JP2016126567A - Content recommendation device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To search for a content by adding words and phrases in another notation which are deeply associated with the search word semantically to an original search word.SOLUTION: An acquisition unit of a content recommendation device acquires primary query data for showing a list of search words. A search word candidate extraction unit extracts words and phrases of search word candidates from a plurality of pieces of corpus data in which the same topic can be described in a different notation. A query extension unit, out of the words and phrases of the search word candidates, selects the words and phrases of the search word candidates in which similarity with one of the search words contained in the primary query data is higher than a predetermined condition, and also, which co-occur with one of the search words different from the search words which have acquired similarity higher than the predetermined condition in the corpus data, and adds them to the search words. A search unit searches for a content by using the search words contained in the primary query data and the search words added by the query extension unit.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a content recommendation device and a program.

  Some content processing apparatuses for viewing content such as television programs have a content recommendation function (see, for example, Patent Document 1 and Patent Document 2). This content recommendation function has an advantage that the user can easily find the content he / she wants to view. The content recommendation function is realized by analyzing and processing the viewing history of individual users in a content reproduction environment constructed by some means.

  For example, the content processing apparatus disclosed in Patent Document 2 accumulates viewing log information including a content reproduction start time and a content reproduction end time of each user and a content ID that identifies the viewed content. The content processing apparatus mechanically extracts viewing log information reflecting individual user preferences from a set of accumulated viewing log information using heuristic rules programmed in advance by a system designer. The content processing apparatus extracts, by some means, a phrase (query) representing a topic that causes individual viewing behavior from language information corresponding to the extracted viewing log information. The language information corresponding to the viewing log information is metadata such as caption text of the content viewed by the user or a program summary sentence obtained from an electronic program guide (EPG) attached to the viewed content.

  On the other hand, there is a technique called social data mining that analyzes an enormous amount of utterance records on social media by an unspecified number of users such as Twitter (see Non-Patent Document 1, for example). In social data mining, it is possible to extract the daily interests of the general public and various topics in social life in a form with specific language expressions. In addition, there is a technology for automatically determining to which content the utterance target in social data is uttered (see Non-Patent Document 2, for example).

JP 2005-348253 A JP 2012-065119 A

M. A. Russell, “Introductory Social Data Techniques for Data Mining, Analysis, and Visualization”, O'Reilly Japan, 2011 Mariko Hirano, Keisuke Kobe, Ken Kobayakawa, “Automatic detection of tweet target programs for comprehensive and continuous detection”, Proceedings of 2013 Winter Conference of the Institute of Image Information and Television Engineers, The Institute of Image Information and Television Engineers , 2013, 3-7

  In setting a query, which is a set of words and phrases used as a content search condition, it is necessary to consider the diversity of notation. This is because a query representing a specific topic does not necessarily have a unique notation. For example, “soccer” and “football” refer to the same ball sport in many contexts. In the case where only one of the different notations representing the same topic is set as a query, it is difficult in principle to search for the content described by the other notation. Therefore, in a device that searches for content, although the surface notation (description) is different, a phrase that is assumed to match the content of the content intended by the user may be used in the query. However, if the word / phrase assumed by the device is different from the content of the content intended by the user, it is theoretically difficult to correctly identify the content corresponding to the user's intention even if the assumed word / phrase is used in the query. It is. As a result, there is a risk that some or most of the content to be presented to the user is missing from the recommendation list.

  As described above, the query used for content recommendation must be set in consideration of the variety of notations. However, Patent Documents 1 and 2 do not show any specific solution to the technical problem of how to incorporate the diversity of query notation into the query setting procedure.

  On the other hand, according to the technology of Non-Patent Document 1, social data mining is used to extract daily interests of the general public and various topics in social life in a form with specific language expressions. Is possible. The linguistic expression extracted in this way becomes a candidate for a search term to be added to a query when searching for content. Further, in the technique of Non-Patent Document 2, it is automatically determined to which content the utterance target is uttered. In content recommendation for content services, it is expected that such a technology will be used to realize query expansion considering diversity. However, the social data mining technology described in Non-Patent Document 1 and the technology for automatically determining the target content of a tweet described in Non-Patent Document 2 are semantically related to any phrase that is a query. There is no specific way to identify certain other words.

  The present invention has been made in consideration of such circumstances, and can search for content by adding words of other notations that are deeply related to the search word to the original search word. A recommendation device and a program are provided.

According to one aspect of the present invention, an acquisition unit that acquires primary query data indicating a list of search terms that are terms used for a search, and a search term candidate phrase from a plurality of corpus data in which the same topic can be described by different notations The search word candidate extraction unit to be extracted and the similarity between the search word candidate and any one of the search words included in the primary query data is higher than a predetermined condition and higher than the predetermined condition. A query expansion unit that selects and adds to the search term a search term candidate phrase that co-occurs in the corpus data with any of the search terms different from the search term that has obtained a high similarity, and the primary query data A content recommendation device comprising: a search unit that searches for content using the search term included in the search term and the search term added by the query expansion unit .
According to the present invention, the content recommendation device acquires a word / phrase as a search word candidate from a plurality of corpus data in which the same topic can be described by different notations. The content recommendation device uses the original search word used when it is determined that the similarity to the original search word is higher than the predetermined condition and the similarity is higher than the predetermined condition from the search word candidate phrases A phrase that co-occurs in the corpus data and the original search word different from is added to the search word. The content recommendation device searches for content using the original search word and the added search word.
As a result, the content recommendation device can search and recommend content by adding, to the original search word, other notation words that are deeply related to the search word.

One aspect of the present invention is the content recommendation device described above, wherein the search word candidate extraction unit extracts words of the search word candidate from the corpus data within a predetermined period.
According to the present invention, the content recommendation device selects a word to be added as a search word from the search word candidate words extracted from the corpus data for a predetermined period.
As a result, the content recommendation device can search and recommend content by adding words having other notations that are deeply related to the search word and reflect current affairs to the original search word.

One aspect of the present invention is the content recommendation device described above, wherein the acquisition unit acquires primary query data including words and phrases extracted from text information related to content viewed by a user.
According to the present invention, the content recommendation device acquires an original search word representing the user's preference based on the history of the content viewed by the user.
Thereby, the content recommendation apparatus can search and present other content that matches the user's preference from the history of the content viewed by the user.

One aspect of the present invention is the content recommendation device described above, wherein the acquisition unit acquires primary query data including words and phrases extracted from text information related to a portion of content reproduced by a user.
According to the present invention, the content recommendation device acquires a search term from text information representing the content portion of the content reproduced by the user.
As a result, the content recommendation device can acquire a search term that well represents the user's preference, and can therefore recommend content that conforms to the content desired by the user.

One aspect of the present invention is the above-described content recommendation device, wherein the search word candidate extraction unit extracts a phrase of the search word candidate from a tag or text of the corpus data.
According to the present invention, the content recommendation device extracts a phrase that is a search word candidate from a tag or text of corpus data.
As a result, the content recommendation device can extract, as a search word candidate, a phrase that well expresses the content of the text of the corpus data while suppressing the processing load by using the tag included in the corpus data. Further, the content recommendation device can extract a search term candidate word from the content of the text of the corpus data even when the tag cannot be used.

  One aspect of the present invention is a search term candidate from a plurality of corpus data in which a computer can describe primary query data indicating a list of search terms that are words used for a search, and a plurality of corpus data in which the same topic can be described by different notations The search word candidate extracting means for extracting the word of the search term, and the similarity between the search word candidate and any one of the search words included in the primary query data is higher than a predetermined condition, and the predetermined word Query expansion means for selecting and adding to the search terms a search term candidate word that co-occurs in the corpus data with any of the search terms different from the search term that has obtained the similarity higher than a condition; Content comprising: search means for searching for content using the search terms included in primary query data and the search terms added by the query expansion means Is a program for functioning as Como device.

  According to the present invention, it is possible to search for content by adding words having other notations deeply related to the search word to the original search word.

It is a functional block diagram which shows the structure of the content recommendation system by one Embodiment of this invention. It is a figure which shows the processing flow of the content recommendation process by the content recommendation apparatus of the embodiment. It is a figure which shows the processing flow of the primary query production | generation process by the viewing history analysis part of the embodiment. It is a figure which shows the processing flow of the secondary query selection process by the query expansion part of the embodiment. It is a figure which shows the processing flow of the recommendation content selection process by the recommendation list production | generation part of the embodiment. It is a figure which shows the example of the user viewing history information which the viewing history recording part of the embodiment outputs. It is a figure which shows the example of the unviewed content information which the unviewed content information recording part of the embodiment outputs. It is a figure which shows the example of the primary query data which the viewing history analysis part of the embodiment outputs. It is a figure which shows the example of the social data which the social data recording part of the embodiment preserve | saves. It is a figure which shows the example of the secondary query candidate list data which the social data analysis part of the embodiment outputs. It is a figure which shows the example of the expansion query data which the query expansion part of the embodiment produces | generates. It is a figure which shows the relationship between the expansion query and unviewed content information of the embodiment. It is a figure which shows the example of the recommendation content list data which the recommendation list production | generation part of the embodiment outputs. It is a figure which shows the example of the recommended content presentation screen which the recommended content presentation part of the embodiment displays on a content display apparatus.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
The content recommendation device according to the present embodiment uses a set of search terms in which user preferences are described by linguistic means such as keywords as a primary query, and a secondary query that is a set of other notation phrases related to the primary query and contents. Is newly added to the primary query to generate an extended query. A secondary query is a set of phrases that are deeply semantically related to the original search term. That is, the secondary query is a set of auxiliary search terms that indirectly expresses the user's preference. The content recommendation device according to the present embodiment searches for content to be recommended to the user (hereinafter also referred to as “recommended content”) using the generated extended query.
As described above, the content recommendation device according to the present embodiment is a content that should be recommended to the user by using the original search word that represents the user's preference and another search word that is deeply related to the original search word. Search for. Therefore, the content recommendation device of the present embodiment enables advanced content recommendation in consideration of the user's potential preference and topic relevance.

  In addition, individual concepts (ie, topics) that constitute user preferences are often influenced by current events and other social situations, and are considered to change from day to day. Therefore, in setting the query, it is necessary to sufficiently consider the topical nature of the topic according to the time when the content recommendation is performed. For example, when the user's interest is a temporary topic that is limited to a very short period of time, such as entertainment topics or incident / accident news, in addition to the user's long-term preference, the user is newly interested It is necessary to reflect a search term that reflects the topic of current events that is highly likely. Therefore, the content recommendation device according to the present embodiment sets the above-described extended query used for the search of recommended content in consideration of the currentity in addition to the variety of notations. Thereby, the content recommendation apparatus of this embodiment can implement | achieve the content recommendation according to the content which the user calculates | requires reflecting the topic of the new thing which appears every day in addition to a user's potential preference. Therefore, the content recommendation device according to the present embodiment is suitable for a broadcasting service that handles a large amount of highly specialized content such as sports broadcasts and a highly current content such as news programs and documentary programs.

  FIG. 1 is a functional block diagram showing a configuration of a content recommendation system according to an embodiment of the present invention, and only functional blocks related to the present embodiment are extracted and shown. As shown in the figure, the content recommendation system includes a content recommendation device 1 and a content display device 3. The content recommendation device 1 and the content display device 3 are connected via a network 9 such as an IP (Internet Protocol) network. Further, the social media service providing apparatus 5 is connected to the network 9. In the figure, only one content display device 3 and one social media service providing device 5 are shown, but a plurality of devices may exist.

  The content recommendation device 1 is realized by, for example, one or more computer devices. The content recommendation device 1 includes a viewing history recording unit 11, an unviewed content information recording unit 12, a viewing history analysis unit 13 (acquisition unit), a social data recording unit 14, and a social data analysis unit 15 (search word candidate extraction). Section), a query expansion section 16, a recommendation list generation section 17 (search section), a recommended content presentation section 18, and a storage section 19. For example, when the content recommendation device 1 is realized by a plurality of computer devices connected via a network, which computer device realizes which functional unit can be arbitrary.

  The viewing history recording unit 11 acquires and records user viewing history information. The user viewing history information indicates the history of the content viewed by the user in the content service that is the processing target of the content recommendation device 1. The user viewing history information includes identification information of content viewed by the user and text information related to the content. The content can be any content data such as a TV program, a moving image, a still image, a web page, a document, text, and an electronic book. For example, the viewing history recording unit 11 acquires user viewing history information to be recorded based on the content viewing information received from the content display device 3.

  The unviewed content information recording unit 12 acquires and records unviewed content information. Unviewed content information includes text information related to content that has not been viewed by the user among content that can be provided to the user in the content service. Content that has not been viewed by the user is also referred to as “unviewed content”. The unviewed content can be any content data such as a TV program, a moving image, a still image, a web page, a document, a text, and an electronic book.

  The viewing history analysis unit 13 analyzes the user viewing history information recorded in the viewing history recording unit 11 and generates primary query data indicating a set of phrases that are primary search terms. A phrase (search term) that is an element of the phrase set indicated by the primary query data is referred to as “primary query phrase”.

  The social data recording unit 14 acquires and records social data posted by an unspecified number of contributors on social media from the social media service providing apparatus 5. An example of social media is Twitter (Twitter). The social data is, for example, data published on social media accessible via the Internet, and includes text information indicating the content of the contributor's remarks and information on the date and time when the remark content is disclosed (posted). In the social data indicating the contents of remarks of an unspecified number of contributors, the same topic can be described by differently expressed words and phrases.

  The social data analysis unit 15 analyzes the social data recorded in the social data recording unit 14 and generates secondary query candidate list data indicating a set of terms of secondary query candidates. A secondary query candidate is a candidate for a phrase that should be an element of a secondary query. The secondary query is a set of terms that are deeply related to the terms of the primary query and serve as search terms to be added to the primary query. That is, the phrase of the secondary query candidate is the phrase of the search word candidate.

  The query expansion unit 16 compares the primary query word / phrase included in the primary query data with the secondary query candidate word / phrase included in the secondary query candidate list data, and selects the primary query from the secondary query candidate words / phrases. A word having a deep semantic relationship with the word is extracted as a word of the secondary query. The query expansion unit 16 generates expanded query data that is a set of search terms obtained by adding the extracted secondary query terms to the primary query terms.

  The recommendation list generation unit 17 calculates a matching score using text information related to the unviewed content indicated by the unviewed content information and the extended query data. The recommendation list generation unit 17 ranks each unviewed content according to the calculated matching score, and selects the recommended content from the unviewed content based on the rank. The recommended list generation unit 17 generates recommended content list data listing information for specifying the selected recommended content.

  The recommended content presentation unit 18 presents information regarding each recommended content indicated by the recommended content list data. That is, the recommended content presentation unit 18 transmits recommended content presentation information in which information related to the recommended content is set to the user's content display device 3 for display.

  The storage unit 19 stores text information regarding each content. The storage unit 19 stores information on content that can be provided to each user in the content service.

  The content display device 3 is, for example, a user's personal computer, a smartphone, a tablet terminal, a television receiver, or the like. The content display device 3 includes an operation unit 31, an acquisition unit 32, an output unit 33, a notification unit 34, and a reception unit 35. The operation unit 31 receives an operation by a user. The operation unit 31 is a device that receives an operation by a key, a button, a mouse, a touch sensor arranged on a touch panel, or a remote controller, for example. The acquisition unit 32 acquires the content selected by the user's operation received by the operation unit 31 from the content provided by the content service available to the user. The output unit 33 is a display or a speaker, and outputs the content acquired by the acquisition unit 32. The notification unit 34 notifies the content recommendation device 1 of content viewing information indicating the content viewed by the user. The receiving unit 35 receives recommended content presentation information from the content recommendation device 1 and causes the output unit 33 to display the recommended content presentation information.

Next, the operation of the content recommendation device 1 will be described. Hereinafter, a case where the content is a television program will be described as an example.
The viewing history recording unit 11 records user viewing history information indicating the history of content viewed by each user in the past in association with a user ID that is user identification information. Specifically, the viewing history recording unit 11 receives content viewing information in which information on content viewed by the user and information specifying the user are set from the notification unit 34 of the content display device 3. The content viewed by the user is content acquired by the acquisition unit 32 based on a user operation received by the operation unit 31 of the content display device 3 and output by the output unit 33. Based on the received content viewing information, the viewing history recording unit 11 associates the user viewing history information including the content ID of the content viewed by the user and the text information describing the content with the user ID. Record. The text information is, for example, a program summary sentence. The storage unit 19 stores a program summary sentence acquired from the broadcast wave received by the content recommendation device 1, and the viewing history recording unit 11 reads the program summary sentence from the storage unit 19.

  As described above, the viewing history recording unit 11 basically records text information indicating the content of the content viewed by the user in units of content based on the user viewing history information. Note that, as described in Patent Document 2, the viewing history recording unit 11 may further set information that records in detail a detailed operation history when a user views content as user viewing history information. In this case, the notification unit 34 of the content display device 3 further notifies the content recommendation device 1 of the operation history when the user views the content. The operation history is, for example, a content reproduction start point and a reproduction end point.

  On the other hand, the social data recording unit 14 accesses the social media service providing apparatus 5 via the network 9. The social media service providing apparatus 5 publishes social data posted by an unspecified number of contributors. The social data includes text information indicating the content of the utterance of the poster and a time stamp indicating the posting date and time of the text information. The social data recording unit 14 acquires and records social data posted by an unspecified number of contributors from the social media service providing apparatus 5 as an access destination.

  The social data recording unit 14 is based on acquiring all social data that can be acquired. That is, the social data recording unit 14 collects data of all utterance records on accessible social media. However, in the case of the purpose of realizing content recommendation with limited content to be analyzed or content recommendation with more importance on current affairs, the social data recording unit 14 further performs a process of classifying and selecting social data to be collected. It's okay. For the classification and selection of social data to be collected, existing techniques for automatically determining the utterance target of social data and auxiliary means capable of specifying the date and time of utterance can be used. The technology of Non-Patent Document 2 can be used as a technology for automatically determining the utterance target of social data. In addition, technologies that comprehensively acquire social data in a form that can specify the date and time of remarks include “Sho Hashimoto,“ tw twitter client on Ruby ”, [online], Internet <URL: http: //shokai.github .io / tw /> "etc. can be used. Thereby, the social data recording unit 14 collects social data whose time stamp indicates a posting date and time within a predetermined period. For example, when reflecting recent topics, the social data recording unit 14 collects social data for a period of days, weeks, months, or years ago from the present, If you want to reflect, collect social data of the specified past period.

  FIG. 2 is a diagram showing a processing flow of content recommendation processing by the content recommendation device 1. By the above processing, before the content recommendation processing is started, the viewing history recording unit 11 has the user viewing history information including the content ID of the content viewed by each user and the text information describing the content of the content. It is recorded. The content recommendation device 1 executes the content recommendation process shown in FIG. 2 for each user.

  The viewing history recording unit 11 outputs the user viewing history information provided with the user ID of the user who recommends the content to the unviewed content information recording unit 12 and the viewing history analysis unit 13 (step S110).

  The unviewed content information recording unit 12 searches for unviewed content of a user who recommends content based on the user viewing history information received from the viewing history recording unit 11. Specifically, the unviewed content information recording unit 12 refers to the content information that can be provided to each user stored in the storage unit 19, the content ID is not set in the user viewing history information, and The content that can be used by the user is searched for as unviewed content. The unviewed content information recording unit 12 reads the text information describing the contents of the unviewed content from the storage unit 19. The text information is, for example, a program summary sentence or the like, and is acquired from the broadcast wave by the content recommendation device 1 and stored in the storage unit 19. The unviewed content information recording unit 12 generates unviewed content information that is a list of text information of unviewed content (step S110).

  As described above, the unviewed content information recording unit 12 has not yet been viewed by the user from all the contents that the user can obtain in a computer-readable state in the content service assumed as the processing target by the content recommendation device 1. Search for content. The available content may be, for example, a program scheduled to be broadcast up to one week in advance, or may be content that can be currently distributed in a content service subscribed to by the user. When it is known in advance that there is a clear bias in the viewing tendency of the user, the unviewed content information recording unit 12 performs a process of limiting the target to be selected as unviewed content according to the content attribute such as the genre. May be.

The viewing history analysis unit 13 receives user viewing history information from the viewing history recording unit 11. The viewing history analysis unit 13 analyzes text information related to the contents of each viewed content described in the received user viewing history information, and generates primary query data indicating a primary set of search terms (step) S115).
Specifically, the viewing history analysis unit 13 divides text information such as a program summary sentence described in the user viewing history information into words using a known morphological analysis technique. The viewing history analysis unit 13 selects a word (word) as a search word from the divided words based on information such as a part of speech given to each word as a result of morphological analysis. For example, the viewing history analysis unit 13 selects semantically important phrases (words) such as proper nouns (for example, personal names) as search terms. The viewing history analysis unit 13 describes the selected word / phrase in the form of a list to obtain primary query data.

  When an operation history is set in the user viewing history information, the viewing history analysis unit 13 performs morphological analysis on text information corresponding to a playback section of content viewed by the user as described in Patent Document 2. You may limit to. The content playback section is indicated by the playback start point and playback end point of the content indicated by the operation history set in the user viewing history information. The text information corresponding to the content playback section is, for example, subtitle data of the content in the playback section. In the content recommendation device 1, the storage unit 19 accumulates the caption information of each content acquired from the broadcast wave, and the viewing history analysis unit 13 reads out the caption data of the content in the playback section from the storage unit 19.

  The social data recording unit 14 outputs the recorded social data to the social data analysis unit 15 (step S120). In consideration of current affairs, the social data recording unit 14 outputs social data for a predetermined period to the social data analysis unit 15. Moreover, the social data recording unit 14 may output predetermined social data to be spoken to the social data analysis unit 15. Further, the social data recording unit 14 may collect social data after the process of step S <b> 115 without collecting social data in advance, and output the social data to the social data analysis unit 15. In this case, the social data recording unit 14 may acquire and record social data using the primary query data obtained in the process of step S115.

  The social data analysis unit 15 analyzes the social data received from the social data recording unit 14 and extracts words of secondary query candidates that are candidates for search terms to be added to the primary query data. The social data analysis unit 15 generates secondary query candidate list data in which the extracted secondary query candidate words are set (step S125).

  When a label attached to the social data main body is available, such as a hash tag in Twitter which is one of social media, the social data analysis unit 15 uses the character strings (words) of the labels as they are. Basically, it is used as a word for the next query candidate. Further, the social data analysis unit 15 may analyze the main body of the social data by the same process as the viewing history analysis unit 13 and perform a process of directly extracting important words such as proper nouns from the main body of the social data. . The main body of the social data is data of the body text describing the content of the poster's remarks in the social data as a character string.

  When each acquired phrase of the secondary query candidate co-occurs with other words of the secondary query candidate in the original social data from which the phrase is obtained, the social data analysis unit 15 selects secondary query candidate list data. In addition, the supplementary information is recorded in addition to the words of the secondary query candidate. In the supplementary information of the secondary query candidate word, all other words of the secondary query candidate that co-occur in the original social data from which the secondary query candidate word is obtained are set. This supplementary information is used in the secondary query selection process in the next step S130.

  After both the primary query data and the secondary query candidate list data are generated as described above, the query expansion unit 16 stores the contents in a plurality of phrases of the primary query from the phrases set in the secondary query candidate list data. Words that have some kind of relationship. Further, the query expansion unit 16 adds a phrase selected from the extracted secondary query candidate phrases to the primary query data, thereby obtaining expanded query data (step S130).

  First, the query expansion unit 16 calculates the similarity between each word of the secondary query candidate set in the secondary query candidate list data and each word of the primary query set in the primary query data by some means. To do. The query expansion unit 16 is based on a process of adopting as a secondary query only for words / phrases whose similarity to the word / phrase of the primary query exceeds a predetermined threshold among words / phrases of the secondary query candidate.

  Note that when supplemental information describing another phrase of a secondary query candidate that co-occurs in social data is added to the phrase of the secondary query candidate, only the phrase of the primary query may be left in the supplemental information. Then, the query expansion unit 16 selects a phrase other than the primary query word that is one of the pair of supplementary information of the secondary query candidate word from the pair of the secondary query candidate word and the primary query word. Pairs that are not set may be excluded from the target of similarity calculation. By performing this processing, the query expansion unit 16 has a deep semantic relationship with at least one of the phrases in the primary query, and has some semantic connection between the phrase of at least one other primary query. It is possible to extract words that are guaranteed to exist. That is, words that only have a semantic connection to a single primary query phrase are excluded from the secondary query candidates, and only words that have a semantic connection to multiple primary query phrases are included in the secondary query. Extracted as a phrase.

  As a technique for quantitatively calculating the similarity between words, there is a technique of vector expression of words reflecting a semantic distance by a multi-layer neural network (for example, see Reference 1). In addition, a technique of word clustering (for example, Reference 2) using the magnitude of a singular value based on the appearance tendency of a word document for weighting a feature amount can be used. However, any calculation method can be used as long as it is a technique that can quantify the semantic similarity of a pair of an arbitrary phrase acquired from the primary query data and an arbitrary phrase selected from the secondary query candidate list data. Well, other techniques may be used.

  (Reference 1) Yasukazu Nishio, “Natural Language Processing with word2vec”, O'Reilly Japan, May 2014

  (Reference 2) Mariko Hirano, Keisuke Kobe, Ken Kobayakawa, “Documents for Overlooking Large-Scale Data and Extracting Target Data—Effectiveness of Singular Value Decomposition of Word Matrix and Weighting by Singular Values”, Natural Language Processing Society, Nature Journal of Language Processing Society, 2013, Vol.20, no. 3, p. 335-365

  The query expansion unit 16 calculates the similarity between words as described above for all logically possible pairs of each word of the primary query and each word of the secondary query candidate, and the similarity is a predetermined threshold value. Basically, a word of a secondary query candidate that satisfies the above condition is selected as a word of a secondary query (search word). At this time, the query expansion unit 16 may assign a score calculated separately based on the value of the similarity calculation result to each word (search term) of the selected secondary query. For the score, for example, a value obtained by normalizing the similarity so that the similarity between the same words becomes 1 can be used. Moreover, you may use similarity degree itself as a score. This score is used in the recommended content selection process in the next step S135.

  The recommendation list generation unit 17 reads unviewed content information from the unviewed content information recording unit 12. The recommendation list generation unit 17 selects the recommended content to be presented to the user from the list of unviewed content described in the unviewed content information using the extended query data. The recommendation list generation unit 17 collects the selected recommended contents in a list format in the order of matching with the extended query data. Therefore, the recommendation list generation unit 17 sets the text information of each unviewed content set in the unviewed content information, and the words of the primary query and the secondary query set in the expanded query data (each of the expanded query data). A matching score is calculated for a pair with (element). The recommendation list generation unit 17 ranks each unviewed content according to the calculated matching score, and shows a list that lists the content IDs of the top N unlisted content (N is an integer of 1 or more). Recommended content list data is generated (step S135).

  The matching score may be the number of times that the search word appears in the text information describing the content of the unviewed content. The recommendation list generation unit 17 is basically based on a method of handling matching scores for each of the primary query and the secondary query equally. Specifically, a sum (simple sum) that calculates a matching score for each of the primary query terms (search terms) and secondary query terms (search terms) independently and treats the matching scores with equal weights. Take. The recommendation list generation unit 17 adds a weighting process to be added to the matching score for the former primary query after multiplying the matching score for the latter secondary query by a weight determined by some method. May be.

  Further, in the case of a method of multiplying the matching score of the secondary query by a weight, a method of using the similarity between the primary query and the secondary query in addition to a method of fixing the weight to be used to an heuristically determined experience value Can be considered. As a specific method of utilizing the similarity between the latter queries, for example, an average value of scores added to each word of the secondary query is set as a weight (weight having a numerical value between 0 and 1). . As described above, the score added to each word of the secondary query is a score calculated by the query expansion unit 16 based on the similarity to the word of the primary query. The recommended content presentation unit 18 sums up the matching scores of each word in the secondary query, and multiplies the total matching score by the average score based on the similarity, and then adds it to the matching score for the primary query.

  Finally, the recommended content presentation unit 18 presents the content of the recommended content described in the recommended content list data to the user. That is, as the recommended content presentation information, information related to each recommended content described in the recommended content list data is read from the storage unit 19, and recommended content presentation information in which the read information is set is generated (step S140). The recommended content presentation unit 18 transmits the generated recommended content presentation information to the user's content display device (step S145). The receiving unit 35 of the content display device 3 causes the output unit 33 to output the received recommended content presentation information.

  The recommended content presentation information is text information for specifying each recommended content described in the recommended content list data. For example, when the recommended content is a broadcast program, the recommended content presentation information describes the broadcast date of the broadcast program, the broadcast start time, the program name, and the like. In addition, when the information for directly accessing the recommended content by the user or the sample video obtained by cutting out a part of the video of the recommended content can be acquired from the storage unit 19 or a computer server connected via the network, the recommended content The presentation unit 18 may set such information as auxiliary information in the content presentation information. As information for the user to directly access the recommended content, for example, link information of Internet distribution content can be used. Moreover, a thumbnail image, a highlight video, a notice video, etc. can be used for the sample video.

  The display format of the recommended content presentation information is basically a text-based static display format that can be displayed as a list on the screen of the computer device. If sample video of recommended content is available, such auxiliary information (moving image) is presented in a predetermined area on the screen in the order described in the recommended content list data (operation playback). A visual device may be implemented separately.

  In the process of FIG. 2, the content recommendation device 1 may execute any process among the process of step S110, the process of step S115, and the process from step S120 to step S125 in parallel.

FIG. 3 is a diagram showing a processing flow of primary query generation processing by the viewing history analysis unit 13, and shows details of the primary query generation processing in step S115 of FIG.
The viewing history analysis unit 13 receives user viewing history information from the viewing history recording unit 11 (step S205). In the basic method, the viewing history analysis unit 13 extracts text information such as a program summary representing the content of the viewed content from the user viewing history information (step S210). As another method, the viewing history analysis unit 13 stores text information such as subtitle text corresponding to the playback section of the viewing content set in the user viewing history information as in the method described in Patent Document 2. Obtained from the unit 19.

  The viewing history analysis unit 13 performs a morpheme analysis process on the character string corresponding to the text information extracted in step S210, and decomposes it into a string of words (morphemes) to which part-of-speech information is assigned (step S215). The text information to be subjected to morphological analysis is a character string corresponding to the entire viewing content or the playback section of the viewing content. As a specific means of morphological analysis, a known technique such as MeCab which is open source morphological analysis software can be used.

  Next, the viewing history analysis unit 13 selects a phrase that can specify the entire viewing content or the topic related to the playback section of the viewing content from the sequence of phrases with part-of-speech information obtained by the morphological analysis in step S215. (Step S220). For example, the viewing history analysis unit 13 selects nouns (proper nouns) whose target objects are limited, such as phrases such as names of people, organizations, regions, and products, based on the part of speech information. Finally, the viewing history analysis unit 13 collects the selected words and phrases in a list format and outputs them as primary query data.

FIG. 4 is a diagram showing a processing flow of the secondary query selection process by the query expansion unit 16, and shows details of the secondary query selection process in step S130 of FIG. Here, an example of processing for selecting a secondary query word / phrase using supplementary information from secondary query candidate words / phrases will be described.
First, the query expansion unit 16 receives primary query data from the viewing history analysis unit 13 and receives secondary query candidate list data from the social data analysis unit 15 (step S305). Next, the query expansion unit 16 determines, for each word / phrase described in the secondary query candidate list data, whether or not the word / phrase is a word / phrase described in the primary query data. The query expansion unit 16 excludes a phrase that matches one of the phrases described in the primary query data from the phrases of the secondary query candidates described in the secondary query candidate list data (step S310). .

  Next, the query expansion unit 16 determines whether or not the word / phrase included in the supplemental information is the word / phrase described in the primary query data for each word / phrase of the secondary query candidate excluding the word / phrase of the primary query in step S305. Judging. The supplementary information added to the phrase of the secondary query candidate indicates another phrase that co-occurs in the social data with the phrase of the secondary query candidate. The query expansion unit 16 excludes words / phrases that do not match any of the words / phrases described in the primary query data from the words / phrases indicated by the supplementary information (step S315). As a result, words that do not appear in the primary query data are excluded from the words of the co-occurrence of the words of the secondary query candidate.

  The query expansion unit 16 removes words other than the words of the primary query from the supplementary information of the secondary query candidate through the process of step S315, and then includes each word and phrase included in the secondary query candidate list data and the primary query data. Each phrase is converted into a vector expression reflecting the semantic distance between words by some means. The query expansion unit 16 calculates the semantic similarity between the words by some means for each possible combination of the words of the secondary query candidate and the words of the primary query (step S320). As a specific method for quantitatively evaluating the degree of similarity, for example, any existing method such as Reference Document 1 and Reference Document 2 described above can be used, but this is not restrictive.

  The phrase of the secondary query candidate is the phrase A, and the phrase of the primary query for which the similarity with the phrase A is calculated is the phrase C. The query expansion unit 16 uses the phrase of the partner whose similarity exceeds a predetermined threshold for all the phrases A of the secondary query candidates whose similarity to the phrase C of any one or more primary queries exceeds a predetermined threshold. Save C by overwriting the secondary query candidate list data. In addition, although the method of fixing to a empirical value can be considered as the setting method of a threshold value, it is not this limitation.

Next, the query expansion unit 16 co-occurs the word A of the secondary query candidate, the word C of the primary query whose similarity with the word A exceeds a predetermined threshold, and the word A from the secondary query candidate list data. The phrase B of the primary query to be read is read (step S325). The phrase B of the primary query in which the phrase A co-occurs is read from the supplementary information of the phrase A.
For each word A of the secondary query candidate, the query expansion unit 16 has the same word B of the primary query in which the word A co-occurs and the word C of the primary query whose similarity with the word A exceeds a predetermined threshold. It is determined whether or not. The query expansion unit 16 does not select the secondary query candidate phrase A having the same phrase B and phrase C as the secondary query phrase, and is added to the phrase A from the secondary query candidate list data. Together with the supplementary information and the phrase C. The query expansion unit 16 leaves the phrase A of the secondary query candidate in which the phrase B and the phrase C are different from each other in the secondary query candidate list data (step S330).

Finally, the query expansion unit 16 adopts the phrase A remaining in the secondary query candidate list data as a secondary query. In other words, the query expansion unit 16 includes a partner query whose similarity exceeds a predetermined threshold in the primary query data, and the partner query is different from the query of the co-occurrence partner primary query. The phrase is adopted as the phrase for the secondary query. The query expansion unit 16 outputs, as expanded query data, a list obtained by concatenating the primary query data including the primary query word list and the adopted secondary query word list (step S335).
In the extended query data, as shown in a specific example of the extended query data in FIG. 11 described later, the value obtained by the previous similarity calculation (similarity between the words of the secondary query and the words of the primary query). May be included in each word, but this is not an essential process.

FIG. 5 is a diagram showing a process flow of recommended content selection processing by the recommendation list generation unit 17, and shows details of recommended content selection processing in step S135 of FIG. Here, it shows about the case where a weighting process is performed to the matching score of a secondary query.
First, the recommendation list generation unit 17 receives unviewed content information from the unviewed content information recording unit 12, and receives expanded query data from the query expansion unit 16 (step S405). The recommendation list generation unit 17 acquires primary query data that is a subset of the extended query data. For each content described in the list of unviewed content information, the recommendation list generation unit 17 calculates a matching score between the text information of the content and the primary query word belonging to the primary query data, and sets it as the primary score ( Step S410). As a specific process for calculating the primary score, for example, a method of simply adding up the appearance frequency of words that match the word of the primary query at the notation level and using the total value of the appearance frequencies as it is, Not so.

  Next, the recommendation list generation unit 17 acquires a list of phrases of the secondary query that is the remaining subset of the extended query data. For each content described in the list of unviewed content information, the recommendation list generation unit 17 calculates a matching score between the text information of the content and the phrase of the secondary query, and sets it as a secondary score (step S415). . As a specific process for calculating the secondary score, the method of using the cumulative value of the appearance frequency of the phrase that matches the phrase of the secondary query at the notation level as in the above-described calculation method of the primary score is considered. Yes, but not so.

  Next, the recommendation list generation unit 17 multiplies each of the primary score and the secondary score calculated for each content by a predetermined weight, calculates a sum of them, and sets the value as a matching score for the content (step) S420). After calculating all the matching scores for each unviewed content, the recommendation list generation unit 17 rearranges the content by some means based on the value of the matching score. As a specific means for rearranging the list of unviewed content, for example, a sort of UNIX (registered trademark) command, which is a well-known technique, can be used, but this is not restrictive. The recommendation list generation unit 17 stores and outputs the content IDs of the top N unviewed contents of the sorted matching scores in the recommended content list data (step S425).

Next, an operation example of the content recommendation device 1 will be described using specific data examples.
FIG. 6 is a diagram illustrating a specific example of user viewing history information output by the viewing history recording unit 11. In the user viewing history information, information for specifying the content viewed by the user and text information indicating the content are described for each content. When the content to be analyzed is a broadcast program, as shown in the figure, the user viewing history information includes a broadcast channel name, a broadcast date and time, a program name, and a program summary sentence of the program viewed by the user in a list format. Described. It should be noted that various information regarding these broadcast programs can be obtained in a computer-readable state from broadcast waves by using commercial software such as SKNET's Monster TV.

  FIG. 7 is a diagram illustrating a specific example of unviewed content information output by the unviewed content information recording unit 12. The unviewed content information is information that is not included in the user viewing history information and that can be used by the user at present and in the future in a form similar to the user viewing history information. When the content to be analyzed is a broadcast program, the unviewed content information includes the program ID, broadcast channel name, broadcast date and time, program name, and program outline of each program scheduled to be broadcast from the start of the content recommendation process to one week ahead. A sentence is described in a list format. It should be noted that the various types of information related to the broadcast-scheduled program can be acquired in a computer-readable state from the broadcast wave by using commercial software such as SKNET's Monster TV described above.

  FIG. 8 is a diagram illustrating a specific example of primary query data output from the viewing history analysis unit 13. In the primary query data, in step S115, the viewing history analysis unit 13 describes the search terms that are the phrases extracted from the user viewing history information in a list format. The primary query data shown in the figure shows a set of search terms including the three phrases “architecture”, “Scotland”, and “politics” extracted from the user viewing history information shown in FIG. 6 by the viewing history analysis unit 13.

  FIG. 9 is a diagram illustrating a specific example of social data stored by the social data recording unit 14. In the social data shown in the figure, the social data recording unit 14 displays the words “architecture”, “Scotland”, and “politics” of the primary query included in the primary query data shown in FIG. The tweet contents obtained by inputting as a search term are shown. A character string in parentheses in each entry represents a tweet speaker and a utterance date. The character string following the parenthesis indicates the content of each tweet. The character string starting with the last “#” symbol of the entry is a label (hash tag) for classifying the content of the tweet.

  FIG. 10 is a diagram illustrating a specific example of secondary query candidate list data output by the social data analysis unit 15. The secondary query candidate list data shown in the figure shows a list of phrases of secondary query candidates extracted by the social data analysis unit 15 using the hash tag from the social data shown in FIG. 9 in step S125. Secondary query candidate phrases “Architecture”, “Scotland”, “Glasgow”, “Narrow”, “Minimal”, “Invitation”, “Politics”, “Military Song” are followed by social data Supplementary information indicating other words co-occurred in is written in parentheses.

  In the process of step S310 in FIG. 4, the query expansion unit 16 determines whether or not the phrase is described in the primary query data for each phrase described in the secondary query candidate list data illustrated in FIG. Determine whether. The query expansion unit 16 excludes the phrases “architecture”, “Scotland”, and “politics” of the primary query included in the primary query data from the phrases described in the secondary query candidate list data. Accordingly, the secondary query candidate list data includes “Glasgow (Architecture, Scotland)”, “Narrow (Architecture, Minimal)”, “Minimal (Architecture, Narrow)”, “Whisky (Scotland)”, “ Scotland) ”and“ Military Song (Politics) ”remain.

  Further, in the process of step S315, the query expansion unit 16 excludes words / phrases that do not appear in the primary query data from the supplementary information set in the secondary query candidate list data. At this time, phrases of secondary query candidates whose supplementary information does not include words that appear in the primary query data are also excluded from the secondary query candidate list data. As a result, the secondary query candidate list data includes "Glasgow (Architecture, Scotland)", "Narrow (Architecture)", "Minimal (Architecture)", "Whisky (Scotland)", "Military song (politics)" remains.

  The query expansion unit 16 includes the secondary query candidate words “Glasgow”, “narrow”, “minimal”, “whiskey”, “residential vote”, “military song”, and the primary query candidate remaining in the secondary query candidate list data. The similarity with each of the query terms “architecture”, “Scotland”, and “politics” is calculated. In step S320, the query expansion unit 16 determines that the similarity between the secondary query word “Glasgow” and the primary query word “politics” is similar to the primary query word “Scotland”. Select the phrase “local referendum” in the secondary query above. The query expansion unit 16 writes, in the secondary query candidate list data, the phrase “Scotland” of the partner's primary query having a similarity equal to or higher than a predetermined level in association with the phrase “Glasgow” of the secondary query candidate. Further, the query expansion unit 16 writes the phrase “politics” of the primary query of the partner having a similarity equal to or higher than a predetermined level in association with the phrase “resident vote” of the secondary query candidate. The query expansion unit 16 uses the secondary query terms “narrow”, “minimal”, “whiskey”, “military” of the secondary query whose similarity to the primary query is lower than a predetermined value from the secondary query candidate list data, Delete supplemental information for a phrase.

In steps S325 to S330, the query expansion unit 16 performs the following processing. That is, the query expansion unit 16 uses the secondary query candidate word / phrase “Glasgow” from the secondary query candidate list data, the supplementary information (architecture, Scotland) of the word, and the primary query word / phrase of the partner having a similarity equal to or higher than a predetermined value Read "Scotland". Since the query expansion unit 16 is set with the phrase “architecture” of the primary query other than the phrase “Scotland” of the primary query of the partner whose similarity is equal to or higher than a predetermined level, the query expansion unit 16 sets the phrase “Glasgow” as the secondary query candidate. As the secondary query term. The query expansion unit 16 leaves the phrase “Glasgow” of the secondary query candidate as it is in the secondary query candidate list data.
In addition, the query expansion unit 16 uses the secondary query candidate list data from the secondary query candidate list data, the phrase “resident vote”, supplementary information (Scotland) of the phrase, and the phrase “of the primary query of a partner whose similarity is equal to or higher than a predetermined value”. Read "politics". Since the query expansion unit 16 sets the phrase “scotland” of the primary query other than the phrase “politics” of the primary query of the partner whose similarity is equal to or higher than a predetermined level in the supplementary information, As the secondary query term. The query expansion unit 16 leaves the phrase “resident vote” of the secondary query candidate in the secondary query candidate list data as it is.

FIG. 11 is a diagram illustrating a specific example of the expanded query data generated by the query expansion unit 16.
The first half of the phrase “architecture”, “Scotland”, and “politics” in the list of phrases set in the expanded query data shown in FIG. 3 are phrases inherited from the primary query data. The latter half of the phrase “Glasgow” and “resident vote” in the list of phrases set in the expanded query data are the phrases that the query expansion unit 16 has adopted as secondary queries from the secondary query candidate phrases in step S335. It is.

  In the figure, the numerical value written together with each word indicates the highest similarity among the similarities between the word calculated by the query expansion unit 16 and each word of the primary query. Note that the similarity between the same words is 1.00. Therefore, the phrase taken over from the primary query data has the highest similarity with the own phrase, and is 1.00.

  FIG. 12 is a diagram for explaining the relationship between the phrase of the extended query and the unviewed content information. In the figure, the terms described in circles R1 to R3 are the primary query terms “architecture”, “Scotland”, and “politics”, respectively. The phrases described in the circles R4 to R5 are secondary query candidate phrases, which are the phrases “Glasgow” and “resident vote” selected for the secondary query. The phrases described in the circles R4 to R9 are the phrases “narrow”, “minimal”, “whiskey”, and “military song” that are candidates for the secondary query but are not selected for the secondary query. The center position of each circle represents the position when a word described in the circle is projected onto a vector space reflecting the semantic similarity of the word by a certain means. That is, the words in circles arranged at close positions in the figure indicate that the semantic similarity is large.

  An arrow attached to each circle represents a relationship in which words described in each circle co-occur in social data. The original phrase of the arrow indicates that the phrase is present in the primary query data, and the phrase at the end of the arrow indicates that the phrase is a secondary query candidate. For example, the phrase “Glasgow” of the secondary query candidate (secondary query) represents a co-occurrence relationship in the social data with the phrases “Architecture” and “Scotland” of the primary query. In addition, the phrase “narrow” of the secondary query candidate represents a co-occurrence relationship in the social data with the phrase “architecture” of the primary query.

  The text information T1 indicates unviewed content information that hits the phrases “Scotland” and “Architecture” of the primary query and the phrase “Glasgow” of the secondary query. The text information T2 indicates unviewed content information that hits the phrase “Scotland” in the primary query and the phrase “resident vote” in the secondary query. That is, these are specific examples of content that is shifted up the recommendation list by using a secondary query.

  The content recommendation device 1 uses information on both the co-occurrence relationship (arrow) between the words of the primary query and the words of the secondary query candidate, and the similarity degree between the words (the proximity of the positions of the circles). Then, it is determined which word / phrase is adopted as the secondary query from the words / phrases of the secondary query candidate. Then, the content recommendation device 1 calculates the matching score with the text information of the unviewed content by using the adopted secondary query phrase together with the primary query. As a result, the content recommendation device 1 can position the content more reflecting the user's potential preference and the topic of the general public at the top of the recommended content list.

  In the following, the criteria for selecting terms in the secondary query will be described in detail with specific examples. As described above, among the phrases of the secondary query candidates, the phrases “Glasgow” and “resident referendum” of the circles R4 and R5 are the phrases adopted as the secondary query, and the phrases “narrow” of the circles R6 to R10, “Minimal”, “whiskey”, and “military song” are rejected phrases as secondary queries. The following two points are used as criteria for determining whether or not to accept the secondary query candidate phrase A as the secondary query phrase.

(1) The word A of the secondary query candidate has a high similarity to the word C of any primary query.
(2) The phrase B of the primary query that is co-occurring with the phrase A is different from the phrase C of the primary query of (1).

  The content recommendation device 1 adopts the secondary query candidate word / phrase A that satisfies the two criteria (1) and (2) as the secondary query word / phrase and satisfies one or both of the conditions. Reject the missing phrase A from the secondary query. The phrase A that satisfies both of the above two criteria is highly similar to the phrase C of the primary query that reflects the user's interest and is different from the phrase C of the primary query that is different from the phrase C and is present in social data. It is synonymous with having some relationship in context. That is, in order for the phrase A to be adopted, it is a phrase that is analogized to indicate the target of the user's potential interest based on the co-occurrence relationship with the phrase B on social data, and the user's explicit It is a condition that the meaning is close to that of the word C indicating the object of general interest. The phrase A that satisfies the above two criteria is used by the content recommendation device 1 by utilizing the co-occurrence relationship with the phrase of the primary query (object of the user's explicit interest) in social data that deals with many current topics. By analogy, it is a target of the user's potential interest, and it is highly likely that it is a search word (secondary query) that reflects the topic of current events in society in general.

  For example, the phrase “Glasgow” (specific example of the phrase A) of the secondary query candidate shown in FIG. 12 has a co-occurrence relationship with the phrase “architecture” (specific example of the phrase B) of the primary query, and “architecture” Is different from “Scotland” (specific example of the phrase C) of the primary query different from “” and is adopted as the phrase of the secondary query. Similarly, the phrase “resident vote” (specific example of the phrase A) is co-occurring with the phrase “Scotland” (specific example of the phrase B) of the primary query, and the phrase “politics” (phrase C) of the primary query. It is adopted as a phrase of the secondary query because it is semantically close to the specific example. On the other hand, the phrases “narrow”, “minimal”, “whiskey”, “military song” (specific examples of the phrase A) are meanings in addition to their co-occurring counterparts “architecture”, “Scotland”, “politics”. Since it does not have a primary query phrase that is close to the target, it is not adopted as a secondary query phrase. If the phrase “sake” that is semantically close to the phrase “whiskey” of the secondary query candidate exists in the primary query, the phrase “whiskey” may be adopted as the secondary query. As shown in FIG. 12, the content obtained as a result of searching using the phrases “Glasgow” and “local referendum” adopted as the secondary query is the potential user's preference (the magnificent architecture remaining in Glasgow). It is also a content that reflects the topic of the current general public (a referendum on Scotland independence).

  FIG. 13 is a diagram illustrating an example of recommended content list data output by the recommendation list generation unit 17. The recommended content list data shown in the figure is data in which the program name, broadcast date and time, and program outline of each recommended content are set.

  FIG. 14 is a diagram illustrating a display example of a recommended content presentation screen that the recommended content presentation unit 18 displays on the content display device. This figure is a GUI (graphic user interface) screen on which the content of the recommended content list shown in FIG. 13 is displayed by a web browser. The “future broadcast schedule”, “system setting 1”, and “system setting 2” tabs displayed at the top of the recommended content presentation screen are respectively a list of unviewed content and the contents of the extended query of the user. And a recommendation list generation unit 17 to display a setting of a weight to be multiplied by the secondary score. These displays are not essential in the present embodiment.

In the above embodiment, the phrase of the primary query is extracted from the user viewing history information. However, the phrase of the primary query may be a keyword input by the user.
Moreover, in the said embodiment, although the phrase of a secondary query candidate is acquired using social data, you may acquire the phrase of a secondary query using other data. Arbitrary data can be used as long as it is computer-usable corpus data that can identify the subject of a topic by time such as a time stamp, etc. Can do.

  In the above description, the case where the content recommendation device 1 and the content display device 3 are connected via a network has been described. However, the content display device 3 may be configured to include the content recommendation device 1. The content display device 3 may be configured to include a part of the functional units of the content recommendation device 1. For example, the content display device 3 may include the viewing history recording unit 11 of the content recommendation device 1, and may further include the unviewed content information recording unit 12 and the viewing history analysis unit 13.

  According to the above-described embodiment, the content recommendation device 1 uses social media provided on the Internet, and has other semantic relationships with a primary query that is a set of search terms describing user preferences. Is extracted as a secondary query phrase. The content recommendation device 1 adds, as a search term, a secondary query word / phrase extracted based on the primary query word / phrase to the primary query data, which is a set of search terms describing user preferences. The content recommendation device 1 searches for content using a set of search terms to which a secondary query term is added. As a result, the content recommendation device 1 can recommend content that conforms to the content requested by the user. In addition, by limiting the time of social media used for the extraction of the words and phrases of the secondary query, the content recommendation device 1 allows the topic of a new topic to appear daily or a past topic in addition to the user's potential preference. Can be recommended.

As described above, according to the content recommendation device 1 of the present embodiment, automatically generating a query (extended query) that appropriately reflects the diversity of phrases and current topics of topics from the original search terms. Can do. And the content recommendation apparatus 1 can implement | achieve the content recommendation which matched the user preference rather than before by searching a content using the produced | generated query.
In addition, according to the content recommendation device 1 of the present embodiment, it is possible to recommend a variety of content based on phrases (extended queries) having a deep semantic relationship with the original search word. As a result, it is possible to recommend content that is difficult to find from only the original search terms and has a high possibility of finding new discoveries and discoveries of the user (serendipity).

  The content recommendation device 1 and the content display device 3 described above have a computer system inside. The process of operation of the content recommendation device 1 and the content display device 3 is stored in a computer-readable recording medium in the form of a program. The computer system reads and executes this program, so that the above processing is performed. Is called. The computer system here includes a CPU, various memories, an OS, and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

1 content recommendation device 11 viewing history recording unit 12 unviewed content information recording unit 13 viewing history analysis unit 14 social data recording unit 15 social data analysis unit 16 query expansion unit 17 recommendation list generation unit 18 recommended content presentation unit 19 storage unit 3 content Display device 31 Operation unit 32 Acquisition unit 33 Output unit 34 Notification unit 35 Reception unit 5 Social media service providing device 9 Network

Claims (6)

An acquisition unit that acquires primary query data indicating a list of search terms that are terms used in the search;
A search word candidate extraction unit that extracts words of a search word candidate from a plurality of corpus data in which the same topic can be described by different notations;
The search in which the similarity with any of the search terms included in the primary query data is higher than a predetermined condition and higher than the predetermined condition from the search term candidate words A query expansion unit that selects and adds to the search terms a search term candidate phrase that co-occurs in the corpus data with any of the search terms different from a word;
A search unit for searching for content using the search term included in the primary query data and the search term added by the query expansion unit;
A content recommendation device comprising: The search word candidate extraction unit extracts words of the search word candidate from the corpus data within a predetermined period.
The content recommendation device according to claim 1. The acquisition unit acquires primary query data including words extracted from text information related to content viewed by a user.
The content recommendation device according to claim 1, wherein the content recommendation device is a content recommendation device. The acquisition unit acquires primary query data composed of phrases extracted from text information related to a portion of content reproduced by a user,
The content recommendation device according to any one of claims 1 to 3, wherein The search word candidate extraction unit extracts words of the search word candidate from a tag or text of the corpus data.
The content recommendation device according to any one of claims 1 to 4, wherein the content recommendation device is characterized in that: Computer
An acquisition means for acquiring primary query data indicating a list of search terms that are terms used in the search;
Search word candidate extraction means for extracting words of a search word candidate from a plurality of corpus data in which the same topic can be described by different notations;
The search in which the similarity with any of the search terms included in the primary query data is higher than a predetermined condition and higher than the predetermined condition from the search term candidate words Query expansion means for selecting and adding to the search terms a search term candidate phrase that co-occurs in the corpus data with any of the search terms different from a word;
Search means for searching for content using the search terms included in the primary query data and the search terms added by the query expansion means;
For functioning as a content recommendation device.

Images