Knowledge data generation device and program

The knowledge data generation device enhances content presentation by processing multimodal data to select accurate keywords through hierarchical clustering, addressing the limitations of existing methods and providing comprehensive content coordination.

JP2024157930AInactive Publication Date: 2024-11-08NIPPON HOSO KYOKAI
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2023072607
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-04-26
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing content presentation methods struggle to provide users with accurate and comprehensive content beyond curriculum guidelines, requiring manual data re-entry and often result in insufficient content coordination due to inaccurate keyword selection.

Method used

A knowledge data generation device that processes multimodal data to generate hierarchical keywords, performs clustering, and selects highly accurate keywords for content presentation, using a hierarchical clustering process to refine and optimize content search.

Benefits of technology

Enables effective content coordination and accurate content presentation across various fields, reducing noise and ensuring comprehensive content recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024157930000001_ABST
    Figure 2024157930000001_ABST
Patent Text Reader

Abstract

To generate data for presenting highly accurate content by actualizing the association of various types of content.SOLUTION: A candidate pair selection part 30 of a knowledge data generation device 1 uses hierarchical data to calculate the distance between keywords etc., and selects a pair of shortest distances as a candidate pair. A gravity center position calculation part 31 calculates a gravity center position G based upon scores of the candidate pair. A clustering processing part 32 deletes a candidate which is far away from the gravity center position G of the candidate pair when it is determined that the gravity center position G has a deviation, and generates a new cluster to update the hierarchical data when it is determined that the gravity center position has no deviation. The clustering processing part 32 repeats the processing until keywords and clusters included in the hierarchical data decrease to one, and generates a clustering result. A keyword selection processing part 23 selects a keyword after clustering from the clustering result according to a previously set limit number, and generates knowledge data.SELECTED DRAWING: Figure 3
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a technique for generating a database for searching for content. [Background technology]

[0002] Conventionally, there has been progress in the development of a technology for linking contents by semantically linking the contents with each other, thereby presenting the content desired by a user. A technology for linking contents is a technology for connecting contents with each other.

[0003] For example, in the field of education, a method has been disclosed for presenting content related to content currently being used in order to assist a user in self-directed learning (see, for example, Patent Document 1 and Non-Patent Document 1).

[0004] In general, it is difficult for an instructor to fully and appropriately grasp a user's level of proficiency, and users are expected to study independently according to their own level of proficiency. On the other hand, with the spread of electronic learning materials, a wide variety of content is being created for use in learning.

[0005] When a user self-study, he / she needs to search for the content suitable for him / herself from among various contents, and therefore a content search method according to the level of proficiency is desired. In other words, a method capable of presenting the content related to the content currently being used for learning is desired to assist the user's spontaneous learning.

[0006] As an example of such a technique, Patent Document 1 discloses a technique in which a learning area is defined based on a curriculum guideline, and recommended teaching material content is presented to the user based on the grade, subject, and textbook unit from within that area.

[0007] In addition, non-patent document 1 discloses a method of structuring a learning order and learning system based on the introduction of each unit written in a textbook, and providing learning support to users using the structured data.

[0008] This content presentation method takes into account the fact that there are large differences in the levels of proficiency among users. For example, this content presentation method is used because a uniform teaching method such as a group lesson cannot provide education that matches the level of proficiency of the user.

[0009] Meanwhile, a method for generating keywords for linking contents has been disclosed (see, for example, Non-Patent Documents 2 and 3). This method extracts objects using images, audio, and subtitles as input data, lists and scores metadata calculated from the objects, and selects the keyword with the highest score. [Prior art documents] [Patent documents]

[0010] [Patent Document 1] JP 2015-18159 A [Non-patent literature]

[0011] [Non-Patent Document 1] Makoto Urakawa, Masaru Miyazaki, Ichiro Yamada, Hiroshi Fujisawa, "A Study on the Systematization of Content Based on the Course of Study," Japanese Society for Artificial Intelligence, Second Type Research Meeting Materials, vol. 2016, no. SWO-038, p. 03, 2016 [Non-Patent Document 2] Yuto Komatsu, Shoko Fujii, Shinya Abe, Satoshi Fujitsu, Kinji Matsumura, Hiroshi Fujisawa, “Development of Keyword Generation Technology for Content Data Integration,” Information Processing Society of Japan, Proceedings of the 21st Forum on Information Science and Technology (FIT2022), No. 4, O-035, p. 387-388, 2022 [Non-Patent Document 3] Yuto Komatsu, Shoko Fujii, Shinya Abe, Satoshi Fujitsu, Hiroshi Fujisawa, “Development of content description generation technology for content data integration”, Institute of Image Information and Television Engineers, Proceedings of the 2021 Winter Conference of the Institute of Image Information and Television Engineers, 32A-4, 2021 Summary of the Invention [Problem to be solved by the invention]

[0012] However, with the method of presenting content based on the curriculum guidelines described in the aforementioned Patent Document 1, and the method of presenting content based on the learning order described in the aforementioned Non-Patent Document 1, it is difficult for users to be presented with content other than teaching materials (e.g., broadcast program content) that is related to the content they are currently using for their studies.

[0013] In other words, these methods present content in areas related to and adjacent to the curriculum guidelines, and the area of ​​the presented content is limited. Therefore, although the user can receive content for learning, the effect is not sufficient.

[0014] Furthermore, after a user has been presented with content, in order to continue receiving content, the user must re-enter data, which requires costs in terms of manpower and work time.

[0015] Here, if the user can receive presentations of content other than the teaching materials, such as broadcast program content, in addition to the teaching material content, it is expected that the user will be able to achieve effective learning.

[0016] To realize such effective learning, it is necessary to build a database that can present content from various fields other than teaching materials, and it is expected to use the methods described in the above-mentioned non-patent documents 2 and 3. With this method, content linkage is realized using keywords and a database is built.

[0017] However, the methods of Non-Patent Documents 2 and 3 extract objects from multimodal images, audio, and subtitles, list and score them, and select the keyword with the highest score. Therefore, the selected keyword may contain noise, and it is not always possible to select a keyword with high accuracy. The use of such keywords leads to insufficient content linkage, and as a result, there is a problem that it is not possible to build an effective database for presenting desired content to users.

[0018] In this way, in order to search and present content related to currently used content from various domains, not only in the field of education but also in fields other than education, it is necessary to build an effective database that links content together. By using the built database, content from various domains can be linked, and users can search for desired content with high accuracy. Hereinafter, the data used to link content together will be referred to as "knowledge data."

[0019] Therefore, the present invention has been made to solve the above-mentioned problems, and its object is to provide a knowledge data generation device and program that realizes the linkage of various contents and generates data for presenting highly accurate contents. [Means for solving the problem]

[0020] In order to solve the above problem, a knowledge data generation device according to claim 1 is a knowledge data generation device for determining keywords related to content and generating knowledge data to be used when searching for desired content, the knowledge data generation device comprising: a data processing unit for detecting character strings of text for each of the multiple types of data by a predetermined process according to the data, the content being composed of multiple types of data, and determining the keywords from the character strings of the text; a score calculation unit for scoring the keywords determined for the multiple types of data by the data processing unit, and calculating a score obtained by integrating all of the predetermined processes for each same keyword; and a keyword selection unit for performing a hierarchical clustering process using hierarchical data composed of multiple hierarchically structured words and including the multiple keywords determined by the data processing unit, to determine whether to delete the keyword, generate a cluster, or delete the cluster based on the score calculated by the score calculation unit, and selecting a clustered keyword from the multiple words including the multiple keywords constituting the hierarchical data, and generating the knowledge data including the clustered keyword.

[0021] Further, the knowledge data generation device of claim 2 is characterized in that in the knowledge data generation device of claim 1, the keyword selection unit performs the hierarchical clustering process, selects keywords after the clustering from the plurality of keywords included in the hierarchical data, and generates the knowledge data.

[0022] Further, the knowledge data generation device of claim 3 is characterized in that in the knowledge data generation device of claim 1, the keyword selection unit performs the hierarchical clustering process, selects clustered keywords from a plurality of words that are superordinate concepts of the plurality of keywords included in the hierarchical data, and generates the knowledge data.

[0023] Further, the knowledge data generation device of claim 4 is the knowledge data generation device of claim 1, wherein the keyword selection unit calculates distances in the hierarchical data between one or more keywords and one or more clusters for which the hierarchical clustering process has been performed, and selects two of the one or more keywords and the one or more clusters for which the distance is shortest as a candidate pair; a center of gravity calculation unit calculates a center of gravity of the candidate pair based on the score of the candidate pair selected by the candidate pair selection unit; and based on the center of gravity calculated by the center of gravity calculation unit, deletes from the hierarchical data a candidate of the candidate pair that is farther from the center of gravity; or the candidate pair selection unit performs the hierarchical clustering process of clustering the candidate pairs to generate new clusters, or the candidate pair selection unit performs the hierarchical clustering process to update the hierarchical data, and repeats the process of selecting the candidate pairs by the candidate pair selection unit, the process of calculating the center of gravity by the center of gravity calculation unit, and the process of performing the hierarchical clustering process to update the hierarchical data until one keyword or one cluster remains in the hierarchical data, thereby obtaining a clustering result based on the updated hierarchical data; and a keyword selection processing unit selects the clustered keywords from the clustering result obtained by the clustering processing unit, and generates the knowledge data.

[0024] Further, the knowledge data generation device of claim 5 is the knowledge data generation device of claim 4, wherein, when performing the hierarchical clustering process, the clustering processing unit calculates a first distance from the center of gravity position to one of the candidate pair and calculates a second distance from the center of gravity position to the other of the candidate pair, and if the longer of the first distance and the second distance is equal to or greater than a predetermined first set value and the shorter of the first distance and the second distance is equal to or less than a predetermined second set value, it determines that there is a bias and deletes the candidate corresponding to the longer of the candidate pair from the hierarchical data, and if the longer distance is less than the first set value or the shorter distance is greater than the second set value, it determines that there is no bias and clusters the candidate pairs to generate new clusters.

[0025] Further, the knowledge data generation device of claim 6 is characterized in that in the knowledge data generation device described in claim 4, the keyword selection processing unit selects the clustered keywords from the clustering result according to a preset limit number, and generates the knowledge data.

[0026] Further, the program of claim 7 is characterized in that the computer constituting a knowledge data generation device for determining keywords related to content and generating knowledge data to be used when searching for desired content is made to function as: a data processing unit which, assuming that the content is composed of multiple types of data, detects character strings of text for each of the multiple types of data by a predetermined process corresponding to the data, and determines the keywords from the character strings of the text; a score calculation unit which scores the keywords determined for the multiple types of data by the data processing unit, and calculates a score that integrates all of the predetermined processes for each same keyword; and a keyword selection unit which uses hierarchical data composed of a plurality of hierarchically structured words and including the multiple keywords determined by the data processing unit to perform a hierarchical clustering process for each of the multiple keywords, based on the score calculated by the score calculation unit, to delete the keyword, generate a cluster, or delete the cluster, select a clustered keyword from the multiple words including the multiple keywords constituting the hierarchical data, and generate the knowledge data including the clustered keyword. Effect of the Invention

[0027] As described above, according to the present invention, it is possible to realize the linkage of various contents and generate data for presenting highly accurate contents. [Brief description of the drawings]

[0028] [Figure 1] 1 is a diagram for explaining an overview of an entire system using a knowledge data generation device according to an embodiment of the present invention; [Diagram 2] 1 is a block diagram showing an example of the configuration of a knowledge data generation device according to an embodiment of the present invention; [Diagram 3] 11 is a block diagram showing an example of the configuration of a keyword selection unit; FIG. [Figure 4] 13 is a flowchart showing an example of processing by a keyword selection unit; [Diagram 5]13 is a flowchart showing an example of a hierarchical clustering process (step S404). [Figure 6] FIG. 4 is a diagram for explaining an example of score calculation by a score calculation unit in the first embodiment. [Figure 7] FIG. 2 is a diagram illustrating an example of hierarchical data in the first embodiment. [Figure 8] FIG. 13 is a diagram showing distances between keywords at the start in the first embodiment. [Figure 9] FIG. 1 is a diagram showing the center of gravity between "skis" and "snowboard." [Figure 10] FIG. 13 is a diagram showing an example of hierarchical data after clustering of "skiing" and "snowboarding." [Figure 11] FIG. 13 is a diagram showing updated keywords and distances in clusters. [Figure 12] FIG. 13 is a diagram showing the center of gravity between the cluster and the "skateboard." [Figure 13] FIG. 13 is a diagram illustrating a process for deleting "skateboard" when it is determined that there is a bias. [Figure 14] FIG. 13 is a diagram showing an example of hierarchical data after "skateboard" has been deleted. [Figure 15] FIG. 13 is a diagram showing a final clustering result in Example 1. [Figure 16] FIG. 11 is a diagram illustrating an example of score calculation by a score calculation unit in the second embodiment. [Figure 17] FIG. 11 is a diagram illustrating an example of hierarchical data in the second embodiment. [Figure 18] FIG. 13 is a diagram showing distances between keywords at the start in the second embodiment. [Figure 19] FIG. 13 is a diagram showing the position of the center of gravity between “acceleration” and “acceleration sensor.” [Figure 20] FIG. 13 is a diagram showing an example of hierarchical data after clustering of “acceleration” and “acceleration sensor”. [Figure 21] FIG. 13 is a diagram showing updated keywords and distances in clusters. [Figure 22] FIG. 13 shows the center of gravity between the clusters and "gravity." [Diagram 23] FIG. 13 is a diagram showing an example of hierarchical data after clustering of "acceleration", "acceleration sensor", and "gravity". [Figure 24] FIG. 13 is a diagram showing updated keywords and distances in clusters. [Diagram 25] FIG. 13 is a diagram showing the center of gravity between "smartphone" and clusters. [Figure 26] FIG. 13 is a diagram illustrating a process of deleting "smartphone" when it is determined that there is a bias. [Figure 27] FIG. 13 is a diagram showing a final clustering result in Example 2. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0029] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. Summary of the Invention First, an overview of the present invention will be described. The present invention uses multimodal data constituting a content and knowledge data related to keywords in a hierarchical structure (hereinafter referred to as "hierarchical data") to generate knowledge data used to connect the content with various other contents.

[0030] 1 is a diagram for explaining an outline of an entire system using a knowledge data generating device according to an embodiment of the present invention. This system is configured with a knowledge data generating device 1, external data 2-1 consisting of a database (DB) in a specific field, external data 2-2 consisting of the contents of library data (data managed by a library), external data 2-3 consisting of the contents of museum data (data managed by a museum), broadcasting station data 3-1 consisting of the broadcasting program contents of program information, broadcasting station data 3-2 consisting of the broadcasting program contents of news, and user terminals 4-1, 4-2, 4-3, etc.

[0031] The knowledge data generation device 1 inputs audio data, image data, and subtitle data constituting broadcast program content from, for example, broadcast station data 3-1, and also inputs hierarchical data from external data 2-1, 2-2, 2-3, etc. Please refer to the following website for examples of hierarchical data. "Wikidata", [online], [searched on February 21, 2023], Internet<URL:https: / / www.wikidata.org / wiki / Wikidata:Main_Page>

[0032] Here, hierarchical data, i.e., knowledge data on hierarchically structured keywords, is made up of a group of words that are hierarchically structured from the viewpoint of meaning, and are arranged in a tree structure from higher-level conceptual words to lower-level conceptual words. It is assumed that external data 2-1, 2-2, 2-3, etc. hold hierarchical data.

[0033] The knowledge data generation device 1 uses multiple types of multimodal data consisting of audio data, image data, and subtitle data, as well as hierarchical data, to find vocabulary with optimal meanings that express the broadcast program content as keywords.Then, the knowledge data generation device 1 generates knowledge data consisting of content information such as information (identification information) for identifying the broadcast program content, and keywords.

[0034] The knowledge data generation device 1 finds keywords using data acquired from broadcast station data 3-1, 3-2, etc. and external data 2-1, 2-2, 2-3, etc., and generates knowledge data, thereby constructing a DB of knowledge data.

[0035] A user operating the user terminal 4-1, 4-2, 4-3, etc., while viewing broadcast program content of a health program in the broadcast station data 3-1, for example, searches the knowledge data DB using a keyword selected from the broadcast program content.

[0036] For example, if a user operating user terminal 4-1 wishes to obtain content related to "DNA diagnosis", a search is performed using the knowledge data DB to obtain content such as "DNA diagnosis of cancer" from broadcast station data 3-2. Also, a user operating user terminal 4-2 obtains content such as "Structure of DNA in Living Things" from broadcast station data 3-1 and "Human Body II Genes" from external data 2-3.

[0037] In this way, by using the knowledge data, it is possible to connect contents in various fields (external data 2-1, 2-2, 2-3, etc. and broadcasting station data 3-1, 3-2, etc.), and to recommend desired contents to the user. In addition, it is possible to present contents in a form that broadens the user's interests.

[0038] In other words, knowledge data for linking various contents and presenting highly accurate contents can be generated by the knowledge data generation device 1. Hereinafter, the external data 2-1, 2-2, 2-3, etc. are collectively referred to as external data 2, and the broadcast station data 3-1, 3-2, etc. are collectively referred to as broadcast station data 3.

[0039] [Knowledge Data Generator 1] Next, a description will be given of the knowledge data generation device 1. Fig. 2 is a block diagram showing an example of the configuration of the knowledge data generation device 1 according to an embodiment of the present invention.

[0040] This knowledge data generation device 1 includes an input unit 10, a data processing unit 11, a memory 12, a score calculation unit 13, and a keyword selection unit 14. The data processing unit 11 includes a voice processing unit 111, an image object processing unit 112, an image text processing unit 113, and a subtitle processing unit 114.

[0041] The knowledge data generation device 1 uses not only image data included in the content, but also audio data, subtitle data, and hierarchical data to find vocabulary with the most appropriate meaning to express the content as keywords.

[0042] The input unit 10 inputs content information, such as information for identifying broadcast program content, and data of the broadcast program content (audio data, image data, and subtitle data), for example, from the broadcast station data 3 out of the external data 2 and the broadcast station data 3. Then, the input unit 10 outputs the content information to the keyword selection unit 14.

[0043] The input unit 10 extracts audio data, image data, and subtitle data from the data of the broadcast program content, and outputs these data to the data processing unit 11. Specifically, the input unit 10 outputs the audio data to the audio processing unit 111, the image data to the image object processing unit 112 and the image text processing unit 113, and the subtitle data to the subtitle processing unit 114.

[0044] The data processing unit 11 inputs audio data, image data, and subtitle data from the input unit 10, and uses this multimodal data to find keywords for the broadcast program content for each frame of the moving images of the broadcast program content, and stores the keywords for each frame in the memory 12.

[0045] Specifically, the voice processing unit 111 receives voice data from the input unit 10, performs voice recognition processing on the voice data for each frame to detect character strings of text about the voice, and obtains keywords from the detected character strings of text. The voice processing unit 111 then stores the keywords obtained by the voice recognition processing in the memory 12. The process of obtaining keywords by voice recognition processing on voice data is known, and therefore a detailed description thereof will be omitted here.

[0046] The image object processing unit 112 receives image data from the input unit 10, and performs object recognition processing on the image data for each frame to detect an area including an object, detect a character string of a sentence including the name of the object from the area, and obtain a keyword from the detected character string of the sentence. For example, if the object is an "automobile", a character string of a sentence including the name of the object is detected such as "A car is running". The image object processing unit 112 then stores the keyword obtained by the object recognition processing on the image data in the memory 12. Since the processing for obtaining a keyword by object recognition processing on image data is known, a detailed description will be omitted here.

[0047] The image character processing unit 113 receives image data from the input unit 10, and performs character recognition and morphological analysis on the image data for each frame to detect character strings of text containing characters, and obtains keywords from the detected character strings of text. The image character processing unit 113 then stores the keywords obtained by the character recognition and morphological analysis of the image data in the memory 12. The process of obtaining keywords by the character recognition and morphological analysis of image data is well known, and therefore a detailed description thereof will be omitted here.

[0048] The subtitle processor 114 receives the subtitle data from the input unit 10, and performs morphological analysis on the subtitle data for each frame to detect character strings of text about the subtitles, and obtains keywords from the detected character strings of text. The subtitle processor 114 then stores the keywords obtained by the morphological analysis of the subtitle data in the memory 12. The process of obtaining keywords by morphological analysis of the subtitle data is well known, and therefore a detailed description thereof will be omitted here.

[0049] As a result, memory 12 stores, in a list form, each keyword obtained for each frame of moving images in the broadcast program content by voice recognition processing on audio data (hereinafter referred to as "voice recognition processing"), object recognition processing on image data, character recognition processing on image data, etc., and morphological analysis processing on subtitle data (hereinafter referred to as "subtitle analysis processing").

[0050] The score calculation unit 13 reads out, for each frame, from the memory 12, each keyword obtained by the voice recognition process, the object recognition process from image data, the character recognition process from image data, etc., and the subtitle analysis process. The score calculation unit 13 also inputs weights (weights for each process) for the voice recognition process, the object recognition process from image data, the character recognition process from image data, etc., and the subtitle analysis process. The weights for each process are set in advance and can be changed by a user operation.

[0051] The score calculation unit 13 counts the number of occurrences of each keyword for each of the voice recognition process, the object recognition process from image data, the character recognition process from image data, and the subtitle analysis process for each of the same keywords for each frame. Then, the score calculation unit 13 multiplies the count value of each process by a weight for each of the same keywords to obtain a multiplication result for each process, and adds (merges) the multiplication results for each process to obtain a score for one frame that integrates each process. The score calculation unit 13 stores the obtained score in the memory 12. An example of score calculation by the score calculation unit 13 will be described later.

[0052] In a first embodiment described later, the score calculation unit 13 stores in the memory 12 the score for one frame for each identical keyword.

[0053] Furthermore, in a second embodiment described later, the score calculation unit 13 calculates a score for N frames (N consecutive frames) by adding up the scores of N consecutive frames set in advance for each of the same keywords. N is an integer equal to or greater than 2. The score calculation unit 13 stores the scores of the N frames for each of the same keywords in the memory 12.

[0054] The keyword selection unit 14 inputs content information from the input unit 10, reads all (multiple) keywords with scores greater than 0 and the scores for each keyword from the memory 12, and inputs hierarchical data from the external data 2. Here, the multiple words constituting the hierarchical data include the multiple keywords read from the memory 12 (for example, in FIG. 7 described later, the hierarchical data includes keywords marked with *).

[0055] The keyword selection unit 14 performs a hierarchical clustering process on the multiple keywords included in the hierarchical data, and selects keywords that have been subjected to the hierarchical clustering process (hereinafter, keywords after clustering) from the multiple words (including multiple keywords) that constitute the hierarchical data. Details of the hierarchical clustering process performed by the keyword selection unit 14 will be described later.

[0056] In this case, the keyword selection unit 14 may select a keyword after clustering from a plurality of keywords included in the hierarchical data (from a plurality of keywords read from the memory 12). The keyword selection unit 14 may also select a keyword after clustering from a word of a higher concept for each of a plurality of keywords included in the hierarchical data (from a word other than the plurality of keywords). The keyword selection unit 14 may also select a keyword after clustering from a plurality of keywords included in the hierarchical data and a word of a higher concept for each of a plurality of keywords included in the hierarchical data.

[0057] The keyword selection unit 14 outputs the content information and the clustered keywords as knowledge data, and stores them in a memory (not shown).

[0058] (Keyword selection section 14) Next, a detailed description will be given of the keyword selection unit 14 shown in Fig. 2. Fig. 3 is a block diagram showing an example of the configuration of the keyword selection unit 14.

[0059] The keyword selection unit 14 includes an input processing unit 20, a memory 21, a clustering unit 22, and a keyword selection processing unit 23. The clustering unit 22 includes a candidate pair selection unit 30, a center of gravity position calculation unit 31, and a clustering processing unit 32.

[0060] The input processing unit 20 inputs content information from the input unit 10, reads from the memory 12 a score for each keyword along with a plurality of keywords having a score greater than 0, and inputs hierarchical data from the external data 2. In a first embodiment described below, a score for one frame for each keyword is read from the memory 12 along with the plurality of keywords, and in a second embodiment, a total score for N frames for each keyword is read along with the plurality of keywords.

[0061] The input processing unit 20 identifies each of the input keywords from all hierarchically structured words included in the input hierarchical data. Then, the input processing unit 20 constructs new hierarchical data consisting of the identified keywords and all superordinate words found by tracing upstream from the keywords (see, for example, Figs. 7 and 17 described later. In Figs. 7 and 17, words marked with * are keywords, and other words are superordinate words).

[0062] The input processing unit 20 stores in the memory 21 the content information, the plurality of keywords and the score for each keyword, as well as the newly constructed hierarchical data.

[0063] The clustering unit 22 reads out the hierarchical data, the keywords, and the scores for each keyword from the memory 21. Then, the clustering unit 22 performs a hierarchical clustering process for each of the keywords based on the score for each keyword using the hierarchical data. In the hierarchical clustering process, the clustering unit 22 deletes a keyword, creates a cluster, or deletes a cluster.

[0064] Here, a cluster refers to a set of two or more keywords. The clustering unit 22 performs hierarchical clustering processing to update the hierarchical data and generate a clustering result. As the clustering unit 22 updates the hierarchical data, it stores the updated hierarchical data in the memory 21. The clustering unit 22 outputs the clustering result to the keyword selection processing unit 23. In the clustering unit 22, the hierarchical clustering processing is performed by a candidate pair selection unit 30, a center of gravity position calculation unit 31, and a clustering processing unit 32.

[0065] The candidate pair selection unit 30 inputs hierarchical data (the original hierarchical data stored by the input processing unit 20, or the updated hierarchical data). Then, the candidate pair selection unit 30 calculates the distance in the hierarchical data between one or more keywords and one or more clusters included in the hierarchical data. This allows the distance between two keywords, the distance between a keyword and a cluster, and / or the distance between two clusters to be calculated.

[0066] Then, the candidate pair selection unit 30 selects a pair of nodes (keywords or clusters) with the shortest distance from among the calculated distances as a candidate pair, and outputs the candidate pair to the center of gravity position calculation unit 31. The candidate pair selected by the candidate pair selection unit 30 is either two keywords, one keyword and one cluster, or two clusters.

[0067] The center of gravity calculation unit 31 inputs the hierarchical data, a plurality of keywords, and the score for each keyword, as well as the candidate pair from the candidate pair selection unit 30. Then, based on the scores of each of the candidate pairs (of two nodes in the candidate pairs), the center of gravity calculation unit 31 calculates the center of gravity G of the candidate pair such that the candidate pairs are closer when the score is high and farther apart when the score is low. The center of gravity calculation unit 31 outputs the candidate pairs and the center of gravity G to the clustering processing unit 32.

[0068] The clustering processing unit 32 inputs the hierarchical data, and also inputs the candidate pairs and the centroid positions G from the centroid position calculation unit 31. The clustering processing unit 32 determines whether or not there is a bias in the centroid positions G of the candidate pairs, updates the hierarchical data depending on whether or not there is a bias in the centroid positions G, and stores the updated hierarchical data in the memory 21.

[0069] When the hierarchical data is updated so that the number of keywords and clusters contained therein becomes one, the clustering processing unit 32 uses this word (containing one keyword or one cluster) to generate a clustering result and outputs the clustering result to the keyword selection processing unit 23.

[0070] The keyword selection processing unit 23 receives the clustering result from the clustering unit 22, as well as a preset limit number, and further reads out the content information from the memory 21. Then, the keyword selection processing unit 23 selects one or more clustered keywords from the clustering result according to the limit number.

[0071] The keyword selection processing unit 23 adds information for specifying the one frame in the first embodiment, and information for specifying the Nth frame in the second embodiment, to the content information.

[0072] A detailed description will now be given of the processing performed by the keyword selection unit 14. FIG 4 is a flowchart showing an example of processing performed by the keyword selection unit 14.

[0073] The input processing unit 20 inputs content information from the input unit 10 (step S401), reads out from the memory 12 a number of keywords having a score greater than 0 as well as the score for each keyword (step S402), and inputs hierarchical data from the external data 2 (step S403).

[0074] The clustering unit 22 reads out the hierarchical data, as well as the multiple keywords and the scores for each keyword from the memory 21, and uses the hierarchical data to perform hierarchical clustering processing for each of the multiple keywords based on the score for each keyword, thereby updating the hierarchical data and generating a clustering result (step S404).

[0075] The keyword selection processing unit 23 receives the clustering result and a preset limit number from the clustering unit 22, and selects one or more post-clustering keywords from the clustering result according to the limit number (step S405).

[0076] The keyword selection unit 14 outputs the content information and the clustered keywords as knowledge data (step S406).

[0077] (Hierarchical clustering process: step S404) Next, details of the hierarchical clustering process in step S404 in Fig. 4 will be described. Fig. 5 is a flowchart showing an example of the hierarchical clustering process (step S404). As described above, the hierarchical clustering process is performed by the candidate pair selection unit 30, the center of gravity position calculation unit 31, and the clustering processing unit 32 of the clustering unit 22 shown in Fig. 3.

[0078] The clustering unit 22 reads out the hierarchical data, a plurality of keywords, and a score for each keyword from the memory 21 (step S501). Then, in steps S502 to S509 described later, the clustering unit 22 performs a hierarchical clustering process of deleting a keyword, generating a cluster, or deleting a cluster based on the score for each keyword until the number of keywords and clusters included in the hierarchical data is reduced to one, updates the hierarchical data, and generates and outputs a clustering result.

[0079] The clustering processor 32 determines whether the number (total number) of keywords and clusters included in the hierarchical data is one (step S502).

[0080] If the clustering processor 32 determines that the number of keywords and clusters included in the hierarchical data is not one (two or more) (step S502: N), the process proceeds to step S503. On the other hand, if the clustering processor 32 determines that the number of keywords and clusters included in the hierarchical data is one (step S502: Y), the process proceeds to step S509.

[0081] The candidate pair selection unit 30 proceeds from step S502(N) and inputs hierarchical data (the original hierarchical data stored by the input processing unit 20, or the updated hierarchical data). Then, the candidate pair selection unit 30 calculates the distances in the hierarchical data between one or more keywords and one or more clusters included in the hierarchical data (step S503).

[0082] Then, the candidate pair selecting unit 30 selects, as a candidate pair, the node pair with the shortest distance from among the distances calculated in step S503 (step S504).

[0083] The center of gravity calculation unit 31 inputs the candidate pair from the candidate pair selection unit 30, and calculates the center of gravity G of the candidate pair based on the scores of each of the candidate pairs (two nodes in the candidate pair) so that the candidate pairs are closer when the score is high and farther apart when the score is low (step S505).

[0084] The clustering processing unit 32 inputs the candidate pairs and the center positions G from the center position calculation unit 31, and determines whether there is a bias in the center positions G of the candidate pairs based on the distance between one of the candidate pairs and the center position G and the distance between the other of the candidate pairs and the center position G (step S506).

[0085] If the clustering processing unit 32 determines in step S506 that there is a bias in the center of gravity G (step S506: Y), it deletes from the hierarchical data the candidate (keyword or node) far from the center of gravity G among the candidate pairs (step S507), and proceeds to step S502. If there is a bias in the center of gravity G, the candidate that exists far from the center of gravity G is considered to be inappropriate as a keyword for the broadcast program content to be actually selected, and is therefore deleted. In this case, the clustering processing unit 32 updates the hierarchical data, and stores the updated hierarchical data in the memory 21.

[0086] On the other hand, if the clustering processing unit 32 determines in step S506 that there is no bias in the center of gravity positions G (step S506: N), it clusters the candidate pairs to generate new clusters (step S508) and proceeds to step S502. In this case as well, the clustering processing unit 32 updates the hierarchical data and stores the updated hierarchical data in the memory 21. The hierarchical data stored in the memory 21 in steps S507 and S508 is used for the processing of the next step S503, etc.

[0087] The hierarchical data update process carried out in steps S503 to S508 is repeated until the number of keywords and clusters is reduced to one in step S502.

[0088] The clustering processing unit 32 proceeds from step S502(Y) to generate a clustering result using the words (including one keyword or one cluster) contained in the updated hierarchical data, and outputs the clustering result to the keyword selection processing unit 23 (step S509).

[0089] Then, as described above, the keyword selection processing unit 23 selects post-clustering keywords from the clustering results according to a preset limit number.

[0090] Example 1 Next, a description will be given of the processing of the embodiment 1 by the knowledge data generation device 1 shown in Fig. 2. In the embodiment 1, for video images of broadcast program contents, a score for one frame is obtained for each identical keyword from multimodal data, keywords according to an arbitrary limited number are selected by clustering of hierarchical data, and knowledge data is generated.

[0091] As described above, in the knowledge data generation device 1, the score calculation unit 13 stores the score for one frame for each identical keyword in the memory 12. As a result, the memory 12 stores the score for one frame for each identical keyword (the total score for each process for each keyword) as shown in FIG.

[0092] 6 is a diagram illustrating an example of score calculation by the score calculation unit 13 in the embodiment 1. It is assumed that a weight of 0.5 is preset for the voice recognition process, a weight of 0.8 is preset for the object recognition process from image data, a weight of 0.7 is preset for the character recognition process from image data, and a weight of 0.9 is preset for the subtitle analysis process.

[0093] It is also assumed that the audio processing unit 111 generates one "snowboard" keyword, and the image object processing unit 112 generates two "ski", one "skateboard" and one "table tennis" keywords. For example, if the image data shows two people skiing, the image object processing unit 112 will find two "ski" keywords. It is also assumed that the image text processing unit 113 generates one "ski" keyword, and the subtitle processing unit 114 generates one "ski" keyword.

[0094] In this case, the score calculation unit 13 multiplies the number of keywords obtained by the object recognition processing from the image data for the keyword "skiing" (2) by the weight of the object recognition processing from the image data (0.8), thereby obtaining a score of 1.6 for the object recognition processing from the image data. The score calculation unit 13 also multiplies the number of keywords obtained by the character recognition processing, etc. from the image data for the keyword "skiing" (1) by the weight of the character recognition processing, etc. from the image data (0.7), thereby obtaining a score of 0.7 for the character recognition processing, etc. from the image data. The score calculation unit 13 also multiplies the number of keywords obtained by the character analysis processing for the keyword "skiing" (1) by the weight of the character analysis processing (0.9), thereby obtaining a score of 0.9 for the character analysis processing.

[0095] Then, the score calculation unit 13 adds up these scores 1.6, 0.7, and 0.9 to obtain a score of 3.2 for one frame for the keyword "skiing."

[0096] Similarly, the score calculation unit 13 obtains a score of 0.5 per frame for the keyword "snowboarding", and a score of 0.8 per frame for the keywords "skateboarding" and "table tennis".

[0097] For keywords such as "ball", the score for one frame is 0. These keywords are preset in the memory 12.

[0098] Hereinafter, a description will be given of the processing of the keyword selection unit 14 shown in Fig. 3. In the keyword selection unit 14, it is assumed that the memory 21 stores content information, a plurality of keywords (a plurality of keywords having a score greater than 0) shown in Fig. 6, a score for one frame for each keyword, and hierarchical data shown in Fig. 7, which will be described later.

[0099] Fig. 7 is a diagram showing an example of hierarchical data in Example 1. This hierarchical data is a hierarchically structured data group consisting of a plurality of keywords (a plurality of keywords obtained from voice data or the like) identified in the original hierarchical data by the input processing unit 20 shown in Fig. 3, and all superordinate words of the keywords when tracing back upstream.

[0100] The multiple keywords are "table tennis," "skiing," "snowboarding," and "skateboarding," which are indicated with an *. The higher-level word for "table tennis" is "ball games," the higher-level word for "skiing" and "snowboarding" is "winter sports," and the higher-level word for "skateboarding" is "roller sports." The higher-level word for these higher-level words is "sports."

[0101] In other words, this hierarchical data is a hierarchically structured data group ranging from the word "sports" which is the most upstream concept to a plurality of downstream keywords such as "table tennis".

[0102] The hierarchical clustering process is performed using the hierarchical data shown in Fig. 7, with the edges between words containing a keyword as the distance and the score for each keyword as the weight. For convenience of explanation, the weight of the edge is set to 1, but other values ​​may be used.

[0103] First, the candidate pair selection unit 30 calculates the distance between two keywords, "table tennis," "skiing," "snowboarding," and "skateboarding," for a plurality of keywords included in the hierarchical data.

[0104] Fig. 8 is a diagram showing the distance between keywords at the start in Example 1. For example, the distance between "skiing" and "snowboarding" is 2. This is because, referring to Fig. 7, there are two edges between "skiing" and "snowboarding" via their superordinate concept "winter sports".

[0105] After calculating the distance between the two keywords shown in FIG. 8, the candidate pair selection unit 30 identifies the shortest distance among these distances (2 in the case of FIG. 8), and selects the two keywords "ski" and "snowboard" at both ends of the shortest distance as a candidate pair.

[0106] When there are multiple candidate pairs with the shortest distance, the candidate pair selection unit 30 sums up the weights, which are the scores of the two keywords that make up the candidate pair, identifies the candidate pair with the smallest sum, and selects the candidate pair with the smallest distance. When there are multiple candidate pairs with the smallest sum, the candidate pair selection unit 30 randomly selects the candidate pair with the shortest distance from among them.

[0107] Next, the center of gravity position calculation unit 31 calculates the center of gravity position G1 for the candidate pair "ski" and "snowboard" based on the weights, which are their scores. Specifically, the center of gravity position calculation unit 31 calculates the center of gravity position G1 so that the larger the weight, the closer the center of gravity position G1 is, and the smaller the weight, the farther the center of gravity position G1 is.

[0108] FIG. 9 is a diagram showing the center of gravity position G1 between "skiing" and "snowboarding". A node with a double frame indicates a candidate pair. From FIG. 6, the weight of "skiing" is 3.2, and the weight of "snowboarding" is 0.5. Therefore, as shown in FIG. 9, the center of gravity position calculation unit 31 calculates the center of gravity position G1 to be 0.27 away from the position of "skiing" and 1.73 away from the position of "snowboarding" via "winter sports".

[0109] Next, the clustering processing unit 32 determines whether or not there is a bias for the candidate pair of "skis" and "snowboard" based on the distance 0.27 between the position of "skis" and the center of gravity position G1, and the distance 1.73 between the position of "snowboard" and the center of gravity position G1.

[0110] If the clustering processing unit 32 determines that there is a bias, it deletes candidates far from the center of gravity position G1. On the other hand, if the clustering processing unit 32 determines that there is no bias, it clusters candidate pairs without deleting candidates and generates new clusters. Then, the clustering processing unit 32 updates the hierarchical data.

[0111] For example, if the longer of the two distances is equal to or greater than a first preset value (e.g., 2) and the shorter of the two distances is equal to or less than a second preset value (e.g., 1), the clustering processing unit 32 determines that there is a bias and deletes the candidate that is far from the center of gravity position G1. On the other hand, if the longer distance is less than the first preset value or the shorter distance exceeds the second preset value, the clustering processing unit 32 determines that there is no bias and clusters the candidate pair.

[0112] In the above example, the clustering processing unit 32 determines that there is no bias because the distance 1.73 between the longer "snowboard" and the center of gravity position G1 is less than the first set value of 2, and clusters "ski" and "snowboard", generating a cluster C1 consisting of "ski" and "snowboard". Then, the clustering processing unit 32 updates the hierarchical data shown in Fig. 7 to hierarchical data including the cluster C1 consisting of "ski" and "snowboard".

[0113] FIG. 10 is a diagram showing an example of hierarchical data after clustering of "skiing" and "snowboarding," showing the updated hierarchical data.

[0114] As shown in Fig. 10, the keyword "table tennis", cluster C1 ("skiing" and "snowboarding"), and keyword "skateboarding" are located at the bottom, and the superordinate word "sports" is located at the top. This hierarchical data is a hierarchically structured data group consisting of these keywords and cluster C1, and all superordinate words such as "ball games" that are found when tracing upstream from these keywords and cluster C1.

[0115] The center of gravity G1 of the cluster C1 is the same as the center of gravity G1 shown in FIG. 9, and is located at a distance of 0.27 from the position of "skiing" and at a distance of 0.73 from the position of "winter sports" toward the position of "skiing".

[0116] Then, as described below, the same processes as those performed by the candidate pair selection unit 30, the center of gravity position calculation unit 31, and the clustering processing unit 32 described above are repeated until the number of keywords and clusters included in the hierarchical data is reduced to 1. In this case, the candidate pair selection unit 30 uses the center of gravity position G1 to calculate the distance of the cluster C1.

[0117] Next, the candidate pair selection unit 30 calculates the distance between the two keywords "table tennis" and "skateboard" included in the updated hierarchical data shown in FIG. 10 and the cluster C1.

[0118] Fig. 11 is a diagram showing the updated keywords and distances in cluster C1. For example, the distance between cluster C1 and "skateboarding" is 3.73. This is because, with reference to Fig. 10, the center of gravity position G1 of cluster C1 is used to calculate the distance, and the distance between cluster C1 and "winter sports" is 0.73, and the distance between "winter sports" and "skateboarding" is 3, totaling 3.73. The distance between cluster C1 and "table tennis" is also 3.73.

[0119] After calculating the distances between the two keywords and the cluster C1 shown in FIG. 11, the candidate pair selection unit 30 identifies the shortest distance (3.73 in the case of FIG. 11) among these distances, and determines that the two candidate pairs located at both ends of the shortest distance are the cluster C1 and "skateboard" and the cluster C1 and "table tennis". Then, since the total weight value of both candidate pairs is the same at 4.5 (=3.2+0.5+0.8), the candidate pair selection unit 30 cannot select a candidate pair based on the total weight value. Therefore, the candidate pair selection unit 30 randomly selects a candidate pair from the two candidate pairs. In this case, the candidate pair selection unit 30 selects the cluster C1 and "skateboard" as a candidate pair.

[0120] Next, the center of gravity position calculation unit 31 calculates the center of gravity position G2 for the candidate pair of cluster C1 and "skateboard" based on the weights, which are their scores. Specifically, the center of gravity position calculation unit 31 calculates the center of gravity position G2 between the cluster C1 and the "skateboard" between the center of gravity position G1 (see FIG. 10) of the cluster C1 and the position of the "skateboard" based on the weight of the cluster C1 (the total value of the weights of "ski" and "snowboard" in the cluster C1) and the weight of the "skateboard".

[0121] FIG. 12 is a diagram showing the center of gravity G2 between the cluster C1 and "skateboard". The double frame in the cluster C1 and the "skateboard" indicates that they are a candidate pair. From FIG. 6, the weight of the cluster C1 (the total value of the weights of "ski" and "snowboard" in the cluster C1) is 3.7, and the weight of the "skateboard" is 0.8. Therefore, as shown in FIG. 12, the center of gravity G2 between the cluster C1 and the "skateboard" is calculated based on the distance 3.73 (see FIG. 10) between the center of gravity G1 of the cluster C1 and the position of the "skateboard" according to the weight of the cluster C1, 3.7, and the weight of the "skateboard", 0.8, such that the larger the weight, the closer the center of gravity G2 is, and the smaller the weight, the farther the center of gravity G2 is.

[0122] In other words, the center of gravity G2 between cluster C1 and "skateboard" is 0.67 from the center of gravity G1 of cluster C1 (0.94 from the position of "ski") and 3.06 away from the position of "skateboard" via "roller sports", "sports" and "winter sports". This center of gravity G2 is the center of gravity of the cluster when cluster C1 and "skateboard" are clustered.

[0123] Next, the clustering processing unit 32 determines whether or not there is a bias for the candidate pair of cluster C1 and "skateboard" based on the distance 0.67 between the center of gravity G1 of cluster C1 and the center of gravity G2 when cluster C1 and "skateboard" are clustered, and the distance 3.06 between the position of "skateboard" and the center of gravity G2.

[0124] Fig. 13 is a diagram for explaining the process of deleting "skateboard" when it is determined that there is a bias. The clustering processing unit 32 determines that there is a bias because the longer of the two distances (between the position of "skateboard" and the center of gravity position G2) is a distance of 3.06 that is equal to or greater than a first preset value (e.g., 2), and the shorter of the two distances (between the center of gravity position G1 of cluster C1 and the center of gravity position G2) is equal to or less than a second preset value (e.g., 1). Then, as shown in Fig. 13, the clustering processing unit 32 deletes "skateboard", which is a candidate far from the center of gravity position G2, and updates the hierarchical data.

[0125] FIG. 14 is a diagram showing an example of hierarchical data after "skateboard" has been deleted, showing the updated hierarchical data.

[0126] As shown in Fig. 14, the keyword "table tennis" and cluster C1 ("skiing" and "snowboarding") are located at the bottom, and the superordinate word "sports" is located at the top. This hierarchical data is a hierarchically structured data group consisting of the keyword "table tennis" and cluster C1, and all superordinate words such as "ball games" that are found when tracing upstream from these.

[0127] Next, the keyword "table tennis" and cluster C1 included in the updated hierarchical data shown in Fig. 14 are processed by the candidate pair selection unit 30, the center of gravity position calculation unit 31, and the clustering processing unit 32, respectively. In other words, since it is determined that there is no bias in "table tennis" and cluster C1, clustering is not performed, "table tennis" is deleted, and the hierarchical data is updated.

[0128] In the updated hierarchical data, cluster C1 ("skiing" and "snowboarding") is located at the bottom, and the superordinate word "sports" is located at the top. This hierarchical data is a hierarchically structured data group consisting of cluster C1 and all superordinate words "winter sports" and "sports" found when tracing upstream from cluster C1.

[0129] When the clustering processing unit 32 determines that the number of keywords and clusters has been reduced to one, it ends the repetitive processing. Then, the clustering processing unit 32 sets a keyword or word of a higher concept at the meeting point in the updated hierarchical data. That is, the clustering processing unit 32 sets "winter sports" at the meeting point, which is a higher concept of "skiing" and "snowboarding", as a higher concept word from the updated hierarchical data consisting of the keywords "skiing" and "snowboarding" constituting cluster C1, their higher concept word "winter sports", and the highest higher concept word "sports".

[0130] The clustering processing unit 32 generates, as a clustering result, a data group consisting of the keywords "skiing" and "snowboarding" and the superordinate word "winter sports."

[0131] 15 is a diagram showing the final clustering result in Example 1. This clustering result shows the relationship between keywords and superordinate and subordinate concepts of words, and shows a data group consisting of the keywords "skiing" and "snowboarding" and the superordinate word "winter sports".

[0132] The clustering result thus generated is used when selecting keywords according to an arbitrary limit number in the keyword selection processing unit 23. For example, the keyword selection processing unit 23 selects keywords with a preset limit number as the number of clusters.

[0133] When the limit number is 1, the keyword selection processing unit 23 selects the superordinate concept "winter sports" from among "skiing," "snowboarding," and "winter sports" in the clustering result shown in Fig. 15. When the limit number is 2 or more, the keyword selection processing unit 23 selects "skiing" and "snowboarding" as keywords. In this case, the superordinate concept "winter sports" is not selected as a keyword.

[0134] When the limit number is three, the keyword selection processing unit 23 may select, in addition to "skiing" and "snowboarding," the superordinate concept "winter sports" as a keyword.

[0135] Example 2 Next, a description will be given of the process of Example 2 by the knowledge data generation device 1 shown in Fig. 2. Example 2 obtains the score of N frames for each identical keyword from multimodal data for video images of broadcast program contents, selects keywords according to an arbitrary limit number by clustering the hierarchical data, and generates knowledge data.

[0136] As described above, in the knowledge data generation device 1, the score calculation unit 13 calculates a score for N frames (N consecutive frames) by adding up the scores of N consecutive frames set in advance for each of the same keywords, and stores the score in the memory 12. N is an integer equal to or greater than 2. As a result, the memory 12 stores the scores for N frames for each of the same keywords (the sum of the scores for each process for each keyword for N consecutive frames) as shown in FIG.

[0137] Fig. 16 is a diagram for explaining an example of score calculation by the score calculation unit 13 in the second embodiment. Fig. 16 shows the scores of one frame for each of the same keywords for the (nx)th, (n-3)th, (n-2)th, (n-1)th, nth (current) and other frames.

[0138] For example, for the keyword "smartphone", the score of the (nx)th frame is 2 and the score of the nth frame is 0.5. For the keyword "accelerometer", the score of the (n-3)th frame is 2 and the score of the (n-2)th frame is 1.

[0139] For example, for the keyword "smartphone", the score calculation unit 13 adds up the scores of N frames to obtain a total score of 10.5 for the N frames. Similarly, for the keyword "acceleration sensor", the score calculation unit 13 adds up the scores of N frames to obtain a total score of 33.2 for the N frames. For "acceleration" and "gravity", the total scores of 49.8 and 36.5 are obtained, respectively. For the keyword "photo", the total score for N frames is 0.

[0140] Hereinafter, a description will be given of the processing of the keyword selection unit 14 shown in Fig. 3. In the keyword selection unit 14, it is assumed that the memory 21 stores content information, a plurality of keywords (a plurality of keywords having a score greater than 0) shown in Fig. 16, scores of N frames for each keyword, and hierarchical data shown in Fig. 17, which will be described later.

[0141] Fig. 17 is a diagram showing an example of hierarchical data in Example 2. As in Fig. 7, this hierarchical data is a hierarchically structured data group consisting of a plurality of keywords specified in the original hierarchical data by the input processing unit 20 shown in Fig. 3 and all superordinate words of the keywords when tracing upstream.

[0142] The multiple keywords are "smartphone," "acceleration sensor," "acceleration," and "gravity," which are indicated with an *. Furthermore, the higher-level word for "smartphone" is "home appliances," the higher-level words for "acceleration sensor" are "sensor" and "acceleration," the higher-level word for "gravity" is "acceleration," and the higher-level word for "acceleration" is "mechanics." Furthermore, for these higher-level words, the most upstream common word for the higher-level concepts is "physics."

[0143] In other words, this hierarchical data is a hierarchically structured data group ranging from the word "physics" which is the most upstream concept, to multiple downstream keywords such as "smartphone."

[0144] The hierarchical clustering process is performed using the hierarchical data shown in Fig. 17, with the edges between words containing a keyword as the distance and the keyword score as the weight. For convenience of explanation, the edge weight is set to 1, but other values ​​may be used.

[0145] First, the candidate pair selecting unit 30 calculates the distance between two keywords, "smartphone," "acceleration sensor," "acceleration," and "gravity," for a plurality of keywords included in the hierarchical data.

[0146] Here, when clustering keywords, etc., the word of the higher concept in the merging hierarchy is used. Therefore, when the candidate pair selection unit 30 calculates the distance between two keywords, the number of round trips of the directed edge is set to within one round trip. For example, in FIG. 17, regarding the distance between "smartphone" and "acceleration", the route of "acceleration" from "smartphone" via "home appliance", "electricity", "physics", and "mechanics" is included, but the route of "acceleration" from "smartphone" via "home appliance", "tool", "part", "sensor", and "acceleration sensor" is not included.

[0147] 18 is a diagram showing distances between keywords at the start in Example 2. For example, the distance between "smartphone" and "acceleration sensor" is 5, and the distance between "smartphone" and "acceleration" is also 5.

[0148] After calculating the distance between the two keywords shown in Fig. 18, the candidate pair selection unit 30 identifies the shortest distance among these distances (1 in the case of Fig. 18) and determines that the candidate pairs located at both ends of the shortest distance are two pairs of "acceleration sensor" and "acceleration", and "acceleration" and "gravity". The candidate pair selection unit 30 then calculates the total weight of "acceleration sensor" and "acceleration" to be 83.0, and the total weight of "acceleration" and "gravity" to be 86.3, and selects "acceleration sensor" and "acceleration", which has the smallest total weight value, as the candidate pair.

[0149] As in the first embodiment, when there are multiple candidate pairs with the shortest distance, the candidate pair selection unit 30 sums up the weights of the two keywords that make up the candidate pair, identifies the candidate pair with the smallest sum, and selects the candidate pair with the smallest distance. When there are multiple candidate pairs with the smallest sum, the candidate pair selection unit 30 randomly selects the candidate pair with the shortest distance from among them.

[0150] Next, the center of gravity position calculation unit 31 calculates the center of gravity position G3 for the candidate pair of "acceleration sensor" and "acceleration" based on the weights, which are their scores. Specifically, the center of gravity position calculation unit 31 calculates the center of gravity position G3 based on the weights of the "acceleration sensor" and the weights of the "acceleration" such that the larger the weights, the closer the center of gravity position G3 is, and the smaller the weights, the farther the center of gravity position G3 is.

[0151] FIG. 19 is a diagram showing the center of gravity position G3 between "acceleration" and "acceleration sensor". A node with a double frame indicates a candidate pair. From FIG. 16, the weight of "acceleration" is 49.8, and the weight of "acceleration sensor" is 33.2. Therefore, as shown in FIG. 19, the center of gravity position calculation unit 31 calculates the center of gravity position G3 to be 0.4 away from the position of "acceleration" and 0.6 away from the position of "acceleration sensor".

[0152] Next, the clustering processing unit 32 determines whether or not there is a bias for the candidate pair of “acceleration” and “acceleration sensor” based on the distance 0.4 between the position of “acceleration” and the center of gravity position G3, and the distance 0.6 between the position of “acceleration sensor” and the center of gravity position G3.

[0153] As in the first embodiment, if the clustering processing unit 32 determines that there is a bias, it deletes candidates far from the center of gravity position G3. On the other hand, if the clustering processing unit 32 determines that there is no bias, it clusters the candidate pairs without deleting the candidates, and generates a new cluster. Then, the clustering processing unit 32 updates the hierarchical data.

[0154] For example, if the longer of the two distances is equal to or greater than a preset first set value (e.g., 2) and the shorter of the two distances is equal to or less than a preset second set value (e.g., 1), the clustering processing unit 32 determines that there is a bias and deletes the candidate that is far from the center of gravity position G3. On the other hand, if the longer distance is less than the first set value or the shorter distance exceeds the second set value, the clustering processing unit 32 determines that there is no bias and clusters the candidate pair.

[0155] In the above example, the clustering processing unit 32 determines that there is no bias because the distance 0.6 between the longer "acceleration sensor" and the center of gravity position G3 is less than the first set value 2, and clusters "acceleration" and "acceleration sensor", generating a cluster C2 consisting of "acceleration" and "acceleration sensor". Then, the clustering processing unit 32 updates the hierarchical data shown in Fig. 17 to hierarchical data including cluster C2 consisting of "acceleration" and "acceleration sensor".

[0156] FIG. 20 is a diagram showing an example of hierarchical data after clustering of "acceleration" and "acceleration sensor", showing the updated hierarchical data.

[0157] As shown in Fig. 20, the keyword "smartphone", cluster C2 ("acceleration" and "acceleration sensor"), and keyword "gravity" are located downstream, and the superordinate word "physics" is located at the top. This hierarchical data is a hierarchically structured data group consisting of these keywords and cluster C2, and all superordinate words such as "home appliances" that are found when tracing upstream from these keywords and cluster C2.

[0158] The center of gravity position G3 of the cluster C2 is the same as the center of gravity position G3 shown in FIG. 19, and is located at a distance of 0.6 from the "acceleration sensor" and at a distance of 0.4 from the position of "acceleration".

[0159] Then, as described below, the same processes as those performed by the candidate pair selection unit 30, the center of gravity position calculation unit 31, and the clustering processing unit 32 described above are repeated until the number of keywords and clusters included in the hierarchical data is reduced to 1. In this case, the candidate pair selection unit 30 uses the center of gravity position G3 to calculate the distance of cluster C2.

[0160] Next, the candidate pair selecting unit 30 calculates the distance between the multiple keywords "smartphone" and "gravity" included in the updated hierarchical data shown in FIG. 20 and the cluster C2.

[0161] Fig. 21 is a diagram showing the updated keywords and distances in cluster C2. For example, the distance between "smartphone" and cluster C2 is 5.4. This is because, referring to Fig. 20, the center of gravity position G3 of cluster C2 is used to calculate the distance, the distance between "smartphone" and "mechanics" is 4, and the distance between "mechanics" and cluster C2 is 1.4, resulting in a total of 5.4.

[0162] After calculating the distance between the two keywords and cluster C2 shown in FIG. 21, the candidate pair selection unit 30 identifies the shortest distance among these distances (1.4 in the case of FIG. 21), and selects cluster C2 and “gravity”, which are located at both ends of the shortest distance, as a candidate pair.

[0163] Next, the center of gravity calculation unit 31 calculates a center of gravity G4 between the cluster C2 and "gravity", which are candidate pairs, based on the weights, which are their scores. Specifically, the center of gravity calculation unit 31 calculates a center of gravity G4 between the cluster C2 and "gravity", between the center of gravity G3 of the cluster C2 (see FIG. 20) and the position of "gravity", based on the weight of the cluster C2 (the sum of the weights of "acceleration" and "acceleration sensor") and the weight of "gravity".

[0164] FIG. 22 is a diagram showing the center of gravity G4 between cluster C2 and "gravity". The double frame in cluster C2 and "gravity" indicates that they are a candidate pair. From FIG. 16, the weight of cluster C2 (the total value of the weights of "acceleration" and "acceleration sensor") is 83, and the weight of "gravity" is 36.5. Therefore, as shown in FIG. 22, the center of gravity G4 between cluster C2 and "gravity" is calculated based on the distance 1.4 (see FIG. 20) between the center of gravity G3 of cluster C2 and the position of "gravity" according to the weight of cluster C2, 83, and the weight of "gravity", 36.5, such that the larger the weight, the closer the center of gravity G4 is, and the smaller the weight, the farther the center of gravity G4 is.

[0165] That is, the center of gravity G4 between cluster C2 and "gravity" is 0.43 away from the center of gravity G3 of cluster C2 via "acceleration" (0.03 away from the position of "acceleration"), and 0.97 away from the position of "gravity." This center of gravity G4 is the center of gravity of the cluster when cluster C2 and "gravity" are clustered.

[0166] Next, the clustering processing unit 32 determines whether or not there is a bias for the candidate pair of cluster C2 and "gravity" based on the distance 0.43 between the center of gravity G3 of cluster C2 and the center of gravity G4 when cluster C2 and "gravity" are clustered, and the distance 0.97 between the position of "gravity" and the center of gravity G4.

[0167] The clustering processing unit 32 determines that there is no bias because the distance 0.97 between "gravity", which is the longer distance, and the center of gravity position G4 is less than the first set value of 2, and clusters cluster C2 and "gravity", generating cluster C3 consisting of "acceleration", "acceleration sensor", and "gravity". Then, the clustering processing unit 32 updates the hierarchical data shown in Fig. 20 to hierarchical data including cluster C3 consisting of "acceleration", "acceleration sensor", and "gravity".

[0168] FIG. 23 is a diagram showing an example of hierarchical data after clustering of "acceleration", "acceleration sensor", and "gravity", showing the updated hierarchical data.

[0169] 23, the keyword "smartphone" and cluster C3 ("acceleration", "acceleration sensor", and "gravity") are located at the bottom, and the superordinate word "physics" is located at the top. This hierarchical data is a hierarchically structured data group consisting of the keyword "smartphone" and cluster C3, and all superordinate words such as "home appliances" that are found when tracing upstream from these keywords and cluster C3.

[0170] The center of gravity position G4 of the cluster C3 is the same as the center of gravity position G4 shown in FIG. 22, and is located at a distance of 0.03 from the position of "acceleration" and at a distance of 0.97 from the position of "gravity".

[0171] Next, the candidate pair selecting unit 30 calculates the distance between the keyword "smartphone" and the cluster C3 included in the updated hierarchical data shown in FIG.

[0172] Fig. 24 is a diagram showing the updated keywords and distances in cluster C3. The distance between "smartphone" and cluster C3 is 5.03. This is because, referring to Fig. 23, the center of gravity position G4 of cluster C3 is used to calculate the distance, the distance between "smartphone" and "mechanics" is 4, and the distance between "mechanics" and cluster C3 is 1.03, for a total of 5.03.

[0173] After calculating the distance between the keyword “smartphone” and cluster C3 shown in FIG. 24, the candidate pair selection unit 30 determines that this distance is the shortest distance and selects “smartphone” and cluster C3, which are located at both ends of the shortest distance, as a candidate pair.

[0174] Next, the center of gravity position calculation unit 31 calculates the center of gravity position G5 between the candidate pair "smartphone" and cluster C3 based on the weights, which are their scores. Specifically, the center of gravity position calculation unit 31 calculates the center of gravity position G5 between the position of "smartphone" and the center of gravity position G4 of cluster C3 (see FIG. 23) based on the weight of "smartphone" and the weight of cluster C3 (the total value of the weights of "acceleration", "acceleration sensor", and "gravity").

[0175] FIG. 25 is a diagram showing the center of gravity G5 between "smartphone" and cluster C3. The double frame in "smartphone" and cluster C3 indicates that they are a candidate pair. From FIG. 16, the weight of "smartphone" is 10.5, and the weight of cluster C3 (the total value of the weights of "acceleration", "acceleration sensor" and "gravity") is 119.5. As shown in FIG. 25, the center of gravity G5 between "smartphone" and cluster C3 is calculated based on the distance 5.03 (see FIG. 23) between the position of "smartphone" and the center of gravity G4 of cluster C3 according to the weight of "smartphone" 10.5 and the weight of cluster C3 119.5, so that the larger the weight, the closer the center of gravity G5 is, and the smaller the weight, the farther the center of gravity G5 is.

[0176] In other words, the center of gravity G5 between "smartphone" and cluster C3 is 4.63 away from the position of "smartphone" via "home appliances", "electricity", "physics" and "mechanics" (0.63 from the position of "mechanics"), and 0.4 away from the center of gravity G4 of cluster C3 via "acceleration" (0.37 from the position of "acceleration"). This center of gravity G5 is the center of gravity of the cluster when "smartphone" and cluster C3 are clustered.

[0177] Next, the clustering processing unit 32 determines whether or not there is a bias for the candidate pair of “smartphone” and cluster C3 based on the distance of 4.63 between the position of “smartphone” and the center of gravity position G5, and the distance of 0.4 between the center of gravity position G4 of cluster C3 and the center of gravity position G5.

[0178] FIG. 26 is a diagram for explaining the process of deleting "smartphone" when it is determined that there is a bias. For example, the clustering processing unit 32 determines that there is a bias because the longer of the two distances (between the position of "smartphone" and the center of gravity position G5) is a distance of 4.63 that is equal to or greater than a first preset value (e.g., 2), and the shorter of the two distances (between the center of gravity position G4 of cluster C3 and the center of gravity position G5) is equal to or less than a second preset value (e.g., 1). Then, as shown in FIG. 26, the clustering processing unit 32 deletes "smartphone", which is a candidate far from the center of gravity position G5, and updates the hierarchical data.

[0179] As shown in Fig. 26, cluster C3 (consisting of "acceleration", "acceleration sensor", and "gravity", with "acceleration" upstream and "acceleration sensor" and "gravity" downstream) is located at the most downstream, and the superordinate word "physics" is located at the most upstream. This hierarchical data is a hierarchically structured data group consisting of cluster C3 and all superordinate word "mechanics" etc. when tracing upstream from cluster C3.

[0180] When the clustering processing unit 32 determines that the number of keywords and clusters has been reduced to 1, it ends the repetitive processing. Then, the clustering processing unit 32 sets a keyword of a higher concept at the junction in the updated hierarchical data. That is, the clustering processing unit 32 sets "acceleration" as a keyword of a higher concept at the junction from the keywords "acceleration", "acceleration sensor", and "gravity" constituting cluster C3.

[0181] The clustering processing unit 32 generates a data group consisting of the higher-level keyword "acceleration" and the lower-level keywords "acceleration sensor" and "gravity" as a clustering result.

[0182] 27 is a diagram showing the final clustering result in Example 2. This clustering result shows the relationship between the superordinate and subordinate concepts of each keyword when the three keywords constituting cluster C3, "acceleration," "acceleration sensor," and "gravity," are positioned at the most downstream.

[0183] In cluster C3 shown in Fig. 26, the superordinate concept of "acceleration sensor" and "acceleration" is "acceleration", and the superordinate concept of "acceleration" and "gravity" is "acceleration". The clustering result shown in Fig. 27 shows the relationship between these superordinate and subordinate concepts. In other words, the clustering result is a data group in which the superordinate concept of "acceleration sensor" and "acceleration" is "acceleration", and the superordinate concept of "acceleration" and "gravity" is "acceleration".

[0184] The clustering result thus generated is used when selecting keywords according to an arbitrary limit number in the keyword selection processing unit 23. For example, the keyword selection processing unit 23 selects keywords with a preset limit number as the number of clusters.

[0185] When the limit number is 1, the keyword selection processing unit 23 selects the highest superordinate concept "acceleration" from the clustering result shown in Fig. 27 as a keyword. When the limit number is 2, the keyword selection processing unit 23 selects "acceleration" and "gravity" as keywords, and when the limit number is 3 or more, the keyword selection processing unit 23 selects "acceleration", "gravity", and "acceleration sensor" as keywords.

[0186] As described above, according to the knowledge data generation device 1 of the embodiment of the present invention, for the moving images of broadcast program content, a score is calculated for each identical keyword from audio data, etc., and by clustering the hierarchical data, keywords according to an arbitrary limit number are selected, and knowledge data is generated.

[0187] Specifically, the data processing unit 11 obtains keywords from the voice data or the like, and the score calculation unit 13 obtains a score for each of the same keywords.

[0188] The candidate pair selection unit 30 identifies multiple keywords obtained by the data processing unit 11 from the multiple words that make up the hierarchical data, calculates the distance in the hierarchical data between one or more keywords and one or more clusters, and selects the pair with the shortest distance as a candidate pair.

[0189] The center of gravity calculation unit 31 calculates the center of gravity G of the candidate pair based on the scores of each of the candidate pairs so that the centers are closer when the scores are high and farther apart when the scores are low.

[0190] The clustering processing unit 32 determines whether or not there is a bias in the center positions G of the candidate pairs, based on the distance between one of the candidate pairs and the center positions G and the distance between the other of the candidate pairs and the center positions G. When it is determined that there is a bias in the center positions G, the clustering processing unit 32 deletes the candidate far from the center positions G from the hierarchical data, and when it is determined that there is no bias in the center positions G, it clusters the candidate pairs to generate new clusters and updates the hierarchical data.

[0191] The candidate pair selection unit 30, the center of gravity position calculation unit 31, and the clustering processing unit 32 repeat the above-mentioned processes until the number of keywords and clusters is reduced to one.

[0192] When the repeated processing is completed, the clustering processing unit 32 generates a clustering result using the updated hierarchical data. The keyword selection processing unit 23 selects clustered keywords from the clustering result according to a preset limit number, and generates and outputs knowledge data.

[0193] As described above, the conventional methods described in Non-Patent Documents 2 and 3 extract objects from multimodal images and the like, list and score them, select the keyword with the highest score, and generate knowledge data.

[0194] The selected keywords may contain noise because they are obtained from multimodal images, etc. Furthermore, the number of selected keywords is limited, so that keywords with high accuracy may be excluded. For example, if there are two highly accurate keywords but only one keyword is selected, one of the two highly accurate keywords will be excluded. As a result, it may not be possible to generate effective knowledge data for presenting desired content to a user.

[0195] In contrast, in the embodiment of the present invention, hierarchical clustering processing is performed using hierarchical data in which multiple words are hierarchically structured from the perspective of meaning, and keywords after clustering are selected. Therefore, the selected keywords are keywords that are close in distance in the hierarchical structure (keywords with similar meanings), and the possibility of including noise can be reduced.

[0196] In addition, in the embodiment of the present invention, since hierarchical data is used, words of higher concepts included in the hierarchical data are selected as keywords according to the limited number. Therefore, comprehensive words including highly accurate keywords that were previously excluded are selected, and highly accurate keywords are not excluded.

[0197] Therefore, it is possible to realize the linkage of various contents and generate knowledge data for presenting highly accurate contents.

[0198] Although the present invention has been described above with reference to the embodiments, the present invention is not limited to the above-described embodiments and can be modified in various ways without departing from the technical concept thereof.

[0199] A normal computer can be used as the hardware configuration of the knowledge data generation device 1 according to the embodiment of the present invention. The knowledge data generation device 1 is configured by a computer equipped with a CPU, a volatile storage medium such as a RAM, a non-volatile storage medium such as a ROM, an interface, etc.

[0200] The functions of the input unit 10, data processing unit 11 (sound processing unit 111, image object processing unit 112, image text processing unit 113, and subtitle processing unit 114), memory 12, score calculation unit 13, and keyword selection unit 14 (input processing unit 20, memory 21, clustering unit 22 (candidate pair selection unit 30, center of gravity position calculation unit 31, and clustering processing unit 32)) and keyword selection processing unit 23 provided in the knowledge data generation device 1 are each realized by causing a CPU to execute a program describing these functions.

[0201] These programs are stored in the storage medium and are read and executed by the CPU. These programs can also be distributed by storing them in a storage medium such as a magnetic disk (floppy disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.), or a semiconductor memory, and can also be transmitted and received via a network. [Explanation of symbols]

[0202] 1. Knowledge data generation device 2,2-1,2-2,2-3 External Data 3,3-1,3-2 Broadcasting station data 4, 4-1, 4-2, 4-3 User terminal 10 Input section 11 Data Processing Section 111 Audio processing unit 112 Image object processing section 113 Image and text processing unit 114 Subtitle Processing Unit 12,21 Memory 13 Score Calculation Section 14 Keyword selection section 20 Input Processing Section 22 Clustering Department 23 Keyword selection processing section 30 Candidate pair selection section 31 Center of gravity position calculation section 32 Clustering Processing Section G,G1,G2,G3,G4,G5 Center of gravity position C1, C2, C3 Cluster

Claims

1. A knowledge data generating device for generating knowledge data to be used when searching for desired content by obtaining keywords related to the content, comprising: The content is composed of multiple types of data, a data processing unit that detects a character string of a sentence for each of the plurality of types of data by a predetermined process corresponding to the data, and obtains the keyword from the character string of the sentence; a score calculation unit that scores the keywords obtained by the data processing unit for the plurality of types of data, and calculates a score obtained by integrating all of the predetermined processes for each same keyword; a keyword selection unit that performs a hierarchical clustering process using hierarchical data composed of a plurality of hierarchically structured words and including a plurality of keywords obtained by the data processing unit, for each of the plurality of keywords, based on the score calculated by the score calculation unit, by deleting the keyword, generating a cluster, or deleting the cluster, selects a clustered keyword from the plurality of words including the plurality of keywords constituting the hierarchical data, and generates the knowledge data including the clustered keyword; A knowledge data generating device comprising:

2. The keyword selection unit:

2. The knowledge data generating device according to claim 1, further comprising: a step of: performing said hierarchical clustering process; selecting said clustered keywords from said plurality of keywords included in said hierarchical data; and generating said knowledge data.

3. The keyword selection unit:

2. The knowledge data generating device according to claim 1, further comprising: a step of: performing the hierarchical clustering process; selecting keywords after the clustering from a plurality of words that are superordinate concepts of the plurality of keywords included in the hierarchical data; and generating the knowledge data.

4. The keyword selection unit: a candidate pair selection unit that calculates a distance in the hierarchical data between the one or more keywords and the one or more clusters that have been subjected to the hierarchical clustering process, and selects two of the one or more keywords and the one or more clusters that have the shortest distance as a candidate pair; a center of gravity calculation unit that calculates a center of gravity of the candidate pair based on the score of the candidate pair selected by the candidate pair selection unit; performing the hierarchical clustering process of deleting a candidate of the candidate pair that is farther from the center of gravity position calculated by the center of gravity position calculation unit from the hierarchical data or clustering the candidate pair to generate a new cluster based on the center of gravity position calculated by the center of gravity position calculation unit, thereby updating the hierarchical data; a clustering processing unit that repeats a process of selecting the candidate pair by the candidate pair selecting unit, a process of calculating the center of gravity by the center of gravity calculating unit, and a process of updating the hierarchical data by performing the hierarchical clustering process until one of the keywords or one of the clusters remains in the hierarchical data, and obtains a clustering result based on the updated hierarchical data; a keyword selection processing unit that selects keywords after the clustering from the clustering result obtained by the clustering processing unit and generates the knowledge data; 2. The knowledge data generating device according to claim 1, further comprising:

5. The clustering processing unit: When performing the hierarchical clustering process, determining a first distance from the centroid location to one of the candidate pair and a second distance from the centroid location to the other of the candidate pair; if the longer of the first distance and the second distance is equal to or greater than a first set value and the shorter of the first distance and the second distance is equal to or less than a second set value, it is determined that there is a bias, and the candidate corresponding to the longer of the candidate pair is deleted from the hierarchical data; 5. The knowledge data generating device according to claim 4, wherein if the longer pair is less than the first set value or the shorter pair is greater than the second set value, it is determined that there is no bias, and the candidate pairs are clustered to generate new clusters.

6. The keyword selection processing unit:

5. The knowledge data generating device according to claim 4, further comprising: a step of selecting said clustered keywords from said clustering result in accordance with a preset limit number, and generating said knowledge data.

7. A computer constituting a knowledge data generating device for obtaining keywords related to content and generating knowledge data to be used when searching for desired content, The content is composed of multiple types of data, a data processing unit that detects character strings of text for each of the plurality of types of data by performing a predetermined process corresponding to the data, and obtains the keyword from the character strings of the text; a score calculation unit that scores the keywords obtained by the data processing unit for the plurality of types of data, and calculates a score that is an integrated result of all the predetermined processes for each same keyword; and A program for functioning as a keyword selection unit that uses hierarchical data composed of a plurality of hierarchically structured words and including a plurality of keywords obtained by the data processing unit, performs hierarchical clustering processing for each of the plurality of keywords, based on the score calculated by the score calculation unit, to select clustered keywords from the plurality of words including the plurality of keywords that constitute the hierarchical data, and generates the knowledge data including the clustered keywords.

Citation Information

Patent Citations

  • Educational teaching material navigation system

    JP2015018159A

  • WO-038、