.

  An overview of the search query matrix or the content matrix can be thought of as a graph structure in which the labels are nodes and the element values in the matrix are edges. At this time, the search query matrix or the content matrix can be considered as a matrix form called an adjacency matrix in the graph theory.

  By calculating the similarity between nodes constituting this graph structure, it is possible to increase the accuracy of the relationship between the search query or content and the vocabulary and the relationship between the vocabularies. For this reason, the accuracy of the calculated similarity can be increased.

  There are several known methods for calculating the similarity between nodes in a graph, such as a method using the von Neumann matrix in the field of graph theory and a method using the Laplacian heat expansion matrix. Can be used.

  For example, the similarity (RL) between nodes in the search query matrix can be calculated using a regularized Laplacian kernel. Specifically, the following formula may be used.

  The value (RL) obtained by the following equation is a matrix, and the element values of this matrix indicate the similarity between nodes, that is, the relationship between the search query and the vocabulary, and the relationship between each vocabulary.

  Similarly, the similarity (RL) between nodes in the content matrix can be calculated. In this case, the value (RL) obtained by the above equation is a matrix. Element values in this matrix indicate the degree of similarity between nodes, that is, the degree of relationship between content and vocabulary, and the relationship between each vocabulary.

  Then, using the search query matrix and the content matrix reconstructed by applying the regularized Laplacian kernel, the similarity can be calculated by the method of step S4135 described above.

  Note that the regularized Laplacian kernel may be applied to either the search query matrix or the content matrix. The regularized Laplacian kernel is applied for the purpose of improving the accuracy of element values (relevance) in the search query matrix or content matrix, and even if the regularized Laplacian kernel is applied, the search query matrix or content matrix is applied. This is because the meanings of the labels and the element values in are not changed.

<Information Search Client 105>
FIG. 21 is a block diagram illustrating a configuration of the information search client 105 according to the embodiment of this invention.

  The information search client 105 has the same configuration as the content analysis subsystem 101 (FIG. 2) described above except that the stored program is different. For this reason, the same components as those of the content analysis subsystem 101 described above are denoted by the same reference numerals, and description thereof is omitted.

  That is, the information search client 105 is a computer that includes a memory 110, a storage device 120, a CPU 130, an output device 140, an input device 150, and a communication interface 160, and these components are connected by a bus 170.

  The memory 110 stores a program executed by the CPU 130. Specifically, the system control program P10 and the search client control program P15 are stored in the memory 110.

  The search client control program P15 is a program that generates a search request to be transmitted to the information search server 104, and includes a search condition input program P151 and a search result display program P152 as subprograms.

  The search condition input program P151 receives an input of search conditions from the user and transmits a search request to the information search server 104. The search result display program P152 displays search results according to instructions from the user.

  The storage device 120 stores various programs D100. The various programs D100 include a system control program P10 and a search client control program P15, and are loaded into the memory 110 when executed by the CPU 130.

  Further, the storage device 120 stores search result data D150. The search result data D150 is a cache in which the search result transferred from the information search server 104 is temporarily stored.

  The information search client 105 generates a search request to be transmitted to the information search server 104 by executing the search client control program P15, and displays the result of the search performed by the information search server 104. Next, details of this processing will be described.

  22A and 22B are flowcharts of processing executed by the information search client 105 according to the embodiment of this invention.

  First, a user authentication screen is displayed to prompt input of user authentication information (S501). Then, user authentication information input by the user is acquired (S502), and the acquired user authentication information is sent to the information search server 104 to request user authentication (S503).

  Thereafter, when the result of user authentication is received from the information search server 104, it is determined whether or not the authentication is successful based on the received result of user authentication (S504). As a result, if the authentication is unsuccessful, the process returns to step S501, the authentication failure is displayed, and the user authentication information is further requested. On the other hand, if the authentication is successful, a search condition input screen is displayed to prompt input of an instruction (command) (S505).

  Thereafter, when a command is input (S506), the input command is analyzed (S507).

  If the analyzed command is a client stop command, the information retrieval process is terminated. On the other hand, if the analyzed command is a content search command, the search condition input program P151 is executed, a search request is generated based on the input data (S508), and the generated search request is transmitted to the information search server 104. (S509).

  Thereafter, when a search result is received from the information search server 104, the received search result is stored in the search result data D150 of the storage device 120 (S510). The content search result includes the search result identifier and all or a subset of the content identifier in the searched content set, and is transmitted from the information search server 104 in step S416 in FIG.

  Thereafter, the search result display program P152 is started, a search result display / instruction input screen is displayed, and input of instruction contents is prompted (S511). Then, the instruction content input by the user is acquired (S512), and the instruction content from the user is analyzed (S513).

  If the analyzed instruction content is a detected content identifier list display instruction, a content search result identifier inquiry request is created, and the created request is transmitted to the information search server 104 (S514). Thereafter, when a set of content identifiers is received from the information search server 104, a list of received content identifiers is displayed (S515).

  If the analyzed instruction content is a content content display instruction, a content transfer request is created, and the created request is transmitted to the information search server 104 (S516). Thereafter, when content data is received from the information search server 104, the received content data is stored in the search result data D150 (S517). Then, the received content data is converted into a predetermined form and displayed (S518).

  If the analyzed instruction content is the end of the search result display, the process returns to step S506 and further receives a command.

  As described above, according to the embodiment of the present invention, not only the relationship between the content and the search keyword included in the search query, but also the relationship between the vocabulary included in the content and the search keyword included in the search query. Consider the relationship of vocabulary. For this reason, even if the vocabulary is ambiguous, the similarity of the content related to the vocabulary used by the searcher is high, and the similarity of the content related to the vocabulary not used by the searcher is low. Become. That is, by using the user vocabulary model, it is possible to correct fluctuation of words having various meanings, display content close to the user's preference, and improve search accuracy.

101 Content Analysis Subsystem 102 User Analysis Subsystem 103 Repository Management Subsystem 104 Information Search Server 105 Information Search Client 106 Network

Claims (13)

An information retrieval system configured by one or more computers to retrieve text information,
The text information is information including at least one of text data included in information content and text data added to the information content,
The calculator is
Means for acquiring a first relationship between words included in the search condition by analyzing the search condition input by the user;
Means for analyzing the stored text information to obtain a second relationship between words included in the text information;
An information search system comprising: means for selecting text information to be output as being matched with an input search condition according to the similarity between the acquired first relationship and second relationship. The information search system manages a user vocabulary model in which relevance of words in text information related to the user is recorded,
The means for acquiring the first relationship is:
Extract words included in the input search condition by morphological analysis,
The search query matrix in which a relationship between the extracted words is defined is generated as a representation of the first relationship by referring to the user vocabulary model. Information retrieval system. The computer includes user vocabulary model generation means for generating the user vocabulary model,
The user vocabulary model generating means includes
Identify text information accessed by the user and text information related to the data accessed by the user;
Extracting words included in the identified text information by morphological analysis,
Collecting words that exist within a predetermined co-occurrence range among the extracted words,
3. The information retrieval system according to claim 2, wherein the relevance of the words is obtained by counting the number of appearances of the collected words after being weighted according to the type of access. The information search system manages a content vocabulary model in which relevance of words included in the stored text information is recorded,
The means for acquiring the second relationship is:
Extracting words contained in the stored text information by morphological analysis,
4. The content matrix in which the relationship between the extracted words is defined is generated as the first relationship by referring to the content vocabulary model. The information search system according to any one of the above. The computer includes content vocabulary model generation means for generating the content vocabulary model,
The content vocabulary model generation means includes:
Extracting words contained in the stored text information by morphological analysis,
Collecting words existing in a predetermined co-occurrence range among words included in the accumulated text information,
5. The information search system according to claim 4, wherein the relevance of the word is obtained by counting the number of appearances of the collected word.   The information search system calculates an inner product between a search query matrix indicating the first relationship and a content matrix indicating the second relationship by calculating a norm of the search query matrix and a norm of the content matrix. 6. The information search system according to claim 1, further comprising means for calculating a similarity between the first relevance and the second relevance by dividing by a product. An information search method in a search system configured by one or more computers and searching for text information,
The text information is information of text data included in information content or text data added to the information content,
The computer includes a processor that executes a program, and a memory that stores the program to be executed.
The method
The computer obtains a first relevance between words included in the search condition by analyzing the search condition input by a user, and writes the first relevance in the memory;
The computer obtains a second relationship between words included in the text information by analyzing the stored text information and writes the second relationship to the memory;
The computer selecting and outputting text information to be output as matching the input search condition according to the similarity between the acquired first relationship and second relationship. Information search method characterized by The method includes the step of the computer managing a user vocabulary model in which word associations in text information associated with the user are recorded;
In the step of obtaining the first relevance, the computer extracts words included in the input search condition by morphological analysis, and refers to the user vocabulary model, thereby extracting the words between the extracted words. The information search method according to claim 7, further comprising: generating a search query matrix in which the relevance is defined as representing the first relevance.   In the method, the computer specifies text information accessed by a user and text information related to data accessed by the user, extracts words included in the specified text information by morphological analysis, and extracts the words. Of words collected within a predetermined co-occurrence range, the number of appearances of the collected words is weighted according to the type of access, and the relevance of the words is obtained. The method according to claim 8, further comprising: generating the user vocabulary model by: The method includes the step of the computer managing a content vocabulary model in which relevance of words included in the stored text information is recorded,
In the step of obtaining the second relevance, the computer extracts words included in the accumulated text information by morphological analysis, and refers to the content vocabulary model to thereby extract the extracted words. The information search method according to any one of claims 7 to 9, wherein a content matrix in which a relevance is defined is generated as representing the first relevance.   In the method, the computer extracts words included in the accumulated text information by morphological analysis, and among the words included in the accumulated text information, the words exist within a predetermined co-occurrence range. The information retrieval method according to claim 10, further comprising: obtaining the relevance of the word by collecting the number of appearances of the collected word and generating the content vocabulary model thereby. Method.   In the method, the computer calculates an inner product between a search query matrix indicating the first relevance and a content matrix indicating the second relevance by using a norm of the search query matrix and a norm of the content matrix. The information retrieval method according to claim 7, further comprising: calculating a similarity between the first relationship and the second relationship by dividing by a product of Method. An information retrieval server that retrieves text information according to a request from a client,
The text information is information of text data included in information content or text data added to the information content,
The information search server
Means for acquiring a first relationship between words included in the search condition by analyzing the search condition input by the user;
Means for analyzing the stored text information to obtain a second relationship between words included in the text information;
An information search server comprising: means for selecting text information to be output as a match with an input search condition based on the similarity between the acquired first relationship and second relationship.

Images