JP2011238174A - Keyword acquisition device, keyword acquisition method and keyword acquisition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To collect keywords with high accuracy in performing a task of acquiring a set of keywords having the same superordinate concept as a designated set of initial keywords.SOLUTION: In a keyword acquisition device 1, a keyword instance selecting unit 111 selects, on the basis of the score of each keyword in a keyword set calculated by a keyword instance score counting unit 110, a plurality of keywords out of the keyword set as the desired keyword set and selects, on the basis of the generality evaluation value of each keyword in the keyword set calculated by the keyword instance generality evaluating unit 200, a keyword set to be subjected to a surface layer pattern extracting unit 105 out of the pertinent keyword set. When the keyword selection is to be executed a prescribed number of times, the selected keyword set and the score of each keyword in the set are subjected to extraction of the surface layer pattern and calculation of the score thereof based on the keyword set to be processed.

Description

  The present invention relates to a technique for acquiring a keyword set having a superordinate concept similar to a keyword set inputted in advance from a document set existing inside a computer or accessible via a computer network. In particular, the present invention relates to a technique for acquiring a keyword set having the above superordinate concept using a search engine log that searches based on a search condition composed of one or more keywords.

  Conventionally, as a method for automatically acquiring a keyword having a specific superordinate concept, there is a method for analyzing and extracting a large amount of text data. For example, according to Non-Patent Document 1, there is a method using a bootstrap method based on a text pattern such as “A succi as B, C, and D” in order to acquire a combination of upper and lower concepts from English text. Proposed.

  Further, Non-Patent Document 2 describes a bootstrap method for extracting a pattern in which keywords frequently appear based on examples of keywords having a specific superordinate concept for a large amount of text information and acquiring the keywords using the patterns. Proposed a method for acquiring a keyword having a superordinate concept of a keyword input in advance.

  Further, in Non-Patent Document 3, by applying the method of Non-Patent Document 2 to a search condition log (hereinafter referred to as query log), a keyword set having a superordinate concept equivalent to the input keyword set is suppressed while reducing calculation cost. An identifying method has been proposed.

Hearst, M.M. : "Automatic Acquisition of Hyponyms from Large Text Corp.", Proc. of CORING, NANTES, AUG, 1992, p. 539-545 Patrick Pantel, Marco Pennacchioti: “Espresso: Leveraging Generic Patterns for Automatic Harvesting Semantic Relations”. Proc. of ACL, July, 2006, p. 113-120 Mamoru Komachi and Kumi Suzuki, “Improvement of semi-supervised acquisition of semantic knowledge from search logs”, Journal of the Japanese Society for Artificial Intelligence, Vol. 23, no. 3, March, 2008, p. 217-225

  However, the keyword acquisition technique of Non-Patent Document 1 is a method for comprehensively acquiring combinations of keywords of various superordinate concepts and subordinate concepts, and is not a method of extracting superordinate concept keywords specified by the user.

  The problem of the keyword acquisition technology of Non-Patent Documents 2 and 3 is the existence of a general pattern. When such a general pattern is adopted when alternately extracting patterns and keywords by the bootstrap method, there is a possibility that an unintended keyword is included in the acquired keyword. Note that a general pattern means a pattern that is not only a keyword having a specific superordinate concept but also highly related to a general keyword.

  Further, in the method of Non-Patent Document 3, measures are taken for this point when using the query log, but this method uses the number of keywords acquired by the pattern and is an essential solution. Absent.

  There are also general keyword issues as well as general pattern issues. This is because a general keyword is used for pattern evaluation, a general pattern is selected, and noise is mixed in the next keyword extraction step.

  The present invention has been made in view of the above problems, and in the process of extracting keywords related to a search, prevents acquisition of a general pattern and a general keyword, and is identical to a keyword set inputted with high accuracy. The goal is to acquire a set of keywords with concepts.

  In the present invention, when determining whether or not to adopt a text pattern of a search keyword based on a search condition log, the text pattern is considered by taking into account the overlap between the keyword set acquired by the pattern and the already acquired keyword set. Assess the generality of This prevents the adoption of a general text pattern. Further, when a keyword is extracted based on a set of text patterns, the generality of the keyword is evaluated based on the overlap of the set of text patterns and the set of text patterns that can be extracted with the keyword. This general keyword is not used for the evaluation of the surface layer pattern.

  As an aspect of the keyword acquisition device of the present invention, there is a keyword acquisition device that acquires a keyword set that has the same superordinate concept as a set of a plurality of keywords having the same superordinate concept and does not belong to the set. A keyword set having the same superordinate concept as the keyword set is provided as a processing target keyword set, and a group of keyword pairs including keywords belonging to the set is combined with a combination of two or more keywords from a plurality of keyword search condition logs. A surface layer that is obtained from log statistical information that is recorded by associating a keyword pair that is configured with its frequency, and that extracts a surface layer pattern that includes the keyword for each keyword that belongs to the processing target keyword set from the obtained keyword pair group Pattern extracting means, all the extracted surface layer patterns and Based on the keyword information used for the extraction of the surface layer pattern and the log statistical information, the score of the surface layer pattern, which is an index for evaluating the ease of acquiring the keyword related to the superordinate concept of the processing target keyword set, is obtained. Included in the surface layer pattern score calculating means for calculating the surface layer pattern, the keyword set consisting of the keyword group corresponding to the extracted surface layer pattern, and the surface layer pattern extracted from the keyword pair group matching the surface layer pattern in the log statistical information A surface pattern generality evaluation unit that performs an operation for calculating a generality evaluation value of the surface layer pattern based on a keyword set including a keyword group, and a generality evaluation of all the calculated surface layer patterns. Based on the value and score A surface layer pattern selecting means for selecting a certain number of surface layer patterns from the screen, and keyword pairs matching the selected surface layer patterns are acquired from the log statistical information and included in the surface layer pattern from the acquired keyword pairs Based on the frequency of the identified keyword and its keyword pair, a keyword instance extraction unit that identifies a set of keywords, and a score that is an index for evaluating the relevance between the keyword and a higher concept of the processing target keyword set A keyword instance score calculating means for calculating all the keywords belonging to the specified set, a surface pattern set comprising a group of surface patterns with the keyword for each keyword specified by the keyword instance extracting means, and the log statistical information The key Keyword instance generality evaluation means for performing an operation for calculating the generality evaluation value of the keyword based on the surface pattern set including the surface pattern group with the keyword extracted from the keyword pair group with the word for all the keywords And a keyword calculated by the keyword instance generality evaluation unit while selecting a plurality of keywords from the keyword set as the keyword set to be acquired based on the score of each keyword of the keyword set calculated by the keyword instance score calculation unit And a keyword instance selection means for selecting a keyword set for use in the surface pattern extraction means from the keyword set based on the generality evaluation value of each keyword of the set.

  As an aspect of the keyword acquisition method of the present invention, there is provided a keyword acquisition method for acquiring a keyword set that has the same superordinate concept as a set of a plurality of keywords having the same superordinate concept and does not belong to the set. A keyword set having the same superordinate concept as the keyword set to be acquired is provided as a processing target keyword set by the extraction means, and two keyword pairs including keywords belonging to this set are extracted from a log of search conditions for a plurality of keywords. Acquired from the log statistical information recorded in association with the frequency of the keyword pairs composed of the above keyword combinations, the keyword is included for each keyword belonging to the processing target keyword set from the acquired keyword pair group Step of extracting surface layer pattern and surface layer pattern score calculation Evaluates the ease of obtaining keywords related to the high-level concept of the processing target keyword set based on all the extracted surface layer patterns, keyword information used to extract the surface layer patterns, and the log statistical information. Calculating a score of a surface layer pattern that is an index to be performed for all the surface layer patterns, a keyword set consisting of keywords corresponding to the extracted surface layer pattern by the surface pattern generality evaluation means, and the surface layer in the log statistical information Performing a calculation for all the surface layer patterns for calculating a generality evaluation value of the surface layer pattern based on a keyword set including a keyword group included in the surface layer pattern extracted from the keyword pair group matching the pattern, and the surface layer pattern All the surface layers calculated by the selection means A step of selecting a certain number of surface layer patterns from all the surface layer patterns based on the generality evaluation value and the score of the turn, and a keyword instance extracting means assigns a keyword pair that matches each of the selected surface layer patterns to the log statistics A step of specifying a set of keywords included in the surface pattern from the acquired keyword pair, and a keyword instance score calculating means, based on the specified keyword and the frequency of the keyword pair, Calculating a score, which is an index for evaluating the relevance of a higher-level concept of the processing target keyword set, for all keywords belonging to the specified set; and a keyword instance generality evaluating unit, the keyword instance extracting unit Identified Based on a surface pattern set consisting of a group of surface patterns with the keyword for each keyword and a surface pattern set consisting of a group of surface patterns with the keyword extracted from the keyword pair group with the keyword in the log statistical information A step of calculating a generality evaluation value of the keyword for all the keywords, and a keyword instance selecting unit from the keyword set based on a score of each keyword of the keyword set calculated by the keyword instance score calculating unit A plurality of keywords are selected as the keyword set to be acquired, and from the keyword set based on the generality evaluation value of each keyword of the set of keywords calculated by the keyword instance generality evaluation unit And a step of selecting a keyword set to be subjected to serial surface pattern extraction means.

  Note that the present invention can also be an aspect of a keyword acquisition program that causes a computer to function as each means constituting the keyword acquisition device.

  According to the above invention, keywords can be collected with high accuracy in the task of acquiring a set of keywords having the same superordinate concept as the designated initial keyword set.

The block block diagram of the keyword acquisition apparatus which concerns on embodiment of invention. The hardware block diagram which mounts the keyword acquisition apparatus which concerns on embodiment of invention. The chart explaining the processing procedure which concerns on embodiment of invention. The chart figure explaining the log collection processing procedure concerning the embodiment of the invention. The flowchart figure explaining the log utilization keyword extraction processing procedure which concerns on embodiment of invention. An example of the data of the log storage device 201. An example of log statistical information stored in the log statistical DB 202. An example of data stored in the output result temporary storage unit 203.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[Overview]
A keyword acquisition apparatus 1 according to the present embodiment shown in FIG. 1 constitutes one component of an information search system including a search engine. The keyword acquisition apparatus 1 evaluates the generality of the surface layer pattern based on the overlap between the keyword set acquired from the surface pattern based on the search log obtained by the search engine and the already acquired keyword set, and adopts the general pattern. prevent. Furthermore, when extracting a keyword based on a set of surface layer patterns, the generality of the keyword is evaluated based on the overlap of the set of surface layer patterns and the set of surface layer patterns that can be extracted with the keyword, thereby further improving the keyword acquisition accuracy. Improve.

[Device configuration]
The keyword acquisition device 1 shown in FIG. 1 includes hardware resources such as the CPU 21, the memory 22, the storage devices 23 and 24, the input device 25, and the output device 26 shown in FIG. Then, these hardware resources cooperate with software resources (OS, applications, etc.) so that the log acquisition unit 100, log processing unit 101, seed instance acquisition unit 102, and output result determination unit 103 shown in FIG. , Result output unit 104, surface layer pattern extraction unit 105, surface layer pattern score calculation unit 106, surface layer pattern generality evaluation unit 107, surface layer pattern selection unit 108, keyword instance extraction unit 109, keyword instance score calculation unit 110, keyword instance selection unit 111, a keyword instance generality evaluation unit 200, a log storage device 201, a log statistics DB (database) 202, and an output result temporary storage unit 203 are implemented. Examples of the memory 22 include a volatile memory such as a RAM. Known storage means such as a hard disk drive may be applied to the storage devices 23 and 24. Hereinafter, each functional unit 100 to 203 will be described.

  The log acquisition unit 100 acquires a search condition (query) log input by the user to the search engine from the search engine and stores it in the log storage device 201.

  The log storage device 201 stores a log of search conditions (queries) input to the search engine acquired by the log acquisition unit 100. An example of the search condition is shown in FIG. The log storage device 201 can be realized by a known storage unit exemplified by a hard disk device, a server device, and the like.

  The log processing unit 101 acquires a query composed of a combination of two words from the log of the search condition of the user stored in the log storage device 201, and calculates its frequency. A query composed of a combination of two words indicates a query in which two keywords are connected by a single-byte or two-byte space, and is called a keyword pair. Log statistical information in which the calculated frequency is recorded in association with the keyword pair is stored in the log statistical DB 202. The frequency calculation method may be a well-known frequency calculation method employed in morphological analysis or the like.

  An example of the log statistics DB 202 is shown in FIG. The log statistics DB 202 is a storage device that records keyword pairs extracted from the search engine log by the log processing unit 101 and their appearance frequencies. The log statistics DB 202 may be realized by the storage device 24 exemplified by a hard disk drive device or a server device.

  The seed instance acquisition unit 102 receives input of a plurality of keywords included in a higher concept that the user wants to acquire vocabulary.

  The surface layer pattern extraction unit 105 accesses the log storage device 201 based on the keyword list (set) received from the seed instance acquisition unit 102 or the output result determination unit 103 and the keyword pair including the keyword from the log statistics DB 202 and its frequency. To get. Then, the surface layer pattern is extracted by replacing the keyword used for the inquiry (search) in the acquired keyword pair with a predetermined character string (## in the following example). Table 1 shows examples of inquiry keywords, acquired keyword pairs and their frequencies, and surface layer patterns. The surface layer pattern extraction unit 105 aggregates all keyword pairs extracted for each keyword, their frequencies, and surface layer patterns, and provides them to the keyword instance score calculation unit 110.

  The surface layer pattern score calculation unit 106 determines whether each surface layer is based on all keyword pairs extracted for each keyword acquired from the surface layer pattern extraction unit 105, their frequencies, and the surface layer pattern and keyword information used for extracting the surface layer pattern. Calculate the pattern score. The score of the surface layer pattern is an index for evaluating the ease of acquiring keywords related to the superordinate concept of the keyword set given as the initial input, and is calculated by the following formula (1), for example. The calculated score of the surface layer pattern is provided to the surface layer pattern selection unit 108 together with the surface layer pattern.

In Formula (1), Score _pattern (p) represents the score of the surface layer pattern p. Score _keyword (k) is keyword information used for extracting the surface layer pattern, and represents the score of keyword k (when the score is not defined, that is, when it is an initial value, 1). Can be calculated. K represents a set of keywords that can be acquired by the surface layer pattern p. Further, pmi (k, p) indicates a mutual information amount indicating the strength of relevance between the keyword k and the surface layer pattern p, and is calculated by the following equation (2) (the same applies to equation (4)).

  In Expression (2), N is the number of all keyword pairs, | k, p | is the frequency of keyword pairs having the keyword k, | k, * | is the number of keyword pairs having the keyword k, and | *, p | This represents the number of keyword pairs having the surface layer pattern p. Such information can be acquired from the log statistics DB 202.

  The surface layer pattern generality evaluation unit 107 is included in the surface layer pattern extracted from the keyword set consisting of the keyword group corresponding to the surface layer pattern extracted by the surface layer pattern extraction unit 105 and the keyword pair group matching the surface layer pattern in the log statistics DB 202. The calculation for calculating the generality evaluation value of the surface layer pattern based on the keyword set including the keyword group is performed on all the extracted surface layer patterns.

  Specifically, the surface layer pattern generality evaluation unit 107 calculates the generality evaluation value of the surface layer pattern by the following processes 1 to 4.

  (Process 1) The surface layer pattern extracted for each keyword is acquired from the surface layer pattern extraction unit 105. An example of the acquired information is shown in Table 2.

  (Process 2) Information indicating which keyword is extracted for each pattern from the information obtained in Process 1 is created. An example of the created information is shown in Table 3. Then, a keyword group corresponding to one surface layer pattern in this information is specified as a keyword set (K1). In the case of Table 3, the keyword set (K1) corresponding to the surface layer pattern “#### Deviation Value” is “ABC University DE University JK University”.

  (Process 3) The log statistics DB 202 is accessed using one surface layer pattern in the information created in Process 2 as a key, and a keyword pair group that matches the surface layer pattern is acquired. Extract keywords included. The extracted keyword group is specified as a keyword set (K2). Table 4 shows examples of surface layer patterns and keyword groups obtained thereby. In the case of Table 4, when the surface layer pattern is “## deviation value”, the keyword group included in the surface layer pattern extracted from the keyword pair group, that is, the keyword set (K2) is “T University K University ABC University DE “University H University”.

  (Process 4) Based on the keyword set (K1) acquired in Process 2 and the keyword set (K2) acquired in Process 3, the generality evaluation value of the one surface layer pattern is calculated. The generality evaluation value is calculated by, for example, the following formula (3). In Formula (3), Generality (p) is a value for evaluating the generality of the surface layer pattern p. K1 indicates a keyword set (K1), and K2 indicates a keyword set (K2). The surface layer pattern and its generality evaluation value are provided to the surface layer pattern selection unit 108.

  The surface layer pattern selection unit 108 selects a keyword based on each surface layer pattern provided from the surface layer pattern score calculation unit 106 and its score, and the generality evaluation value of each surface layer pattern provided from the surface layer pattern generality evaluation unit 107. Select a certain number of surface layer patterns for selection. Specifically, a certain threshold value is set for the generality evaluation value, a group of surface layer patterns having a generality evaluation value equal to or greater than this threshold value is identified as a group of uncommon surface layer patterns, and these identified surface layer patterns are identified as A certain number of patterns are acquired in descending order of the score, in descending order of score.

  The keyword instance extraction unit 109 acquires from the log statistics DB 202 a keyword pair that matches each surface layer pattern in the surface layer pattern set provided from the surface layer pattern selection unit 108 and the frequency thereof, and converts the acquired keyword pair into the surface layer pattern. Identify the set of included keywords. All the keywords extracted for each surface layer pattern and the frequency of the keyword pairs are provided to the keyword instance score calculation unit 110.

  The keyword instance score calculation unit 110 calculates the score of each keyword based on the set of keywords included in each surface pattern provided from the keyword instance extraction unit 109 and the frequency of each keyword pair belonging thereto. The keyword instance (the set of keywords) and the score of each keyword belonging to the keyword instance are provided to the keyword instance selection unit 111.

  The score of the keyword instance is an index that evaluates the relevance of the keyword and the superordinate concept of the keyword set specified in advance based on the relevance with the surface layer pattern used to acquire the keyword. The following formula (4) is given as an example of the score calculation formula.

In equation (4), Score _keyword (k) represents the score of keyword k. P represents a set of surface layer patterns in which the keyword k appears, and p represents one surface layer pattern in the surface layer pattern set. Such information can be acquired from the log statistics DB 202. The score _pattern (p) is a value calculated by the equation (1).

  The keyword instance generality evaluation unit 200 extracts, for each keyword specified by the keyword instance extraction unit 109, a surface layer pattern set including a group of surface layer patterns with the keyword and a keyword pair group with the keyword in the log statistics DB 202. An operation for calculating the generality evaluation value of the keyword based on a surface pattern set including a group of surface patterns with keywords is performed for all the keywords.

  Specifically, the keyword instance generality evaluation unit 200 calculates the generality evaluation value of the keyword by the following processes 11 to 14.

  (Process 11) A set of keywords included in each surface pattern specified by the keyword instance extraction unit 109 is acquired from the extraction unit 109. An example of the acquired information is shown in Table 5.

  (Process 12) Information on the surface layer pattern associated with the keyword is created for each keyword from the information obtained in Process 11. An example of the created information is shown in Table 6. A group of surface layer patterns with one keyword in this information is specified as a surface layer pattern set (P1). In the case of Table 6, the surface layer pattern set (P1) with the keyword “ABC University” is “## Deviation Value, ## Faculty of Science and Engineering, ## T City”.

  (Process 13) The log statistics DB 202 is accessed using one keyword in the information created in Process 12 as a search key to acquire a keyword pair group with the keyword, and the surface layer with the keyword included in the acquired keyword pair group Extract a group of patterns. The extracted surface layer pattern group is specified as a surface layer pattern set (P2). Table 7 shows an example of the one keyword and a group of surface layer patterns obtained thereby.

  (Process 14) Based on the surface layer pattern set (P1) acquired in Process 12 and the surface layer pattern set (P2) acquired in Process 13, a generality evaluation value of the one keyword is calculated. The generality evaluation value is calculated by, for example, the following formula (5). In Expression (5), Generality (k) is a value for evaluating the generality of the keyword k. P1 represents a surface layer pattern set (P1), and P2 represents a surface layer pattern set (P2). The keyword and its generality evaluation value are provided to the keyword instance selection unit 111.

  The keyword instance selection unit 111 selects a plurality of keywords from the keyword set as an output keyword set based on the keyword score of the keyword set calculated by the keyword instance score calculation unit 110. Specifically, the keywords provided from the keyword instance score calculation unit 110 are arranged in descending order of their scores, and a certain number of keywords are selected in descending order of score. The fixed number of selected keywords becomes an output keyword set (keyword set to be acquired).

  Further, the keyword instance selection unit 111 selects a keyword set for surface pattern evaluation from the keyword set based on the generality evaluation value of each keyword of the keyword set calculated by the keyword instance generality evaluation 200. Specifically, a keyword whose generality evaluation value calculated by the evaluation unit 200 exceeds a preset threshold is selected from the keyword set provided by the keyword instance generality evaluation unit 200. The selected keyword group becomes a keyword set for evaluating the surface layer pattern (a keyword set for use in the surface layer pattern extracting unit 105).

  The output result determination unit 103 acquires an output keyword list with a score and a keyword list for surface layer pattern evaluation from the keyword instance selection unit 111. Information on the output keyword list with scores is stored in the output result temporary storage 203 as a keyword extraction result. If there is a keyword extraction result that has already been stored in the output result temporary storage unit 203, it is stored in such a manner that it is added to the existing extraction result.

  Further, the output result determination unit 103 determines whether or not the number of keyword extractions performed for one input has reached a predetermined number of times, and the number of times reaches the predetermined number of times. If so, the result output unit 104 is notified that the result is output together with the score. On the other hand, if the number of times does not reach the predetermined number, the surface pattern evaluation keyword group and the score of each keyword are transmitted to the surface pattern extraction unit 105. The number of times is arbitrary.

  The output result temporary storage unit 203 is a storage device that temporarily stores the data transmitted from the output result determination unit 103. The output result temporary storage unit 203 can also be accessed from the result output unit 104. An example of the output result data is shown in FIG. The output result temporary storage unit 203 may be realized by the memory 22 shown in FIG.

  When the result output unit 104 receives that the result is output from the output result determination unit 103, the result output unit 104 accesses the output result temporary storage unit 203, acquires all stored keyword extraction results, and outputs the acquired extraction results. To do.

[Description of processing procedure]
A processing procedure by the keyword acquisition device 1 will be described. As shown in FIG. 3, the processing procedure includes a log collection processing procedure (S1) and a log use keyword extraction processing procedure (S2).

  In the log collection processing procedure (S1), a search condition log is acquired from a search engine, stored in the log storage device 201, and analysis (frequency calculation) of queries stored in the log storage device 201 is executed in a batch manner. The analysis result (log statistical information) is stored in the log statistical DB 202.

  In the log use keyword extraction processing procedure (S2), a keyword set having the same superordinate concept as the keyword set specified by the user using the analysis result contained in the log statistics DB 202 obtained in the log collection processing procedure (S1). Keywords not included in the keyword set designated by the user are extracted.

  Specific procedures of the above log collection processing procedure (S1) and log use keyword extraction processing procedure (S2) will be described with reference to FIGS.

  First, the log collection processing procedure (S1) will be described. The processing procedure (S1) includes steps S11 and S12 shown in FIG.

  S11: The log acquisition unit 100 acquires a search condition log input by a user to a search device such as a Web search engine from the search device and stores the log in the log storage device 201. For example, a search condition log as shown in FIG. 6 is acquired and stored in the log storage device 201.

  S12: The log processing unit 101 obtains a keyword pair composed of a combination of two or more keywords from the search condition log stored in the log storage device 201, and records the log pairs in association with the frequency. Are stored in the log statistics DB 202. For example, the log statistical information illustrated in FIG. 7 is stored in the log statistical DB 202.

  Next, the log use keyword extraction processing procedure (S2) will be described. The processing procedure (S2) includes steps S21 to S36 shown in FIG.

  S21: The seed instance acquisition unit 102 receives a keyword set having the same superordinate concept as the keyword to be acquired from the input device 25 as a processing target keyword set.

  S22: The seed instance acquisition unit 102 supplies the processing target keyword set received in S21 to the surface layer pattern extraction unit 105.

  S23: The surface layer pattern extraction unit 105 accesses the log statistics DB 202 based on the information of the processing target keyword set provided from the seed instance acquisition unit 102, and acquires the keyword pair including the keyword of the keyword set and its frequency. For example, as shown in Table 1, the keyword pair “ABC university deviation value” including the keyword and the frequency “2000” are acquired.

  S24: The surface layer pattern extraction unit 105 extracts the surface layer pattern by replacing the keyword used in the inquiry (search) with the predetermined character string (for example, ##) among the keyword pairs acquired in S23. For example, the surface layer pattern “#### deviation value” as shown in Table 1 is acquired. Then, the surface layer pattern extraction unit 105 supplies all the keyword pairs extracted for each keyword, their frequencies, and surface layer patterns to the surface layer pattern score calculation unit 106 and the surface layer pattern generality evaluation unit 107.

  S25: The surface layer pattern score calculation unit 106 calculates the score of each surface layer pattern by using the surface layer pattern extracted for each keyword provided from the surface layer pattern extraction unit 105 for the calculation of Expression (1). The calculated scores of all surface layer patterns are provided to the surface layer pattern selecting unit 108.

  S26: The surface layer pattern generality evaluation unit 107 supplies the surface layer pattern extracted for each keyword provided from the surface layer pattern extraction unit 105 to the above-described (Process 2) to (Process 4), and the generality evaluation value of each surface layer pattern Is calculated. Then, the calculated generality evaluation value of each surface layer pattern is provided to the surface layer pattern selection unit 108.

  S27: When the surface layer pattern selection unit 108 receives the generality evaluation value of each surface layer pattern from the surface layer pattern generality evaluation unit 107, the surface layer pattern selection unit 108 selects a surface layer pattern group that is a generality evaluation value equal to or greater than a threshold value as an uncommon surface layer pattern. Identify as a group. Next, the identified surface layer pattern group is arranged in descending order of the score calculated in S25, and a certain number of surface layer patterns are acquired in descending order of score. The specified set of surface layer patterns is provided to the keyword instance extraction unit 109.

  S28: Upon receiving a set of surface layer patterns from the surface layer pattern selection unit 108, the keyword instance extraction unit 109 acquires from the log statistics DB 202 the keyword pairs that match each surface layer pattern in the set and the frequency thereof. Then, a set of keywords included in the surface layer pattern is specified from the acquired keyword pair. All the keywords extracted for each surface layer pattern and the frequency of the keyword pairs are provided to the keyword instance score calculation unit 110.

  S29: The keyword instance score calculation unit 110 calculates the score of each keyword by using the keyword set and keyword pair frequency provided from the keyword instance extraction unit 109 for the calculation of Expression (4). The calculated keyword and the score of each keyword are provided to the keyword instance selection unit 111.

  S30: The keyword instance generality evaluation unit 200 uses the set of keywords included in each surface layer pattern provided from the keyword instance extraction unit 109 for the above (Process 12) to (Process 14), and evaluates the generality of each keyword. Calculate the value. Then, the calculated generality evaluation value of each keyword is provided to the keyword instance selection unit 111.

  S31: The keyword instance selection unit 111 selects an output keyword from the keyword set provided by the keyword instance score calculation unit 110 based on the score of each keyword. For example, the keywords provided from the keyword instance score calculation unit 110 are arranged in descending order of their scores, and a certain number of keywords are selected in descending order of score. The selected fixed number of keywords are provided to the output result determination unit 103 as an output keyword set.

  In S31, the keyword instance selection unit 111 selects a keyword for surface pattern evaluation based on the generality evaluation value of each keyword from the set of keywords provided from the keyword instance generality evaluation unit 200. For example, a keyword whose generality evaluation value calculated by the evaluation unit 200 exceeds a preset threshold is selected from the keyword set provided by the keyword instance generality evaluation unit 200. The selected keyword group is provided to the output result determination unit 103 as the surface pattern evaluation keyword set.

  S 32: The output result determination unit 103 stores the set of keywords selected in S 31 in the output result temporary storage unit 203.

  S33: The output result determination unit 103 determines whether or not the keyword extraction process of S23 to S32 has been executed a predetermined number of times.

  S34: If the output result determination unit 103 determines that the keyword extraction process has been executed a predetermined number of times in step S33, the output result determination unit 103 provides a message to the result output unit 104 indicating that the set of keywords selected in S31 is output.

  S35: Upon receiving the message from the output result determination unit 103, the result output unit 104 acquires a set of keywords stored in the output result temporary storage unit 203. Then, the output keyword set is output from the output device 26 as an appropriate keyword set for the user. For example, the output result data shown in FIG. 8 is output. In the illustrated example, the score calculated in S29 is attached to each keyword to be output.

S36: When the output result determination unit 103 determines that the keyword extraction process has not been executed the determined number of times in step S33, the output result temporary storage unit 203 stores the surface pattern evaluation keyword set stored in S23. This is provided to the surface layer pattern extraction unit 105 according to FIG. Further, the output result determination unit 103 provides the score of each keyword in the keyword set to the surface layer pattern extraction unit 105 as the score of the surface layer pattern in S25. The score of each keyword provided to the surface pattern extraction unit 105 is used as the score _keyword (k) of the equation (1) related to the calculation for calculating the score of the surface layer pattern executed by the surface layer pattern score calculation unit 106 in S25. The

[Effect of this embodiment]
As described above, the keyword acquiring apparatus 1 extracts a surface layer pattern highly relevant to a given keyword set in a task of acquiring a set of keywords having the same superordinate concept as the initial keyword set specified by the user. In addition, by evaluating the generality of the surface layer pattern based on the overlap between the keyword set and the keyword set that can be extracted by the pattern, a highly general surface layer pattern can be grasped. As a result, it is possible to prevent the use of keywords having a general surface layer pattern, and it is possible to suppress the extraction of keywords that are not related to the superordinate concept of the input keyword set. Therefore, search keywords can be collected with high accuracy.

  In information extraction using bootstrapping, there is a problem of semantic drift that, if a broad pattern is selected for acquiring an instance, an unintentional instance is extracted. In the keyword acquisition device 1, when selecting the surface layer pattern, the generality evaluation value that is the index that becomes the highest when the overlap between the instance set extracted so far and the instance set obtained by the surface layer pattern becomes half of the latter set Improves the keyword acquisition accuracy.

  Further, when the keyword acquisition device 1 extracts a keyword based on a set of surface layer patterns, the generality of the keyword is evaluated based on the overlap of the set of surface layer patterns and the set of surface layer patterns that can be extracted with the keyword. This general keyword is not used for the evaluation of the surface layer pattern. As a result, it is possible to prevent the adoption of a highly general text pattern that extracts many keywords other than the purpose, and the accuracy of keyword acquisition is further improved.

  Further, in the keyword acquisition device 1, the output result determination unit 103 determines whether or not the number of selections of the keyword has reached a predetermined number, and the number of selections reaches the predetermined number. If it is determined that the number of keywords selected by the keyword instance selection unit 111 is output, the selection unit 111 determines that the number of selections of the keyword has not reached a predetermined number. By setting the selected keyword set as the processing target keyword set and providing the keyword set and the score of each keyword to the surface layer pattern extraction unit 105, a more appropriate set of keywords can be obtained.

  Further, the surface layer pattern selection unit 108 identifies a group of surface layer patterns that are generality evaluation values equal to or greater than a threshold value among the calculated generality evaluation values as a group of uncommon surface layer patterns, and specifies the identified surface layer By arranging a group of patterns in descending order of the score of each surface layer pattern and selecting a certain number of surface layer patterns in descending order of the score, the surface layer patterns with higher generality can be more reliably excluded.

  Further, the keyword instance selection unit 111 selects a predetermined number of keywords from a set of keywords with a high score from the set of keywords whose scores are calculated, thereby further enhancing the ability to exclude a general surface layer pattern.

[Aspect as Program of the Present Invention]
The present invention can be implemented as a keyword acquisition program that causes a computer to function as the function units 100 to 111 and 200 to 203 included in the keyword acquisition device 1 in addition to those realized by dedicated hardware. A computer-readable recording medium storing this program is also an embodiment of the present invention. Examples of the recording medium include a flexible disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, other known recording media, and a storage device such as a hard disk device built in a computer system. Further, as the recording medium, a program system that dynamically holds a program for a short time (transmission medium or transmission wave) as in the case of transmitting a program via the Internet, and a computer system that serves as a server in that case Some of them hold a program for a certain period of time, such as an internal volatile memory.

DESCRIPTION OF SYMBOLS 1 ... Keyword acquisition apparatus 101 ... Log processing part 102 ... Seed instance acquisition part 103 ... Output result determination part 104 ... Result output part 105 ... Surface layer pattern extraction part (surface layer pattern extraction means)
106 ... Surface pattern score calculation unit (surface pattern score calculation means)
107 ... surface layer pattern generality evaluation section (surface layer pattern generality evaluation means)
108 ... Surface layer pattern selection section (surface layer pattern selection means)
109: Keyword instance extraction unit (keyword instance extraction means)
110: Keyword instance score calculation unit (keyword instance score calculation means)
111 ... Keyword instance selection unit (keyword instance selection means)
200: Keyword instance generality evaluation unit 202: Log statistics DB (log statistics information)

Claims (9)

A keyword acquisition device that acquires a keyword set that has the same superordinate concept as a set of a plurality of keywords having the same superordinate concept and does not belong to the set,
A keyword set having the same superordinate concept as the keyword set to be acquired is provided as a processing target keyword set, and a group of keyword pairs including keywords belonging to this set is searched for from two or more keyword search condition logs. Acquires keyword pairs composed of combinations from log statistical information recorded in association with their frequencies, and extracts a surface layer pattern including the keywords for each keyword belonging to the processing target keyword set from the group of acquired keyword pairs. Surface layer pattern extraction means to
An index for evaluating the ease of acquiring keywords related to the superordinate concept of the processing target keyword set based on all the extracted surface patterns, keyword information used for extracting the surface patterns, and the log statistical information Surface layer pattern score calculating means for calculating the score of the surface layer pattern for all the surface layer patterns;
A keyword set consisting of a keyword group corresponding to the extracted one surface pattern and a keyword set consisting of a keyword group included in the surface pattern extracted from the keyword pair group matching the one surface pattern in the log statistical information; A surface layer pattern generality evaluation means for performing an operation for calculating the generality evaluation value of the one surface layer pattern based on the above-mentioned all surface layer patterns;
Surface layer pattern selecting means for selecting a certain number of surface layer patterns from all the surface layer patterns based on the generality evaluation values and scores of all the calculated surface layer patterns,
Keyword instance extraction means for acquiring a keyword pair that matches each of the selected surface layer patterns from the log statistical information, and identifying a set of keywords included in the surface layer pattern from the acquired keyword pairs;
Based on the frequency of the specified keyword and its keyword pair, a score, which is an index for evaluating the relevance between the keyword and the superordinate concept of the processing target keyword set, is calculated for all keywords belonging to the specified set. A keyword instance score calculating means;
For each keyword specified by the keyword instance extraction means, a surface pattern set including a surface pattern group with the keyword and a group of surface patterns with the keyword extracted from a keyword pair group with the keyword in the log statistical information A keyword instance generality evaluation unit that performs an operation for calculating the generality evaluation value of the keyword based on the surface pattern set of
Based on the score of each keyword of the keyword set calculated by the keyword instance score calculating means, a plurality of keywords are selected from the keyword set as the keyword set to be acquired, and the keyword set calculated by the keyword instance generality evaluating means And a keyword instance selecting means for selecting a keyword set for use in the surface pattern extraction means from the keyword set based on the generality evaluation value of each keyword. It is determined whether or not the number of times of selection of the keyword has reached a predetermined number of times, and if it is determined that the number of times of selection has reached the predetermined number of times, the keyword instance selection means has selected While outputting a set of keywords, if it is determined that the number of selections of the keyword has not reached a predetermined number, the keyword set selected by the keyword instance selection means is set as the processing target keyword set. 2. The keyword acquisition apparatus according to claim 1, further comprising output result determination means for providing a set and a score of each keyword to the surface layer pattern extraction means. The surface layer pattern selection means identifies a group of surface layer patterns that are generality evaluation values equal to or greater than a threshold value among the calculated generality evaluation values as a group of surface layer patterns that are not general, and 3. The keyword acquiring apparatus according to claim 1, wherein groups are arranged in descending order of scores of the respective surface layer patterns, and a predetermined number of surface layer patterns are selected in descending order of score. 4. The keyword instance selecting unit selects a predetermined number of keywords from a set of keywords with a high score from the set of keywords for which the score is calculated. 5. Keyword acquisition device. A keyword acquisition method for acquiring a keyword set that has the same superordinate concept as a set of a plurality of keywords having the same superordinate concept and does not belong to the set,
A keyword set having the same superordinate concept as the keyword set that the surface layer extraction means wants to acquire is provided as a processing target keyword set, and a group of keyword pairs including keywords belonging to this set is retrieved from a log of search conditions for a plurality of keywords. A keyword pair composed of a combination of two or more keywords is acquired from log statistical information recorded in association with the frequency, and the keyword is assigned to each keyword belonging to the processing target keyword set from the acquired keyword pair group. Extracting the included surface pattern; and
Based on all the extracted surface layer patterns, the keyword information used for extracting the surface layer pattern, and the log statistical information, the surface layer pattern score calculation means obtains keywords related to the superordinate concept of the processing target keyword set Calculating a score of a surface layer pattern, which is an index for evaluating easiness, for all the surface layer patterns;
The surface layer pattern generality evaluation means is included in the surface layer pattern extracted from the keyword set consisting of the keyword group corresponding to the extracted one surface layer pattern and the keyword pair group matching the one surface layer pattern in the log statistical information Performing an operation on all the surface layer patterns to calculate a generality evaluation value of the one surface layer pattern based on a keyword set consisting of a keyword group that includes:
A step of selecting a certain number of surface layer patterns from all the surface layer patterns based on the generality evaluation values and scores of all the surface layer patterns calculated by the surface layer pattern selecting means;
A keyword instance extraction unit acquires a keyword pair that matches each of the selected surface layer patterns from the log statistical information, and identifies a set of keywords included in the surface layer pattern from the acquired keyword pairs;
The keyword instance score calculation means assigns a score, which is an index for evaluating the relevance between the keyword and a higher concept of the processing target keyword set, to the specified set based on the frequency of the specified keyword and the keyword pair. Calculating for all the keywords to which it belongs;
The keyword instance generality evaluation means is extracted from a surface pattern set consisting of a group of surface patterns with the keyword for each keyword specified by the keyword instance extraction means and a keyword pair group with the keyword in the log statistical information Performing an operation on all the keywords to calculate a generality evaluation value of the keyword based on a surface pattern set consisting of a group of surface patterns with the keyword;
The keyword instance selecting means selects a plurality of keywords as the keyword set to be acquired from the keyword set based on the score of each keyword of the keyword set calculated by the keyword instance score calculating means, and the keyword instance generality evaluating means Selecting a keyword set for use in the surface pattern extraction means from the keyword set based on the generality evaluation value of each keyword of the keyword set calculated by the method. When the output result determination means determines whether the number of times of selection of the keyword has reached a predetermined number of times and determines that the number of times of selection has reached the predetermined number of times, the keyword instance While outputting a set of keywords selected by the selection means, if it is determined that the number of times of selection of the keyword has not reached a predetermined number of times, the set of keywords selected by the keyword instance selection means is 6. The keyword acquiring method according to claim 5, further comprising a step of providing the keyword set and a score of each keyword as processing target keyword sets to the surface layer pattern extracting means. In the step of selecting the surface layer pattern, a group of surface layer patterns having a generality evaluation value equal to or greater than a threshold value among the calculated generality evaluation values is specified as a group of uncommon surface layer patterns, and the specified The keyword acquisition method according to claim 5 or 6, wherein a group of surface layer patterns is arranged in descending order of the score of each surface layer pattern, and a certain number of surface layer patterns are selected in descending order of score. 8. The method according to claim 5, wherein in the step of selecting as a keyword set to be acquired, a predetermined number of keywords are selected from a set of keywords having a high score from the set of keywords for which the score is calculated. The keyword acquisition method according to item 1.   5. A keyword acquisition program for causing a computer to function as each means constituting the keyword acquisition device according to claim 1.

Images