JP2013025421A - Information retrieval system and information retrieval method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information retrieval system and an information retrieval method capable of outputting a retrieval result reflecting popularity, topicality and accuracy of description contents or the like of a Web page in a short time.SOLUTION: In the information retrieval system for retrieving a Web page on the basis of a keyword inputted in a terminal which has executed retrieval, a transition probability is computed from the association of a Web page to which a Web page to be a retrieval object is linked and a terminal connected to the Web page to be the retrieval object at present or within a fixed period. The Web page to be the retrieval object is sequenced on the basis of a computed value and outputted to the terminal.

Description

  The present invention relates to a technique for efficiently and easily acquiring target information in a WWW (World Wide Web) information search system.

  When obtaining target information from a vast amount of information on the WWW, it is common to use a search system to identify the corresponding Web page. In such a search system, Web pages corresponding to the search keyword input by the user are specified and collected, and the collected Web pages are given importance and presented in descending order of importance.

  Here, as a technique for assigning importance to collected Web pages, a technique for ranking Web pages based on the number of appearances of nouns in a Web page, a link structure, and the like is widely known. Specifically, for example, a high importance is given to a Web page in which a search keyword input by the user is used in an important part such as a title, or a Web page in which a search keyword input by the user appears moderately many times. The method of attaching is known. Also, there is known a method of assigning a high importance to a Web page linked from many other Web pages or a Web page linked from a Web page having high content relevance.

  However, if the importance of the Web page is defined based on only the content and link structure of the Web page in this way, if the user cannot select a search term well, the Web page that discloses the target information is displayed at the top. Therefore, there has been a problem that a Web page in which information intended by the user is disclosed cannot be found. Furthermore, even if the Web page in which the target information is disclosed can be identified, there is a problem that the user cannot obtain a deep understanding if the target information is difficult to understand.

  Therefore, in view of such problems, the inventors assign importance to Web pages based on the popularity and topicality of each Web page collected as a search result, accuracy of description contents, etc. An information search system and an information search method that can be presented in order are proposed in the previous application (Patent Document 1).

  Specifically, in the information search system proposed by the present inventors, a plurality of Web pages collected as search results are rearranged in descending order of the number of users browsing each Web page at the time of search execution. Search results can be displayed. That is, the Web page that many users are paying attention to when executing a search is a Web page with high topicality (necessity) at the time of executing the search, and by displaying such a Web page at the top of the search results. This makes it easy to obtain information that is highly necessary when performing a search.

  In the information search system proposed by the present inventors, information on all users who use the information search system can be acquired in advance, and the acquired information can be reflected in the search result. For example, when a specific user performs a search, the information search system determines which user is browsing each Web page collected as a search result, and a user who satisfies a specific condition browses The Web page being displayed is displayed at the top. For example, if the user who executed the search performs a search using the keyword “program”, the Web page viewed by the user whose “profession” is “programmer” is displayed at the top. In this way, if the web page collected as a search result is a web page being browsed by a user who is an expert in a field related to the search keyword, the web page is displayed at the top. In other words, Web pages that are viewed by many experts are Web pages that are rich in content (content accuracy is high), and by displaying such Web pages at the top of the search results, the user has high accuracy. It is easy to obtain information.

  The information search system proposed by the present inventors can communicate with other users who are browsing the same Web page when browsing the specified Web page after executing the search. As a result, even if the content of the specified Web page is difficult, the user can deepen understanding of the content of the Web page by asking other users questions.

JP 2011-39625 A

  As described above, according to the information search system that assigns importance to Web pages based on the popularity and topicality of each Web page collected as a search result, the accuracy of description, etc., and presents them in descending order of importance. Information that the user really needs can be acquired more easily. Therefore, according to the present invention, it is possible to calculate the importance of the Web page reflecting the popularity, the topicality, the accuracy of the description content, and the like more accurately, and the search result reflecting the calculated importance of the Web page can be obtained in a short time. It is an object of the present invention to provide an information search system and an information search method that can be output.

One aspect of the present invention for solving the above problems is information for searching a Web page based on a keyword input at a terminal that has executed a search and outputting the search result to the terminal for display. The search system includes information ordering means for ordering Web pages to be searched. The information ordering means includes a Web page linked to the Web page to be searched, a search target, An information search system having a function of performing an operation for estimating a transition tendency from a relation with a terminal that is currently connected to a Web page or within a certain period of time.
The “calculation for estimating a transition tendency” includes a calculation performed after modifying an element necessary for calculating the transition probability in addition to the calculation of the actual transition probability. In addition, “transition” refers to displaying another Web page by tracing a link from a specific Web page with the Web page or terminal as a vertex. Furthermore, when a specific web page is output by a specific terminal, a transition similar to that in which a link is formed between two web pages including the specific web page is assumed.
In addition, “transition similar to that in which a link is formed between two Web pages including a specific Web page” assumed at this time is “a mutual link is formed between the two Web pages. It is desirable to use “transition to be”. This is because the Web that is viewed by the user compared to the case where “a transition in which a link is formed from one Web page to the other Web page between two Web pages” is assumed. This is because an increase in the importance of the page is more accurately reflected.

  In the second aspect of the present invention, the calculation of transition probability sets a square matrix by a configuration including a Web page to be searched and a terminal that is currently connected to the Web page to be searched or within a certain period of time. The information search system according to claim 1, further comprising: calculation content for calculating an eigenvector of the square matrix.

  According to a third aspect of the present invention, the elements of the square matrix are based on a value based on the number of link destinations to which the web page is linked and a value based on the number of web pages to which the terminal is actually connected within a certain period. The information search system according to claim 2, wherein the information search system is configured.

  In a fourth aspect of the present invention, the element of the square matrix is A based on the number of link destinations to which the Web page is linked, and B is the number of Web pages that the terminal has actually connected within a certain period. The information search system according to claim 2 or 3, wherein A / (A + B), B / (A + B) or a value obtained by multiplying these by a system number.

The fifth aspect of the present invention has information acquisition means for acquiring terminal information, which is information relating to a terminal outputting a Web page to be searched, at the time of search execution or before search execution, The ordering means identifies a Web page that satisfies a predetermined condition based on at least terminal information, and an output terminal that is outputting or outputting the Web page. When it is assumed that a mutual link is formed between the web page output from the output terminal and the virtual web page, the link source is based on the link structure formed by the web page and the virtual web page. Set a transition probability matrix with a transition from the display of the Web page to the display of the linked Web page as a transition, and transition based on the set transition probability matrix It calculates the rate is information retrieval system according to any one of claims 1 to 3, characterized in that ranks the Web pages to be searched according to the result.
The link structure includes a virtual link, and the link source Web page or the link destination Web page includes a virtual page.

  According to such an information search system, it is possible to more accurately calculate the importance level of a Web page that reflects popularity, topicality, accuracy of description contents, and the like, and a search result that reflects the calculated importance level of the Web page Can be output in a short time. Therefore, even if the amount of calculation becomes enormous due to an increase in the items to be used as the sorting reference, the search result can be acquired in a short time.

According to a sixth aspect of the present invention, based on the terminal information, a user rank value indicating a user's importance with respect to a user's specific item and / or user's search ability using the output terminal is calculated. Is,
The element of the square matrix and the transition probability matrix and / or at least one of the calculated square matrix or the eigenvector of the transition probability matrix is changed based on the user rank value. The information search system according to any one of 2 to 5.

  In the sixth aspect of the present invention, it is possible to display a Web page with a high content that is viewed by a user who has a high knowledge of a specific item or a user who has a high search capability. This makes it possible for the user to easily acquire information that is highly necessary when performing a search.

  The seventh aspect of the present invention has an information acquisition function capable of acquiring information related to the contents of a Web page on a network, and the relation rank indicating the importance of the contents of the Web page based on the information acquired by the information acquisition function A value is calculated, and at least one of the square matrix and the transition probability matrix and / or the calculated square matrix or the eigenvector of the transition probability matrix is changed based on the relation rank value. The information search system according to claim 2, wherein the information search system is an information search system.

  In the seventh aspect of the present invention, it is possible to display a Web page having contents highly relevant to the search condition at the top. This makes it possible for the user to easily acquire information that is highly necessary when performing a search.

An eighth aspect of the present invention is an information search system for searching a database based on a keyword input at a terminal, and outputting and displaying the search result on the terminal. An information ordering unit that has information acquisition means for acquiring terminal information that is information related to a terminal that outputs a Web page to be searched before performing search, and ranks the Web pages to be searched The information ordering means identifies a Web page that satisfies a predetermined condition based on at least terminal information and an output terminal that is outputting or outputting the Web page, and the Web page A square matrix is set based on the output terminal, and the square matrix is formed from row elements and column elements corresponding to the Web page and the output terminal. Each element P (i, j) satisfies the relationship of the following expression (1), calculates the eigenvector of the set square matrix, and searches based on the magnitude of the calculated eigenvector component It is an information search system characterized by ordering target Web pages.
P (i, j) = αX / Y (1)
α = a number other than 0
X = number of links from the web page corresponding to column “j” to the web page corresponding to row “i”, or at the output terminal corresponding to row “i” of the web page corresponding to column “j” Number of displays, or number of web pages displayed corresponding to row “i” on the output terminal corresponding to column “j”, or 0
Y = sum of the number of Web pages displayed on the output terminal corresponding to the column “j” and the number of links formed on the Web page corresponding to the column “j”

  According to a ninth aspect of the present invention, there is provided an information search method for searching a Web page based on a keyword input at a terminal that has executed a search, and outputting and displaying the search result on the terminal. Performing a calculation for estimating the transition tendency from the relationship between the Web page linked to the Web page to be searched and the terminal currently connected to the Web page to be searched or within a certain period of time. This is a characteristic information search method.

  According to a tenth aspect of the present invention, terminal information, which is information related to a terminal outputting a Web page to be searched, is acquired at the time of search execution or before search execution, and at least a predetermined condition based on the terminal information And the output terminal that is the terminal that has output or has output the Web page, the output terminal is a virtual Web page, and the Web page that is output from the output terminal and the virtual When a mutual link is formed with the Web page, the display from the link source Web page to the link destination Web page is displayed based on the link structure formed by the Web page and the virtual Web page. Set a transition probability matrix with switching as a transition, calculate the transition probability based on the set transition probability matrix, and search based on the calculation result Information search method according to claim 9, characterized in that ranks the eb page.

  In the ninth aspect and the tenth aspect of the present invention, there are Web pages that have popular information at the time of search execution, users who have similar interest in things, users who are familiar with the field of search purposes, etc. Since the web page to be browsed is displayed at the top of the search result, it is easy to acquire highly necessary information or high quality information. Moreover, such a search result can be output in a short time.

  The present invention provides an information search system for searching a Web page based on a keyword input at a terminal that has executed a search, and outputting and displaying the search result on the terminal. There is an effect that the search result reflecting the accuracy of the description content can be output more quickly.

It is a conceptual diagram which shows the structure of the information search system which concerns on the 1st Embodiment of this invention. It is a figure which shows an example of the functional block of the information search system which concerns on the 1st Embodiment of this invention. It is a figure which shows an example of the hardware constitutions of the information search system which concerns on the 1st Embodiment of this invention. It is a flowchart which shows an example of the search procedure by the information search system which concerns on the 1st Embodiment of this invention. It is explanatory drawing which shows an example of the link structure of the web page used as a search object, and the browsing condition of the said web page when it searches with the information search system which concerns on the 1st Embodiment of this invention. It is explanatory drawing which shows an example of the matrix for rank value calculation set when implementing a search with the information search system which concerns on the 1st Embodiment of this invention. It is explanatory drawing which shows an example of the matrix for rank value calculation set when implementing a search with the information search system which concerns on the 2nd Embodiment of this invention. It is explanatory drawing which shows the matrix used in order to verify the speed of eigenvalue calculation in the Example of this invention, (a) shows a Laplacian matrix, (b) shows a transition probability matrix. In the Example of this invention, it is a graph which shows the processing time ratio when changing the number of processors in the eigenvalue calculation of a Laplacian matrix. In the Example of this invention, it is a graph which shows the processing time in the eigenvalue calculation of a Laplacian matrix. In the Example of this invention, it is a graph which shows the processing time ratio when changing the number of processors in the eigenvalue calculation of a transition probability matrix. In the Example of this invention, it is a graph which shows the parallelization efficiency in the eigenvalue calculation of a transition probability matrix. In the Example of this invention, it is a graph which shows the influence of the network performance in the eigenvalue calculation of the Laplacian matrix in multiple nodes.

  An information search system and an information search method providing method according to an embodiment of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a conceptual diagram showing an outline of a configuration of an information search system 1 according to the first embodiment. As shown in FIG. 1, the information retrieval system 1 is composed of a retrieval device 2 and a client terminal 4 (indicated by reference numerals 4a to 4d and... In FIG. 1), which are connected via a telecommunication network. ing. On the client terminal 4, browsers 5 (indicated by reference numerals 5a to 5d and... In FIG. 1) are activated.

  The information search system 1 of the first embodiment can acquire input information. Specifically, this input information is acquired by the terminal-side information acquisition device 7 incorporated in the browser 5 by a predetermined means such as a plug-in format and transmitted to the search device 2. Note that the input information is information related to a search keyword or the like input by the user via the browser 5 or a program (for example, a plug-in toolbar) linked to the browser 5.

  Further, the terminal-side information acquisition device 7 can transmit not only the input information described above but also usage information that is information related to the output of the Web page via the browser 5 to the search device 2. Specifically, the terminal-side information acquisition device 7 may transmit, as usage information, information related to the Web page display time, such as the specified Web page display start time and display end time in the browser 5, to the search device 2. it can. Further, as usage information, when a link is traced from a specified Web page and another Web page is displayed, information indicating the transition path of the Web page such as the URL of the Web page of the movement source and the URL of the Web page of the movement destination Can be transmitted to the search device 2.

  Thus, the terminal-side information acquisition device 7 transmits the input information and usage information as described above to the search device 2 as terminal information. At this time, the transmission of information by the terminal-side information acquisition device 7 is performed for a specified time after the transmission of the previous information is completed, when a predetermined operation such as a search is executed, such as when the browser 5 is activated. This is automatically performed at a predetermined timing such as when elapses.

  On the other hand, the search device 2 stores the terminal information transmitted from the terminal-side information acquisition device 7 in the search database 3. That is, the terminal information transmitted from the client terminals 4 (4a to 4d,...) On the network to the search device 2 is stored in the search database 3 of the search device 2 (FIG. 2).

  Further, the search device 2 has an information collection unit 12. The information collecting unit 12 has a function of a so-called crawler and indexer. The information collecting unit 12 circulates on the WWW, acquires information about resources, and collectively stores the information in the search database 3. When storing in the search database 3, it is easy to search by analyzing, selecting, and classifying the information about the link structure of the Web page and the information about the contents such as the text, title, and various tags described for each feature. Keep it like that.

  When a search is executed in any of the client terminals 4, the information search system 1 creates a search query from the input search keyword and transmits it to the search device 2 that is a server.

  When the search device 2 receives the search query, the search execution unit 13 of the search device 2 specifies the Web page to be searched from the Web pages stored in the search database 3. In addition, the number of users browsing or browsing each identified web page is identified. As will be described in detail later, eigenvalues of the rank value calculation matrix formed based on these pieces of information are calculated, and the calculated eigenvalues are used as rank values (importance values) of each Web page to be searched. .

Further, the search execution unit 13 rearranges the search result Web pages based on the acquired rank value. Then, the search result output unit 15 of the search device 2 transmits the search results to the client terminal 4 on which the search has been executed, and the search result Web pages are displayed in the order in which the search results are sorted.
This makes it possible to display a Web page that is popular and topical at the time of executing the search at the top of the search results.

  Hereinafter, the information search system 1 of the present embodiment will be described in more detail.

First, a functional configuration of the information search system 1 of the present embodiment will be described.
Note that each functional block of the information search system 1 of the present embodiment described below can be realized as hardware, software, or both hardware and software. Specifically, if a computer is used, a CPU, a main memory, a bus, a nonvolatile memory, a so-called external storage device (hard disk or floppy disk (registered trademark), MO, CD, DVD, BD, magnetic) To control storage devices such as tape and reading devices for these media), printing equipment, display devices, other external peripheral devices, etc., I / O ports for the external peripheral devices, and hardware Driver programs, other application programs, and user interfaces used for information input.

  Then, the arithmetic processing of the CPU according to the program developed on the main memory, the data input through the interface and stored in various memories and hard disks, etc. are processed and accumulated, and the above hardware and software are installed. An instruction for control is generated. Further, the present invention can be realized not only as a system but also as a method. In addition, a part of the present invention can be configured as software. A software product used for causing a computer to execute such software and a recording medium in which the product is fixed to a recording medium are naturally included in the technical scope of the present invention (the same applies throughout the present specification). Is).

  FIG. 2 is a diagram illustrating an example of functional blocks in the information search system 1 of the present embodiment. As shown in this figure, the search system 1 of the present embodiment is mainly composed of a terminal-side information acquisition device 7 and a search device 2.

  The terminal-side information acquisition device 7 has at least a terminal information acquisition unit 10 and a terminal information transmission unit 11.

The terminal information acquisition unit 10 has a function of acquiring terminal information including at least a browsing URL from the browser of the client terminal 4. Here, the browsing URL is information indicating the location on the Web (on the network) of the Web page displayed by the browser (which the user has browsed on the client terminal 4). For example, “http: // /.../ABC.jpg ".
Here, the means for acquiring the terminal information from the browser is not particularly limited, and an appropriate method may be used. Examples of these methods include API (Application Program Interface), DDE (Dynamic Data Exchange), OLE (Object Linking and Embedding), and a file system.

In addition to the browsing URL, the terminal information acquired by the terminal information acquiring unit 10 may include various types of information related to browsing a web page on a browser. These information include, for example, the system user's user ID, browser ID, Web page browsing start time, Web page browsing end time, or “movement” when browsing two or more Web pages by following links. For example, migration information such as “original web page” and “destination web page” may be used.
Further, the terminal information may include input information such as a search keyword input by the user on a specific Web page.

The user ID indicates information for identifying the user, and includes, for example, an ID when logging in to the client terminal 4 and an ID assigned in units of services or software. Similarly, the browser ID is information for identifying the browser. For example, such information can be acquired by using a product ID recorded in a browser program or a toolbar program in advance.
Since the client terminal 4 can be identified by using these identification information, the browsing information can be managed and analyzed in units of the client terminal 4 (user). The ID information is held by an appropriate method such as a method held by the terminal-side information acquisition device 7 itself or a method using a cookie function of the browser.

  The browsing start time refers to the time when a web page is accessed on the web by the browser, the time when the web page is acquired by the browser by the access, the time when the web page is displayed, or the like. Specifically, the time is, for example, the time when the loading completion of the resource requested by the browser is detected, the HTTP status code of the Web server (for example, a code described as “200: OK”), etc. For example, the time when the signal is responded.

  The browsing end time refers to the time at which the display of the Web page is ended in the browser. For example, the time when the browser UI window is closed, the time when the window becomes inactive, and the mouse pointer is detected in the window area. The time when it disappears (the time when the mouse pointer moves out of the window area), and the like. In addition, since there is no operation for a certain period of time, when the OS of the client terminal 4 automatically performs a logoff process, or when a screen saver or the like is activated, the time when these operations are performed may be set as the browsing end time.

  The migration information is information indicating the “movement source web page” and the “movement destination web page” when two or more web pages are browsed continuously. For example, the migration information is a transition (transition) by clicking a hyperlink. Other examples include migration information in which two or more URLs successively input in the URL input field of a browser or a toolbar are “move destination” and “move source”.

The terminal information transmission unit 11 has a function of transmitting the acquired terminal information to a predetermined search device 2. Specifically, it has a function of automatically transmitting the terminal information to a predetermined search device 2 (server device) via a communication line or the like using the communication I / F of the client terminal.
Note that, for example, a method of specifying the transmission destination address of the predetermined search device 2 as the transmission destination of the terminal information by being described in advance in the program of the terminal side information acquisition device 7 may be used.
The timing at which the terminal information is transmitted from the terminal information transmission unit 11 can be set to an appropriate timing. For example, the timing of sending each time the Web page displayed on the browser is switched, the timing of sending in batch processing when a certain time or time arrives, or the timing of sending when the toolbar is started or ended Etc.

  In this way, terminal information for the output (display) of the Web page via the browser 5 at each client terminal 4 is acquired by the terminal-side information acquisition device 7 and transmitted to the search device 2. Then, by collecting the terminal information by the search device 2, it is possible to specify the Web page that is being browsed or browsed by the client terminal 4. In other words, the browsing history of each client terminal 4 can be easily grasped.

  Next, the functional configuration of the search device 2 will be described.

  The search device 2 is a server device, and includes a search database 3, an information collection unit 12, a search execution unit 13, and a search result output unit 15.

  The search database 3 is a database having a function of receiving, managing, and storing terminal information transmitted from the client terminal 4.

  As described above, the information collection unit 12 can circulate on the WWW, acquire information on the structure and content of the Web page, and create an index file based on the acquired content. The created index file can be stored and saved in the search database 3.

  When a search is executed in any one of the client terminals 4, the search execution unit 13 executes a search by referring to information stored in the search database 3 based on a search keyword from the user. Specifically, by referring to the search database 3, a Web page to be searched is specified, and in addition, the number of terminals outputting each specified Web page is specified. As will be described in detail later, eigenvalues of the rank value calculation matrix formed based on these pieces of information are calculated, and the calculated eigenvalues are used as rank values (importance values) of each Web page to be searched. . Still further, the Web sorting that is the search result is executed using the acquired rank value.

  The search result output unit 15 has a function of displaying the search results extracted from the search database 3 and rearranged on the browser 5 of the client terminal 4. Specifically, the transmission destination address of the client terminal is acquired by an appropriate method. Here, as an appropriate method, there are a method of receiving simultaneously with a search query, a method of collating a user ID with a pre-registered IP address, and the like. Then, a search result screen is displayed on the browser 5 of the identified client terminal 4 via the terminal-side information acquisition device 7.

Next, a hardware configuration of the information search system 1 in the present embodiment will be described.
FIG. 3 is a schematic diagram showing an example of a configuration in the information search system 1 when the above functional components are realized as hardware. The operation of each hardware configuration in the search system 1 of the present embodiment will be described using this figure. As shown in FIG. 3, the hardware configuration of the information search system 1 according to the present embodiment includes a terminal-side information acquisition device 7 incorporated in the client terminal 4 and a search device 2 arranged on the network as a server device. Composed.

  The terminal-side information acquisition device 7 and the search device 2 are connected to each other via an electric communication line to transmit and receive information. The telecommunication line includes the Internet.

  The terminal-side information acquisition device 7 includes a CPU (Central Processing Unit) and a RAM that realize the terminal information acquisition unit 10 and performs various other arithmetic processes, and a communication I / F that is a terminal information transmission unit 11. Also, a UI (User Interface) that is an input device such as a keyboard and a mouse, a VRAM for displaying resources processed by the browser program, and a display device such as a display are provided. Then, they are connected to each other by a data communication path such as a system bus, and perform transmission / reception and processing of information.

  A browser program and a plug-in program are stored on the RAM of the terminal-side information acquisition device 7, and various arithmetic processes of the CPU related to acquisition of terminal information, transmission processing, and other processes are executed in accordance with these programs. . The plug-in program secures an area for storing terminal information such as a browsing URL and a user ID at a predetermined address on the RAM.

  On the other hand, the search device 2 includes a CPU that performs various arithmetic processes, a RAM, a communication I / F, and an external storage device such as a hard disk for storing a large amount of terminal information. Then, they are connected to each other by a data communication path such as a system bus, and perform transmission / reception and processing of information.

  The RAM of the search device 2 stores programs for realizing the functions of the information search system 1 such as a program for realizing the functions of the search execution unit and the search result output unit and a database management program (not shown). In accordance with the program, various arithmetic processes of the CPU related to the terminal information management process and other processes are executed.

  Here, the terminal-side information acquisition device 7 receives an operation input via the user's UI, and the following processing is executed according to the browser program and the plug-in program on the RAM.

  For example, when an instruction to access a Web page is input by a browser operation via the UI, the URL of the Web page specified by the access instruction is stored at a predetermined address in the RAM. Then, an HTTP request is transmitted from the client terminal 4 to the search device 2 for the URL specified according to the browser program. The client terminal 4 receives the response to the HTTP request as the content of the access resource through the communication I / F. Subsequently, an HTML file, an image file, and the like included in the content received by the client terminal 4 are subjected to a rendering process by a CPU arithmetic process according to an HTML rendering engine or the like included in the browser 5. The processing result is transferred to the VRAM, and the Web content is displayed on the display device. Then, the user browses resources such as the displayed Web page on the WWW.

  At the same time, the plug-in program executes the following process in response to the access instruction input from the browser. That is, the URL stored at a predetermined address in the RAM is automatically transmitted from the communication I / F to the search device 2 at the predetermined timings as described above as terminal information.

  On the other hand, the search device 2 receives the terminal information transmitted from the terminal-side information acquisition device 7 of each client terminal 4 through the communication I / F and stores it in a predetermined address of the RAM. Then, according to the program of the search device 2, for example, the URL of the terminal information or the user ID is used as a parameter to create a file of the search database 3 by the arithmetic processing of the CPU or For example, it is added to a file and stored in an external storage device such as an HDD.

  In addition, when a search engine (Web page for search) is accessed by a browser operation using a UI and a search query is input and transmitted, or when a search query is input and transmitted to a toolbar, the client terminal A case where the search is executed at 4 will be described. At that time, the search device 2 receives an HTTP request including the search query through the communication I / F and stores it at a predetermined address in the RAM. Then, the search device 2 executes the following processing according to the search server program. In other words, Web pages that satisfy the search query condition are extracted from the index file database stored in the external storage device using the search query stored in the RAM as a key. Subsequently, by comparing and calculating the extracted Web page information and the terminal information stored in the external storage device (terminal information transmitted from the terminal-side information acquisition device 7) by the CPU comparison calculation process, The terminal that outputs the extracted Web page is specified. As will be described in detail later, a rank value calculation matrix is set from the link structure of the extracted Web page and the information of the terminal from which the extracted Web page is output. Furthermore, the rank value of each extracted Web page is obtained by calculating the eigenvalue of the set rank value calculation matrix based on a prescribed algorithm, and the extracted Web page is obtained by the comparison operation processing of the CPU. Are sorted in descending order of rank. In addition, the search device 2 generates a search result in which the Web pages are arranged in the form of an item by calculation processing of the CPU, and transmits the search result to the browser 5 of the client terminal 4 that is the transmission source of the search query by the communication I / F. That is how it is said.

Next, the flow of processing at the time of search execution of the information search system 1 of the present embodiment will be described.
FIG. 4 is a flowchart showing an example of the flow of processing at the time of executing the search of the information search system 1 of the present embodiment. Note that the steps shown below may be processing steps that are recorded on a medium and constitute a program for controlling a computer.

  When a search is executed in any one of the client terminals 4 (Yes in Step 101) and a search query is transmitted from the client terminal 4 to the search device 2, the search device 2 that has received the search query receives a search query from the search database 3. Web pages that satisfy the search query conditions are extracted (step 102). That is, the corresponding index file is searched using the search query as a keyword.

  Further, the search device 2 compares the extracted web page data with the terminal information stored in the search database 3 to identify the terminal that outputs the extracted web page (step 103). That is, by referring to the information relating to the output of the Web page transmitted from each client terminal 4 via the browser 5 or the past output, the terminal that has output the extracted Web page or the terminal that has output the Identify.

  Next, the search device 2 includes information related to the Web page that satisfies the search query condition (information acquired in Step 102) and information related to the terminal that outputs the extracted Web page (information acquired in Step 103). A rank value calculation matrix is set from (step 104). The rank value calculation matrix setting process, which is a characteristic configuration of the present invention, will be described in detail below with reference to FIGS.

The rank value calculation matrix is set corresponding to a link structure made up of a plurality of Web pages published on the network and users, and each “column” is changed from the Web page corresponding to the “column”. This indicates a link to the Web page. That is, each element belonging to the “column” indicates a link to another Web page. Furthermore, each “line” indicates a link from another Web page to a Web page corresponding to the “line”. That is, each element belonging to “row” indicates a link from another web page to a web page corresponding to “row”.
More specifically, for example, in the case of a rank value calculation matrix P as shown in FIG. 6, each element of the matrix P includes the Web pages A, B, C, D and the user shown at the top of the matrix. (Web page viewer) From E, F, G, H, Web pages A, B, C, D displayed on the left part of the matrix and links to users E, F, G, H, or Web page browsing by users It is a numerical value indicating the situation of
That is, the square matrix (rank value calculation matrix P) shown in FIG. 6 is for the Web page to which the Web page to be searched is linked and the Web page to which the Web page to be searched is linked. It shows the relationship with the terminal that is actually connected within a certain period.

Here, the rank value calculation matrix will be specifically described using a model. Of course, the rank value calculation matrix of the present invention is not limited to the model.
As shown in FIG. 5, four Web pages A, B, C, and D are held on a Web server and can be browsed via a browser on a computer connected to a network such as the Internet or a LAN. And That is, it is assumed that four Web pages A, B, C, and D are open on the network. Here, the first Web page A is in a state of being linked to Web page B and Web page C.
Here, the “linked state” means that the URL of another Web page that is the link destination is associated with the specific Web page that is the link source. For example, a predetermined part on the Web page that is the link source By clicking a predetermined operation such as clicking, the linked Web page is displayed on the browser.

  Further, the second Web page B is in a state where a link is made to the Web page D. The third Web page C is in a state of being linked to the Web page D. The fourth Web page D is in a state where links are made to Web page A and Web page C, respectively.

In other words, the first Web page A is linked to the fourth Web page D.
Here, the “linked state” refers to a state in which another Web page that is a link source links to a specific Web page that is a link destination. That is, a URL of a specific Web page is associated with another Web page, and a specific operation is performed on the other Web page, so that the specific Web page is displayed on the browser. .

  Similarly, the second web page B is linked to the web page A. The third Web page C is in a state in which links are made to Web page A and Web page D, respectively. The fourth Web page D is in a state in which links are made to Web page B and Web page C, respectively.

  Here, it is assumed that the user E is browsing the Web page A. That is, the browser 5 of the client terminal 4a used by the viewer E can acquire data from a Web server holding the Web page A via a network such as the Internet or a LAN, and can browse the Web page A. Assume that the status is output.

  Similarly, it is assumed that the users F, G, and H are browsing the web page B, respectively. That is, it is assumed that the browsers 5 of the client terminals 4b, 4c, and 4d output the Web page B in a state where it can be browsed.

  In the situation described above, it is assumed that a search is executed in any one of the client terminals 4 and the Web pages A, B, C, and D are all Web pages that satisfy the search query condition. At this time, the search device 2 identifies the client terminals 4a, 4b, 4c, and 4d that output the Web pages A, B, C, and D, and the users E, F, G, and H specify the Web pages A, B, Assuming that C and D are viewed, a rank value calculation matrix P as shown in FIG. 6 is set.

  When attention is paid to each “column” element of the rank value calculation matrix P, the left part RL is an element relating to the Web pages A, B, C, and D, and the right part RR is an element relating to the users E, F, G, and H. It has become. That is, an element group composed of elements related to a plurality of Web pages A, B, C, D and an element group composed of elements related to a plurality of users E, F, G, H are arranged next to each other.

  Thus, the search device 2 of this embodiment sets the rank value calculation matrix P with the users E, F, G, and H as virtual Web pages E, F, G, and H. That is, “browsing a specific web page by a specific user” is reflected in the rank value calculation matrix P as a state where “a virtual web page and a specific web page are linked to each other”. In other words, in the square matrix shown in FIG. 6, the Web pages A, B, C, and D and the viewers E, F, G, and H are treated equally and are reflected in each element of the matrix.

  More specifically, it is assumed that the user E is browsing the web page A as described above (see FIG. 5). At this time, the search device 2 assumes that there is a virtual web page E that does not actually exist in the link structure composed of the web pages A, B, C, and D, and from the web pages A, B, C, D, and E. The link structure is as follows. Then, in addition to “rows” and “columns” corresponding to the Web pages A, B, C, and D, new “rows” and “columns” corresponding to the virtual Web page E are added to calculate rank values. Set the matrix P. In other words, new “rows” and “columns” are added to the rank value calculation matrix P on the condition that the web page A is output from the client terminal 4a used by the user E.

  Therefore, when the four users E, F, G, and H are browsing any of the Web pages A, B, C, and D to be searched as described above, the Web pages A, B, C , D, E, F, G, H, and rank value by adding four new “rows” and “columns” corresponding to virtual web pages E, F, G, H. A calculation matrix P is set. That is, the sum of the number of Web pages to be searched and the number of terminals outputting the Web pages to be searched is the number of rows and columns of the rank value calculation matrix P.

Here, each “column” of the rank value calculation matrix P corresponds to the above-described Web page to be searched or a virtual Web page defined on the condition that the Web page to be searched is output. . Each element belonging to the “column” indicates a link from the corresponding Web page or virtual Web page (hereinafter also simply referred to as a Web page) to another page. Specifically, the element P (i, j) belonging to the “column” is expressed by the following formula (1):
P (i, j) = αX / Y (1)
α = a number other than 0
X = number of links from the web page corresponding to column “j” to the web page corresponding to row “i”, or at the output terminal corresponding to row “i” of the web page corresponding to column “j” Number of displays, or number of web pages displayed corresponding to row “i” on the output terminal corresponding to column “j”, or 0
Y = sum of the number of Web pages displayed on the output terminal corresponding to the column “j” and the number of links formed on the Web page corresponding to the column “j”

It satisfies the relationship.
That is, when the value based on the number of link destinations to which the Web page is linked is A, and the number of Web pages that the terminal is currently connected to within a certain period is B, each element of the rank value calculation matrix P is , A / (A + B), B / (A + B) or a value obtained by multiplying these by the system number α.

  Here, as described above, the search device 2 of the present embodiment links the Web page A to the virtual Web page E on condition that the Web page A is output from the client terminal 4a used by the user E. And rank value calculation matrix P is set assuming that a link is established from virtual web page E to web page A. In other words, the browsing of the web page A by the user E is reflected in the rank value calculation matrix P as a mutual link between the web page E and the web page A that does not actually exist (a state in which they are linked to each other).

  That is, as shown in FIG. 5, it is assumed that a specific Web page A is linked to the other two Web pages B and C, and the user E is browsing the specific Web page A. At this time, it is assumed that the search device 2 links a specific Web page A to a virtual Web page E in addition to the two Web pages B and C. At this time, as shown in FIG. 6, when attention is paid to each element P (i, 1) forming a column (leftmost column in FIG. 6) corresponding to a specific Web page A of the rank value calculation matrix P. , P (2,1), P (3,1), and P (5,1) that are elements corresponding to links to the web pages B and C and the virtual web page E are all 1/3. It has become. These indicate that the total number of links of the specific Web page A is 3, and that one link is provided from the specific Web page A to each of the Web pages B and C and the virtual Web page E. Yes. Therefore, the sum of the elements of each “column” of the rank value calculation matrix P is 1.

  When attention is paid to each “row” element of the rank value calculation matrix P, the upper part LH is an element related to the Web pages A, B, C, and D, and the lower part LL is the user E, F, It is an element related to G and H. That is, an element group composed of elements related to a plurality of Web pages A, B, C, and D and an element group composed of elements related to a plurality of users E, F, G, H are continuously arranged in the vertical direction. ing.

From the above, as shown in FIG. 6, the rank value calculation matrix P includes the Web pages A, B, C, and D to be searched and the virtual Web pages E, F, G, and H. In the Web page group, each Web page A, B, C, D, E, F, G, H is a vertex, and tracing the link is a transition (transition), and each link in the Web page Therefore, it can be said that the transition probability matrix (transition probability matrix) is assumed to follow one link with a uniform probability.
Here, “follow a link” means that the URL of another Web page that is the link destination is associated with the specific Web page that is the link source. For example, a predetermined portion on the Web page that is the link source It is assumed that the linked Web page is displayed on the browser by a predetermined operation such as clicking. Further, a link defined between an actual web page and a virtual web page is handled in the same manner as a link between actual web pages.

  At this time, paying attention to the rank value calculation matrix P, as shown in FIG. 6, each element belonging to the area E1 located in the upper left is formed between the Web pages A, B, C, and D that actually exist. Shows the links to be. Each element belonging to the area E2 located in the upper right indicates a link from any one of the virtual Web pages E, F, G, and H to the actually existing Web pages A, B, C, and D. ing. Furthermore, each element belonging to the area E3 located at the lower left has a link from any of the existing Web pages A, B, C, and D to any of the virtual Web pages E, F, G, and H. Show. An area E4 located in the lower right indicates a virtual link formed between the virtual Web pages E, F, G, and H.

  When such a rank value calculation matrix P is set, as shown in FIG. 4, the process proceeds from step 104 to step 105, and the search device 2 calculates an eigenvalue of the rank value calculation matrix P. That is, the transition probabilities are calculated for each of the Web pages A, B, C, and D and the virtual Web pages E, F, G, and H that are the vertices of the transition probability matrix. Then, the search device 2 rearranges the Web pages to be searched based on the calculated eigenvalues of the rank value calculation matrix P (step 106).

More specifically, for example, in the rank value calculation matrix P shown in FIG. 6, the eigenvalue is expressed by the following equation (2):
[A, B, C, D, E, F, G, H] = [0.408, 0.544, 0.408, 0.544, 0.136, 0.136, 0.136, 0.136 ] ... (2)
Meet the relationship.
Therefore, the eigenvalues of Web pages A, B, C, and D shown in FIG. 5 are 0.408, 0.544, 0.408, and 0.544, respectively. Here, if the Web pages A, B, C, and D are rearranged in descending order of eigenvalues, the Web page B, Web page D, Web page A, and Web page C are arranged in this order.

The search device 2 transmits the sorted Web pages to be searched to the client terminal 4 as a search result (step 107).
This is the end of the description of the flow of processing when the information retrieval system 1 performs the retrieval.

  In the above-described embodiment, when setting the rank value calculation matrix P, the virtual web pages E, F, G, and H are set with the users E, F, G, and H as uniform users. That is, the number of virtual links defined between the virtual web page and the actual web page is the same regardless of which user E, F, G, or H is the user who viewed the web page. . In other words, any user has a configuration in which one mutual link is defined between one virtual web page and one real web page.

  However, the information search system of the present invention is not limited to such a configuration. That is, when the depth of knowledge for a specific item, search skill, etc. are different among the users E, F, G, H, the users E, F, G, It may be configured to reflect differences in knowledge and skills of H. Furthermore, the configuration may be such that the contents of the Web pages A, B, C, and D such as the number of appearances of the search keyword are reflected in the rank value calculation matrix P.

  Hereinafter, the second embodiment and the third embodiment will be described. Unless otherwise specified, the same parts as those of the information search system 1 in the first embodiment are denoted by the same reference numerals and overlapped. Description is omitted.

  In the second embodiment, when the rank value calculation matrix is set, it depends on the difference in knowledge and search skills of the users E, F, G, and H and the difference in the contents of the Web pages A, B, C, and D. Thus, each element of the rank value calculation matrix is changed.

  As in the first embodiment, the information search system 20 according to the second embodiment includes the search device 2, each client terminal 4, and the browser 5 and the terminal-side information acquisition device 7 that operate on each client terminal 4. (FIG. 1).

  The terminal information acquisition device 7 of the second embodiment calculates the user rank value of each user by using the acquired terminal information. In this specification, the “user rank value” indicates a value obtained by quantifying the importance of the user, which is determined based on the depth of knowledge of each user regarding a specific item, search skill, and the like. The calculated user rank value is stored as terminal information in the search database 3.

The user rank value is calculated by, for example, the user via the browser 5 based on the terminal information acquired by the terminal information acquisition device 7 and the information on the Web page collected by the information collection function (information collection unit 12). The web page may be calculated based on the type of web page output, the browsing time of the web page, the transition history when the link is followed, and the like.
There is also a method for acquiring personal information such as a user's occupation and hobbies when the terminal information acquisition device 7 is incorporated in a browser or when software is installed. Further, there is a method of acquiring information on shortcuts to frequently viewed web pages recorded in the browser 5. That is, when a user browses a Web page to be searched when the search is executed and the user's occupation or hobby is related to the content of the Web page, the user rank value of the user It is a method of raising the.
Furthermore, there is a method of acquiring data from a BBS (specific Web page) as necessary when a user answers a question related to a specific item using a BBS or the like. Specifically, information on evaluation of other users with respect to answers made by the user is acquired as necessary. And it is the method of raising the user rank value of the user who made the reply which got high evaluation from the other user.
Further, by analyzing the search keyword, the web page browsed after the search, the time when the web page was browsed after the search, the page linked from the web page after the search, etc., the user's search skill is determined, and based on it. It is conceivable to assign a user rank value to.
From such methods of evaluating users based on the depth of knowledge of search contents, search skills, etc., at least one appropriate one is selected, and the terminal information acquisition device 7 and the information collection unit 12 collect information. The user rank value is determined by executing an analysis process in the search device 2.

  In the terminal information acquisition device 7 of the second embodiment, the relation rank value of each Web page is calculated based on the information collected by the information collection unit 12. In this specification, the “relation rank value” is based on information such as how many specific words are described in the contents of each Web page and whether specific words are described in a predetermined location such as a title. A value obtained by quantifying the degree of relevance between the content of the Web page and the specific item determined in the above. The calculated relation rank value is stored in the search database 3 as terminal information.

Next, the flow of processing at the time of search execution of the information search system 1 of the present embodiment will be described.
When a search is executed in any one of the client terminals 4, the search device 2 sets a rank value calculation matrix in the same procedure as in the first embodiment.

  Here, in the present embodiment, the relation rank values of the Web pages A, B, C, and D extracted by the search and the users E, F, G, and H who are browsing the extracted Web pages. The value of each element of the rank value calculation matrix is changed with reference to the rank value.

More specifically, it is assumed that the user F is a user with a higher user rank than the users G and H in the model shown in FIG. At this time, the element values of the rank value calculation matrix P shown in FIG. 6 are changed to the rank value calculation matrix P shown in FIG. That is, the values of P (4,2), P (7,2), and P (8,2) are changed from 1/4 to 1/5, and the value of P (6,2) is changed from 1/4. It has been changed to 2/5. Thus, the importance of the link from the web page B to the virtual web page F is made higher than the importance of each link from the web page B to each of the virtual web pages G and H.

  As described above, by appropriately changing the value of each element of the created rank value calculation matrix based on either or both of the user rank value and the relation rank value, the existing Web pages A, B, The importance of virtual web pages E, F, G, and H defined by C, D, and users E, F, G, and H is reflected in the rank value calculation matrix. At this time, even if the value of each element of the rank value calculation matrix is changed, the sum of the elements of each “column” of the rank value calculation matrix P is 1. Therefore, when the element P (6,2) value is changed from the rank value calculation matrix P shown in FIG. 6, the value is changed from 1/4 to (1 + α) / (4 + α) (α is a number other than 0). The

  Further, the eigenvalue of the rank value calculation matrix in which the value of each element is changed as necessary is calculated. That is, by calculating after changing the value of each element of the transition probability matrix, the Web pages A, B, C, D and the virtual Web pages E, F, G, H which are the vertices of the transition probability matrix The calculation which estimates the transition tendency based on the transition probability in each is implemented. Then, the extracted Web pages are rearranged based on the calculated eigenvalues.

  Further, the search device 2 transmits the sorted Web pages to be searched to the client terminal 4 as a search result. Above, description of the flow of the process at the time of search execution of the information search system 1 of this embodiment is completed.

Next, a third embodiment will be described.
The third embodiment is characterized in that after a rank value calculation matrix is set and eigenvalues are calculated, values based on user rank values and relation rank values are added to each eigenvalue or matrix element.
More specifically, a rank value calculation matrix is set and eigenvalues are calculated by the same procedure as in the first embodiment or the second embodiment. Then, the calculated eigenvalues are changed, and the web pages are rearranged with the changed eigenvalues.

  For example, it is assumed that the calculated eigenvalues of Web pages A, B, C, and D are 0.408, 0.544, 0.408, and 0.544, respectively. At this time, it is assumed that a user with a high user rank value is browsing Web page A. Then, the search device 2 adds a predetermined value to the unique value of the Web page A to obtain 0.608. Then, the Web pages A, B, C, and D are rearranged based on the eigenvalues 0.608, 0.544, 0.408, and 0.544 corresponding to the changed Web pages A, B, C, and D. . In this way, after changing the eigenvalues corresponding to the Web page being browsed by the user with the higher user rank value or the Web page with the higher relation rank value, the Web pages are rearranged. In the above-described example, an example is shown in which values are assigned to eigenvalues corresponding to Web pages viewed by users with a high user rank value or Web pages with a high relationship rank value. It is not limited. For example, the eigenvalue value corresponding to the Web page having a low relation rank value may be subtracted. The change of the eigenvalue may be appropriately changed according to the user rank value or the relation rank value.

  In each of the above-described embodiments, the rank value calculation matrix P is set in each Web page group including the Web pages A, B, C, and D to be searched and the virtual Web pages E, F, G, and H. It is a transition probability matrix having vertices of Web pages A, B, C, D, E, F, G, and H. Therefore, the rank value calculation matrix P is set by, for example, creating an adjacency matrix having apexes of the Web pages A, B, C, D, E, F, G, and H, based on the created adjacency matrix. A transition probability matrix may be derived. The derivation method for setting the rank value calculation matrix P may be changed as appropriate.

In each of the above-described embodiments, when a search is executed, a Web page that satisfies the search query condition is extracted, a user browsing the extracted Web page is specified, and the extracted Web page and the Web page are browsed. The matrix P for rank value calculation corresponding to the user who is present is set. However, the processing at the time of search execution of the search system of the present invention is not limited to this.
For example, a rank value calculation matrix P2 corresponding to all Web pages on the network and users browsing the Web page is derived in advance, and the rank value previously derived when the search is executed. The rank value calculation matrix P may be set by extracting necessary elements from the “row” and “column” elements from the calculation matrix P2. That is, by extracting and combining the necessary “row” and “column” elements from the previously derived rank value calculation matrix P2, a Web page that satisfies the search query condition and browsing it are browsed. The rank value calculation matrix P corresponding to the user who is present may be set.

  Similarly, a rank value calculation matrix P3 corresponding to a Web page satisfying a specific condition and a user viewing the Web page is derived in advance, and when a search is executed, the rank is calculated in advance. The rank value calculation matrix P may be set by extracting necessary elements from the “row” elements and “column” elements from the rank value calculation matrix P3.

  In each of the above-described embodiments, in the row element and the column element of the rank value calculation matrix, the element group corresponding to the Web pages A, B, C, and D and the element group corresponding to the users E, F, G, and H are The rank value calculation matrix formed adjacently has been described as an example, but the rank value calculation matrix of the present invention is not limited to this. For example, the order of the row elements and the column elements of the rank value calculation matrix may be appropriately switched. Further, for convenience of explanation, the rank value calculation matrix is represented in the matrix format, but the handling of the matrix when the search system actually calculates may be set as appropriate. In other words, each element forming the matrix or only elements necessary for the calculation among the elements forming the matrix may be stored and held, and the necessary calculation may be executed. It is not always necessary to deal with the matrix shape as described above.

  In each of the above-described embodiments, when the search is executed, the terminal information acquired in advance by the terminal-side information acquisition device 7 and the information related to the web page structure acquired in advance by the information collection unit 12 are retrieved from the search database 3. Although the rank value calculation matrix P is acquired and set, the method for setting the rank value calculation matrix is not limited to this. For example, when the search is executed, the search device 2 may acquire information related to browsing of the Web page from the client terminal 4. Further, the acquired information may be used for setting the rank value calculation matrix without being held in the database. That is, transmission of terminal information from the client terminal 4 to the search device 2 may be performed in a timely manner. Then, the storage of information in the database may be performed in a timely manner or may not be performed in some cases.

  In each of the embodiments described above, the Web pages are rearranged in descending order of eigenvalues. However, the present invention is not limited to this, and for example, the Web pages may be rearranged in ascending order of eigenvalues. However, according to the method of rearranging Web pages in descending order of eigenvalues, Web pages with high topicality (for example, Web pages output from a large number of client terminals 4) are displayed at the top when performing search.

  Further, in each of the above-described embodiments, when the Web pages that are the search results are rearranged, if the unique values of the Web pages are the same, pages with a large number of viewers are displayed at the top. However, the search system of the present invention is not limited to this. For example, when the eigenvalues of Web pages are the same, a page with fewer viewers may be displayed at the top.

  In each of the above-described embodiments, as illustrated in FIG. 3, a configuration in which the search device 2 includes a function of receiving and managing information transmitted from the search database 3 and the client terminal 4 is exemplified. The search system 1 of the invention is not limited to this. For example, the configuration may be such that the search database 3 is provided in a separate server device on the network, and information is acquired from the search database 3 of the server when the search device 2 executes a search. That is, it is only necessary to realize the functions in the entire system, and the number of servers to be installed, the functions (programs), and the arrangement of the databases can be appropriately changed depending on the capabilities of the servers.

  Note that the programs and systems of the above-described embodiments are downloaded to a computer through various recording media such as an optical disk such as a CD-ROM, a magnetic disk, and a semiconductor memory, or via a telecommunication network. Can be installed or loaded.

  By the way, the calculation for obtaining the eigenvalue of the rank value calculation matrix described above may be performed with an appropriate hardware configuration and an appropriate calculation algorithm. However, the following verification was performed in order to obtain eigenvalues of the rank value calculation matrix at higher speed. The information search system of the present invention may select a hardware configuration and an arithmetic algorithm with reference to the following verification.

  In performing the verification, classical methods such as the LR method and the QR method generally require a large amount of labor. Therefore, the calculation time is reduced by converting to a matrix having the same eigenvalue and a smaller size than the original matrix. We verified the method based on the projection method. A method based on the projection method such as Lanzcos, Arnoldi, Krylov-Schur, etc. is a method comprising an orthogonalization phase based on the Gram-Schmidt method, an eigenvalue calculation based on the multi-shift QR method, and a Schur decomposition phase. This is a method that is obtained from a set of desired eigenvalues and eigenvectors such as the maximum eigenvalue and minimum eigenvalue by rearranging the blocks. Therefore, in an application that requires only the eigenvector corresponding to the maximum eigenvalue, the calculation time may be significantly reduced. Therefore, the method based on the “projection method” is considered suitable for the calculation for obtaining the eigenvalue of the rank value calculation matrix.

  Among libraries that implement these projection methods, SLEPc corresponds to the calculation of a large-scale matrix. In SLEPc, the latest eigenvalue calculation method is implemented using existing numerical calculation libraries such as PETSc, BRAS, LAPACK, etc., and a significant reduction in calculation time can be expected. A numerical calculation library called PETSc used by SLEPc has a function of performing parallel processing of vectors and matrices in a plurality of processes using OpenMPI. PETSc is written in C language and can perform calculations at high speed. In addition, OpenMPI supports various methods such as shared memory, InfiniBand, TCP / IP as inter-process communication methods, and automatically switches internally so that optimal parallelization is possible according to the environment that can be prepared It can be easily realized.

  In contrast, a general search service always builds a system on the premise of processing in a large-scale distributed environment. These middlewares are characterized in that a large amount of batch processing can be stably executed in an environment where node failures or the like frequently occur. However, the overhead for redundancy is large, and high performance cannot be achieved in processing involving a large amount of communication.

  In other words, because there is a possibility that tightly coupled processing can be performed with higher performance in small and medium-sized computer environments, in this verification, we decided to investigate middleware for clusters implemented based on MPI.

  First, the performance of SLEPc implemented with OpenMPI was investigated. Specifically, it was investigated how much resources are required to achieve the scalability required by the information retrieval system of the present invention. More specifically, we investigated the performance assuming a case where a calculation of about 100,000 entries is performed for the Web page and the viewer.

The evaluation environment for verification is as follows.
Calculator: PowerEdge R410 × 2 (DELL)
CPU: Intel (R) Xeon (R) L5520 (2.26 GHz, 8 MB cache, 5.86 GT / s QPI) x 2 (4 cores)
Memory: 32GB (4GB x 8 / 2R / 1066MHz / DDR3 RDIMM)
Disk: RAID6 (PERC6i), 600GB, 15,000 RPM SAS
Network: Broadcom NetXtreme II BCM5716 1000Base-T PCI Express
Switch: Cisco Catalyst 2960G-8TC-L
OS: Ubuntu Linux 10.04
Library: PETSc 3.0.0, SLEPc 3.0.0, LAPACK 3.2.1, BRAS 1.2

  In the performance evaluation, eigenvalues and eigenvectors were calculated using a Laplacian matrix as an example of a symmetric matrix and a transition probability matrix generated assuming a model that randomly transitions between elements as an example of an asymmetric matrix. The Laplacian matrix used in this experiment is a matrix as shown in FIG. 8A, in which the diagonal elements are 2 and the elements on both sides are 1. Further, the transition probability matrix is a matrix as shown in FIG. 8B generated on the assumption that it moves randomly on the triangular lattice. In this verification, first, parallel processing performance in a single node is evaluated, and then evaluation is performed on a plurality of nodes using TCP / IP.

  Before the evaluation with the actual machine, the implementation of Arnoldi, Krylov-Schur was investigated from the source code of SLEPc.

  In PETSc used in SLEPc, the OpenMPI collective communication function is used to divide the matrix into specified ranges, such as from the first row to the n1 row and from the n1 + 1 row to the n2 row. Implements a function to generate and process matrices. In the orthogonalization calculation, it is necessary to exchange the intermediate calculation results of submatrix and vector distributed in each process with functions such as gather, scatter, reduce, etc., so the communication overhead during parallelization will increase. .

  On the other hand, in the multi-shift QR and Schur decomposition processes, processing can be performed by paying attention only to the submatrix that is focused on, and it is considered that no communication overhead occurs. The above points were confirmed by actually measuring performance.

  PETSc provides a function for analyzing the execution time of a program, the number of times of calling a preset location, the execution time, etc., with the option of -log summary. In the following experiment, performance was measured using the function of this -log summary.

  Furthermore, in collective communication such as all gather and all reduce in MPI, it is assumed that the number of processors is an even number, and in the case of an odd number, communication of the remaining nodes cannot be performed efficiently, so collective communication is frequently used. It is considered that an even number of processors should be used in the calculation. When actually using an odd number of n processors, only a performance equivalent to the number of even number n1 processors was obtained. Therefore, in the following evaluation, the performance of an even number of processors was evaluated.

  Since the code actually executed in parallel differs depending on the value of the matrix to be calculated, the parallelization rate cannot be obtained directly from the source code, but the parallelization rate with the specified parameters is calculated from the performance measurement results The library performance can be evaluated. The parallelization rate obtained below is the parallelization rate with the specified parameters. In SLEPc, iterative calculation is continued until the calculation result falls within the specified error by specifying the allowable error of the eigenvalue to be obtained. In the following experiment, the tolerance was calculated as 10 −7.

  First, we investigate the computation time in the Laplacian matrix.

  In the investigation, the eigenvalue calculation time of a Laplacian matrix having a size of 100,000 × 100,000 was measured. At this time, first, the processing times at the same number of iterations were compared. If the number of iterations is limited to about 12,500, the eigenvalues are not yet converged in any method of this size matrix, so that the same number of iterations is calculated. FIG. 9 shows the processing time ratio T1 = Tn when the number of processors is changed, where T1 is the processing time when calculated with a single processor and Tn is the processing time when calculated with the number of processors n. It was.

  As the hardware used, an 8-core processor can be used per node, and a 16-core processor can be virtually used by Intel's Hyper Threading technology. However, as shown in FIG. 9, the performance is degraded when the number of processors increases from 8 to 9. Therefore, when the Hyper Threading function was turned off and the processing time of each method was measured up to 8 processes, the same changes were observed up to 8 processors. For this reason, it is considered that the performance is degraded due to Hyper Threading rather than using multiple CPUs. It is possible that the implementation of Ubuntu is influential, but we have not investigated the detailed cause. In response to this, in the following experiments, evaluation was performed with the Hyper Threading function turned off.

  In the results up to eight processors, the increase in the processing time ratio has slowed as the number of processors increases as Amdahl's law shows. If the ratio (parallelization ratio: Parallel Portion) in the program is r (0 <= r <= 1) and the number of processors is n, the processing time of the parallel processing part ignoring the overhead is Since rT1 = n and the processing time of the sequential processing portion is (1-r) T1, the processing time ratio Speedup is as shown in the following [Equation 1].

  When r is estimated from the measured value by the least square method, Lanzcos, Arnoldi, and Krylov-Schur are approximately 0.9653 (96.53%), 0.9720 (97.20%) 0.9583 (95.83), respectively. %). The straight line in FIG. 9 represents ideal performance, and the curve represents an estimated curve. Since the estimation is performed by ignoring the overhead, there is a difference between the observed value and the estimated curve. Since the measured value with 8 processors is below the approximate curve, the parallelization rate is considered to decrease if the performance of 10 processors or more is measured and recalculated. Next, the parallelization efficiency (Parallel Efficiency) with 8 processors in each method was examined. Since the parallelization efficiency PE is a value obtained by dividing the processing time ratio by the number of processors, it is expressed by the following [Equation 2] when expressed as a percentage.

  The parallelization efficiency was approximately 74.58%, 76.11%, and 68.87% for Lanzcos, Arnoldi, and Krylov-Schurb, respectively, and Arnoldi showed the best values.

  For all three methods, the parallelization efficiency in the symmetric matrix is less than 50% if it is up to 8 processors, but if it exceeds that, the parallelization efficiency may decrease. For example, in the case of a large-scale system, there is a case where a job less than 50% parallelization efficiency is refused in order to maintain the system utilization efficiency. Therefore, in the future, it has been found that 8 nodes or more are required to realize the processing performance required by the information retrieval system of the present invention, and that it is found that the parallelization efficiency falls below 50% with 8 processors or more. In order to obtain the return on investment by adding hardware, it is possible to use other libraries and calculation methods.

  FIG. 10 shows a graph comparing the calculation times of 12,500 iterations of the three methods. It can be seen that the calculation time of Lanzcos is the shortest in the symmetric matrix. It can be seen that the remaining two methods that can also calculate the asymmetric matrix have a shorter calculation time for Krylov-Schur. However, these are the calculation times when the number of iterations is made equal, and actually, the number of iterations required until the eigenvalue converges in each method is different. For example, in this investigation, Arnoldi has converged after 36534 iterations.

  Next, the calculation time in the generated transition probability matrix was investigated.

  Since the generated matrix is an asymmetric matrix and is a transition probability matrix, the matrix is closer to the link structure of the target Web page than the Laplacian matrix. There is a constraint condition for generating the matrix so that the sum of the values in the column direction of the matrix becomes 1, and the size of the matrix becomes a halfway value. This time, a matrix with a size of 281625 was used. In addition, the number of repetitions was limited to 500 times in which eigenvalues did not converge in each method. Since Lanzcos is a calculation method for a symmetric matrix, two methods, Arnoldi and Krylov-Schur, were compared for an asymmetric matrix. In Krylov-Schur, when the matrix is symmetric, the calculation algorithm is partially switched to a calculation method based on Lanzcos.

  FIG. 11 shows the result of the processing time ratio T1 = Tn when the number of processors is changed. The parallelization rates estimated by the least square method are 0.817591 (81.75%) and 0.839298 (83.92%) for Arnoldi and Krylov-Schur, respectively, and Arnoldi and Krylov-Schur for the asymmetric matrix In both cases, the parallelization rate is lower than in the case of the symmetric matrix. Similarly, FIG. 12 shows the result of parallel efficiency when the number of processors is changed. Both Arnoldi and Krylov-Schur have 8 processors and the parallelization efficiency is below 50%. Considering the above example, there is a possibility that the current implementation cannot effectively use the CPU.

  However, since the transition probability matrix is a sparse matrix, eigenvalues converge quickly and the number of iterations until convergence is small. Although the number of iterations actually required for convergence varies depending on the method and the number of processors, it has been confirmed that Krylov-Schur converges within 1157 times. The processing time is about 102.2 seconds and 56.77 seconds with two processors, which is about 1/4 of the processing time in the case of a Laplacian matrix. Therefore, it is considered that the number of required processors is reduced.

  Next, in order to investigate the influence of communication overhead, we compared the case of calculation with a single node and multiple nodes. Specifically, the case of executing on one node and the case of executing on two nodes were compared. In addition, evaluation was implemented by TCP / IP.

  Since the transition probability matrix is a sparse matrix, eigenvalues converge quickly and the number of calculation iterations decreases. Therefore, we confirmed the result of the Laplacian matrix, which increases the number of iterations. FIG. 13 shows the processing time when the number of repetitions is changed by calculation using two processors.

  Compared to processing with a single node, the MPI-related function call time is longer, but there are cases where the performance is better when a plurality of nodes are used due to the processing time of other parts. These may depend on the implementation of the Open MPI driver. As a result of this verification, if it becomes clear that the performance required by the information retrieval system will not be realized with the number of processors of a single node in the future, inter-node communication is used as long as the performance improvement rate investigated this time is within the range. There is a possibility that it can be supported by adding a processor.

  This is the end of the description of the verification regarding the processing speed of the eigenvalue calculation including the middleware for clusters implemented on the basis of MPI.

  In the above verification, the Power method implemented in SLEPc is not verified, but this method may also be suitably used in the information search system of the present invention.

  In the above verification, only the speeding up of the eigenvalue calculation part of the ranking function was examined. However, in practice, it is necessary to acquire a link structure between pages and input it to the eigenvalue calculation program as a transition probability matrix. As the matrix size increases, I / O can become a bottleneck. Here, the MPI has a parallel input / output function, and by using this, there is a possibility that performance degradation due to input / output can be avoided. In addition, since PETSc supports a data structure for efficiently storing a sparse matrix, the memory usage can be suppressed.

  From the above verification, it was confirmed that eigenvalues and eigenvectors of the transition probability matrix of hundreds of thousands of scales can be completed in tens of seconds on a PC of about 8 cores by using the parallel calculation library SLEPc.

1 Information Retrieval System 10 Terminal Information Acquisition Unit (Information Acquisition Unit)
13 Search execution part (information ordering means)

Claims (10)

An information search system for searching a Web page based on a keyword input at a terminal that has executed a search, and outputting and displaying the search result on the terminal,
Having information ordering means for ordering Web pages to be searched;
The information ordering means is an operation for estimating a transition tendency from a relation between a Web page linked to a Web page to be searched and a terminal currently connected to the Web page to be searched or within a certain period of time. An information retrieval system having a function of implementing   The transition probability is calculated by setting a square matrix by a configuration including a Web page to be searched and a terminal currently connected to the Web page to be searched or within a certain period, and calculating an eigenvector of the square matrix. The information search system according to claim 1, further comprising a calculation content.   The elements of the square matrix are configured by a value based on the number of link destinations to which a web page is linked and a value based on the number of web pages to which the terminal is actually connected within a certain period. The information search system according to claim 2.   The element of the square matrix is A / (A + B) where A is a value based on the number of link destinations to which the Web page is linked, and B is the number of Web pages that the terminal is currently connected to within a certain period. B / (A + B) or a value obtained by multiplying these by a system number. It has information acquisition means for acquiring terminal information, which is information related to a terminal outputting a Web page to be searched, at the time of search execution or before search execution,
The information ordering means specifies a Web page that satisfies a predetermined condition based on at least terminal information, and an output terminal that is a terminal that has output or has output the Web page,
When the output terminal is a virtual web page and a mutual link is formed between the web page output by the output terminal and the virtual web page, the link formed by the web page and the virtual web page Based on the structure, set a transition probability matrix with a transition from the display of the link source Web page to the display of the link destination Web page as a transition,
The information search system according to any one of claims 1 to 3, wherein a transition probability is calculated based on a set transition probability matrix, and Web pages to be searched are ranked based on a calculation result. Based on the terminal information, to calculate a user rank value indicating the importance of the user with respect to the knowledge and / or user search ability of the user using the output terminal,
The element of the square matrix and the transition probability matrix and / or at least one of the calculated square matrix or the eigenvector of the transition probability matrix is changed based on the user rank value. The information search system according to any one of 2 to 5. It has an information acquisition function capable of acquiring information related to the content of the web page on the network, and calculates a relation rank value indicating the importance of the content of the web page based on the information acquired by the information acquisition function.
The element of the square matrix and the transition probability matrix and / or at least one of the calculated square matrix or eigenvector of the transition probability matrix is changed based on the relation rank value. The information search system according to any one of 2 to 6. An information search system for searching a database based on a keyword input at a terminal and outputting and displaying the search result on the terminal,
It has information acquisition means for acquiring terminal information, which is information related to a terminal outputting a Web page to be searched, at the time of search execution or before search execution,
Having information ordering means for ordering Web pages to be searched;
The information ordering means identifies a Web page that satisfies a predetermined condition based on at least terminal information, an output terminal that is outputting or outputting the Web page, and based on the Web page and the output terminal Set a square matrix,
The square matrix is formed from a row element and a column element corresponding to a Web page and an output terminal, and each element P (i, j) satisfies the relationship of the following formula (1),
An information search system, wherein eigenvectors of the set square matrix are calculated, and Web pages to be searched are ordered based on the magnitudes of the calculated eigenvector components.
P (i, j) = αX / Y (1)
α = a number other than 0
X = number of links from the web page corresponding to column “j” to the web page corresponding to row “i”, or at the output terminal corresponding to row “i” of the web page corresponding to column “j” Number of displays, or number of web pages displayed corresponding to row “i” on the output terminal corresponding to column “j”, or 0
Y = sum of the number of Web pages displayed on the output terminal corresponding to the column “j” and the number of links formed on the Web page corresponding to the column “j” An information search method for searching a Web page based on a keyword input at a terminal that has executed a search, and outputting and displaying the search result on the terminal,
Performing an operation of estimating a transition tendency from a relationship between a Web page linked to a Web page to be searched and a terminal that is currently connected to the Web page to be searched or within a certain period of time. Information retrieval method. At the time of search execution or before execution of search, terminal information that is information related to the terminal that outputs the Web page to be searched is acquired,
A web page that satisfies a predetermined condition based on at least terminal information, and an output terminal that is the terminal that has output or has output the web page,
When the output terminal is a virtual web page and a mutual link is formed between the web page output by the output terminal and the virtual web page, the link formed by the web page and the virtual web page Based on the structure, set a transition probability matrix with a transition from the display of the link source Web page to the display of the link destination Web page as a transition,
The information search method according to claim 9, wherein the transition probability is calculated based on the set transition probability matrix, and the Web pages to be searched are ranked based on the calculation result.

Images