JP2010224623A - Method and program for recommending related article - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a related article recommendation method for preventing content from being substantially overlapped, and for precisely recommending a related article matched with tastes/interests of each user as an article related with a certain article which a user is interested. <P>SOLUTION: The related article recommendation method includes; a first step of calculating the weighting value of a featured word in an original article 401 to obtain featured word data; a second step of calculating the weighting value of the featured word in each subscription article 204 to obtain featured word data; a third step of calculating similarity between the featured word data in the original article 401 and the featured word data of each subscription article 204; a fourth step of classifying the related article 402 from the subscription article 204 based on the similarity; a fifth step of calculating the mean value of the weighting value of the featured word in each read article 207 to obtain featured word data indicating user tastes; a sixth step of calculating similarity between the featured word data indicating the user tastes and the featured word data of each related article 402; and a seventh step of preferentially presenting the related article 402 whose similarity is a high order to the user. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a text information filtering technique, and more particularly to a technique effectively applied to a related article recommendation method and a related article recommendation program that automatically recommends other articles having contents related to the contents of a specific article. It is.

  In recent years, with the spread of the Internet and the like, it is possible to obtain a great variety of information using a computer. For example, a lot of information can be collected using various news sites and search engines. Also, a lot of information can be obtained by e-mail or the like. Further, not only on the Internet, for example, a large amount of information can be obtained from various in-house materials that are stored electronically on an in-house server. In order to make effective use of this large amount of electronic information, users can not only find information that matches their interests with high accuracy, but also other related information that they are not aware of by themselves. It is necessary to have an information filtering mechanism that can discover the information of the user and support the spread of the user's interest.

  In response to such a request, for example, a technology for presenting similar articles as related articles by determining similarity of article contents using natural language processing or the like for information such as news articles that the user is interested in Several have been proposed. In such a technology, in order to maintain the efficiency when the user refers to related articles, articles that are substantially the same as the contents of the original article referenced by the user and that overlap are excluded from the related articles. It is necessary not to present it. For this reason, articles with substantially overlapping contents are specified, so that the user can recognize them by excluding them or listing them together in a list.

  As a technique for presenting such a related article, for example, in Japanese Patent Laid-Open No. 9-101990 (Patent Document 1), the similarity between articles is calculated by comparing article expressions between articles by natural language processing. Discloses a technology that determines articles presented to the user according to their similarity and related articles related to the article, and classifies a set of articles with high similarity and different information sources as a set of duplicate articles. Has been.

  Further, for example, in Japanese Patent Laid-Open No. 2005-352857 (Patent Document 2), articles that have substantially the same content even though the distribution of words that appear in the classification method such as Patent Document 1 are not similar. In consideration of the fact that it may not be possible to identify a set of articles, the difference in the date and time of sending is smaller than a pre-registered threshold for multiple articles that contain keywords that represent topics or the like that the user wants to grasp trends, and the senders are different from each other A technique for identifying an article as an article having substantially the same content is disclosed.

Japanese Patent Laid-Open No. 9-101990 JP 2005-352857 A

  When the user refers to and uses information such as news articles of interest, as described above, the user desires to discover unexpected related articles that match his interests, interests, and preferences. However, with the techniques described in Patent Documents 1 and 2, it is possible to identify duplicate articles when presenting related articles, but for which user the related articles presented for any article are However, it is the same, and it is not possible to recommend related articles that match the preferences and interests of each user.

  Accordingly, an object of the present invention is to provide a related article recommendation method that accurately recommends related articles that do not substantially overlap as articles related to articles that the user is interested in, and that match the preferences and interests of each user, and The related article is to provide a recommended program. The above and other objects and novel features of the present invention will be apparent from the description of this specification and the accompanying drawings.

  Of the inventions disclosed in this application, the outline of typical ones will be briefly described as follows.

  According to a related article recommendation method according to a typical embodiment of the present invention, a related article whose contents are related to an original article composed of text data referred to by a user is recommended by the computer system. The computer system holds, for each user, a plurality of subscribed articles that are a reference target of the user and a reference history for each subscribed article, and a predetermined extraction condition from the original article One or more words are extracted as feature words based on the feature word, and for each of the extracted feature words, a weight value indicating the importance of the feature word in the original article is calculated based on a predetermined calculation condition, and the original word is calculated. A first step of using as feature word data of an article, and extracting one or more words as feature words from each of the subscribed articles based on the predetermined extraction condition; For each feature word, based on the predetermined comparison condition, a second step of calculating the weight value of the feature word in each subscribed article based on the predetermined calculation condition and making it the feature word data of each subscribed article The third step of calculating the similarity between the feature word data of the original article calculated in the first step and the feature word data of each subscribed article calculated in the second step, and the third step The subscribed articles in which the calculated similarity is higher than a predetermined threshold and the difference in issue date and time from the original article is equal to or greater than a predetermined time interval are classified as the related articles, and the target articles calculated in the second step The fourth step of using the feature word data of each subscribed article as the feature word data of each related article, and acquiring the user's read articles based on each subscribed article and the reference history, One or more words are extracted as feature words from the read article based on the predetermined extraction condition, and the feature words in the read articles are extracted based on the predetermined calculation condition for each extracted feature word. The weight value is calculated, the average value of all the read articles is calculated and used as feature word data representing the user's preference, and the fifth step based on the predetermined comparison condition A sixth step of calculating the similarity between the feature word data representing the user's preference calculated in step 4 and the feature word data of each related article classified in the fourth step, and the similarity calculated in the sixth step. And a seventh step of preferentially presenting the related article with a higher degree to the user.

  Among the inventions disclosed in the present application, effects obtained by typical ones will be briefly described as follows.

  According to the exemplary embodiment of the present invention, by using the history of articles that have been referred to in the past for each user, as the article related to the article that the user is interested in, the content does not substantially overlap, In addition, it is possible to accurately recommend related articles that match the preferences and interests of each user.

It is a figure explaining the outline | summary of the example of a processing flow of the related article extraction part when the user in one embodiment of this invention refers to an original article. It is the figure which showed the outline | summary of the structural example of the information collection management system to which the related article recommendation method which is one embodiment of this invention is applied. It is the figure which showed the outline | summary of the table structure example of the database in one embodiment of this invention. It is a figure explaining the example of the process which extracts the characteristic word from the text data of the article in one embodiment of this invention, and calculates characteristic word data. It is a figure explaining the example of the process which calculates the similarity degree of the feature word data of the original article, and the feature word data of each subscription article in one embodiment of this invention. It is a figure explaining the example of the process which classify | categorizes a related article and a duplicate article from the subscription article in one embodiment of this invention. It is a figure explaining the example of the process which calculates the feature word data showing user preference from the read article in one embodiment of this invention. It is a figure explaining the example of the process which calculates the similarity degree of the feature word data showing the user preference in one embodiment of this invention, and the feature word data of each related article.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Note that components having the same function are denoted by the same reference symbols throughout the drawings for describing the embodiment, and the repetitive description thereof will be omitted.

  An information collection management system to which a related article recommendation method according to an embodiment of the present invention is applied, may include text information such as various websites, e-mails, and electronic documents (hereinafter collectively referred to as “articles”). ) Is collected in one place and can be referred to by each user, the action history of each user using and referring to the information is accumulated, and the accumulated action history is reused to recommend valuable information for each user. It is a system that provides a mechanism that allows the accumulated information to be used productively.

  In this embodiment, when referring to articles such as news that the user is interested in from the subscribed articles collected based on the designation of each user's subscription, the related articles that match the user's interests and preferences are extracted. By presenting and recommending, it is possible to improve the possibility of finding an unexpected related article that matches the user's preference, and to support the spread of the user's interest and interest. Moreover, when presenting a related article, by excluding duplicate articles whose contents substantially overlap from the related articles, it is possible for the user to efficiently perform information utilization activities.

[System configuration]
FIG. 2 is a diagram showing an outline of a configuration example of an information collection management system to which a related article recommendation method according to an embodiment of the present invention is applied. The information collection management system includes an information collection management server 100 and a database 200. The database 200 may be mounted on the information collection management server 100 or may be mounted on a device such as another database server.

  The information collection management server 100 collects text data as new articles from, for example, a web site on the web server 310 or an in-house document server 320 by the new article collection unit 110. As a method for collecting newly arrived articles, for example, an RSS (RDF Site Summary) feed can be used. By the new article collection unit 110 operating as an RSS reader, not only articles such as news on the website but also electronic data such as in-house materials stored in the document server 320, e-mails, etc. are converted to RSS. Can be collected.

  The RSS feed information is stored in the user subscription information 203 of the database 200 as a subscription designation 206 that is information about a website or the like that each user wants to subscribe for each user who uses the information collection management system of the present embodiment. Retained as part. New arrival articles collected by the new arrival article collection unit 110 based on the contents of the subscription designation 206 are stored as articles 202 in the article group 201 of the database 200.

  In the user subscription information 203, information specifying the article 202 collected based on the subscription designation 206 of the target user among the articles 202 is held as a subscription article 204. That is, all articles 202 collected based on the subscription designation 206 of all users are stored in the article group 201, and the subscribed articles 204 that are the reference targets of each user are stored in the user subscription information 203, respectively. Is equivalent to The article 202 and the subscribed article 204 include not only those collected by the subscription designation 206 such as an RSS feed but also those registered by the user for later reference by directly clipping the text data. .

  In the user subscription information 203 of the database 200, a history that the target user actually referred to the subscription article 204 is further stored as a reference history 205. Based on the subscribed article 204 and the reference history 205, it is possible to identify a read article that the target user has referred to in the past.

  The information collection management server 100 that has received a request to refer to an article of interest (original article 401) from the user via the client terminal 400 acquires the original article 401 from the subscribed article 204 and does not illustrate it. It is presented to the client terminal 400 via a Web server program or the like. Further, the information collection management server 100 determines the degree of similarity between the contents (feature words) of each subscribed article 204 and the contents (feature words) of the original article 401 by the related article extraction unit 120, thereby relating to the original article 401. The related article 402 having the content to be extracted is extracted and similarly presented to the client terminal 400.

  At this time, by determining the degree of similarity between the content (feature word) of the entire read article of the target user and the content (feature word) of each related article 402, the user's interest / preference is more enhanced among the related articles 402. Suggested relevant articles 402 are recommended. Also, duplicate articles whose contents are substantially duplicated are excluded so that they are not presented as related articles 402.

  The new article collection unit 110 and the related article extraction unit 120 are implemented as software programs that run on the information collection management server 100, and can be implemented as, for example, applications that run on a Web server (not shown). The related article extraction unit 120 includes, for example, a feature word data calculation unit 121, a related article classification unit 122, and a similarity calculation unit 123, and the functions of the related article extraction unit 120 as described above are described in detail later. Realize.

[Table structure]
FIG. 3 is a diagram showing an outline of a table configuration example of the database 200. The table stored in the database 200 includes, for example, user information 210, subscription information 220, feed list 230, reaction information 240, and article data 250. The arrows between the tables in the figure indicate that when A → B, for example, A: B = 1: n (A has many Bs).

  The user information 210 is a table that holds information about each user, and includes items such as a user ID, a password, and a user name. The user information 210 is used at the time of authentication when the user logs in to the information collection management system. The feed list 230 is a table that holds basic information of an RSS feed for automatically collecting various information (articles), and includes items such as a feed ID, a site name, and a URL (Uniform Resource Locator). As described above, not only the RSS feed provided from the news site on the Web server 310 but also various electronic documents and e-mails stored in the document server 320 are converted to RSS so that the information can be changed. The newly arrived article collection unit 110 can automatically circulate and collect.

  The subscription information 220 is a table that holds information on which RSS feed each user subscribes to, and includes items such as a user ID, a feed ID, a subscription start date and time, and an unread number. The subscription information 220 corresponds to the subscription designation 206 in FIG. Each user can subscribe to multiple RSS feeds, and each RSS feed can be subscribed to from multiple users.

  The article data 250 is a table that holds the contents of collected articles such as articles included in each RSS feed, articles clipped from the website by the user, and electronic documents. For example, article data, feed ID, issue date, article It has items such as contents. The article data 250 corresponds to the article 202 in FIG. 2, and the entry of the article data 250 specified by the feed ID value of the subscription information 220 corresponds to the subscribed article 204 in FIG. The issue date / time item represents the date / time when the article 202 specified by the article ID item is issued or transmitted, and the article content item is specific text data of the article 202 specified by the article ID item. It is.

  The reaction information 240 is a table that holds information (action history) about what reaction each user has performed on each subscribed article 204. For example, the user ID, article ID, reference date, tag, memo, high It has items such as a light range. The item of article ID and reference date corresponds to the reference history 205 in FIG. As a reaction to the subscribed article 204, in addition to referring to the subscribed article 204, for example, to the subscribed article 204, a tag for classification indicating the content is added, a text memo is added, an arbitrary range, etc. Can be highlighted, and these contents can be held in tags, memos, and highlight range items, respectively. The items in each table described above are examples, and other items may be included.

[Processing flow]
FIG. 1 is a diagram illustrating an outline of a processing flow example of the related article extraction unit 120 when the user refers to the original article 401. When the information collection management server 100 receives a reference request for the original article 401 by the user from the client terminal 400 via a Web server program (not shown) or the like, the related article extraction unit 120 starts a process of extracting the related article 402.

  First, the feature word data calculation unit 121 extracts one or more words as feature words from the text data of the original article 401 by natural language processing. Further, a TF-IDF value (Term Frequency-Inverse Document Frequency), which will be described later, is calculated for each extracted feature word and used as the feature word data of the original article 401 (step S101). Similarly, the feature word data calculation unit 121 extracts one or more words as feature words from the text data of all the subscribed articles 204 by natural language processing. Further, a TF-IDF value is calculated for each extracted feature word and used as feature word data of each subscribed article 204 (step S102).

  Next, the similarity calculation unit 123 converts the feature word data of the original article 401 calculated in step S101 and the feature word data of each subscribed article 204 calculated in step S102 into a vector space as described later. Further, the inner product of the vector for the feature word data of the original article 401 and the vector for the feature word data of each subscribed article 204 is obtained, and the feature word data of the original article 401 and each subscribed article 204 are calculated based on the inner product value. The similarity with the feature word data is calculated (step S103).

  Next, the related article classifying unit 122 converts the subscribed article 204 in which the similarity calculated in step S103 is higher than a predetermined threshold and the difference between the date and time of issue with the original article 401 is equal to or greater than a predetermined time interval. And the feature word data of each target subscribed article 204 calculated in step S102 is used as the feature word data of each related article 402 (step S104). At this time, the duplicate articles whose contents calculated in step S103 are higher than a predetermined threshold and whose difference in issue date and time from the original article 401 is smaller than a predetermined time interval are substantially the same. Classify as 403 (step S104).

  Next, the related article extraction unit 120 acquires the user's already read article 207 based on the user's subscription article 204 and the reference history 205. Further, the feature word data calculation unit 121 extracts one or more words as feature words from the text data of all the read articles 207 by natural language processing. Further, a TF-IDF value is calculated for each extracted feature word, and an average value of TF-IDF values in all the read articles 207 is calculated for each feature word to obtain feature word data representing the user's preference (step) S105).

  Next, similarly to step S103, the feature word data representing the user's preference calculated in step S105 and the feature word data of each related article 402 classified in step S104 are respectively converted into vector spaces by the similarity calculation unit 123. To do. Furthermore, the inner product of the vector about the feature word data representing the user's preference and the vector about the feature word data of each related article 402 is obtained, and the feature word data representing the user's preference based on the inner product value and each related article The similarity with the feature word data 402 is calculated (step S106). Finally, the related articles 402 are rearranged in the descending order of similarity calculated in step S106 and presented to the user, and the process ends (step S107).

[Calculating article feature word data]
FIG. 4 is a diagram for explaining an example of processing in the feature word data calculation unit 121 for extracting feature words from article text data and calculating feature word data. The processing here includes the above-described feature word data calculation processing of the original article 401 (step S101), feature word data calculation processing of each subscribed article 204 (step S102), and feature word data calculation processing representing user preferences. (Step S105).

  First, compound nouns are extracted from the text data of the target article by morphological analysis. Morphological analysis is generally performed in natural language processing, and various morphological analysis engines and software are provided, and these may be used. In the process of calculating feature words representing user preferences (step S105), as will be described later, compound nouns are extracted from all read articles 207 and merged to form feature words.

  Next, for each extracted compound noun, a TF-IDF value that is a weighting value for the importance of the word in the article is calculated. It is common practice to extract characteristic words (words regarded as important) in a sentence based on TF-IDF values. The feature word data is calculated using this method also in the related article recommendation method of the present embodiment. However, the feature word data calculation method is not limited to this, and can be evaluated as a numerical value (weighting value) for each word. Any technique can be used.

  Here, TF (Term Frequency) is the frequency of appearance of a word (compound noun) in an article, and it is considered that the larger this value, the better this word represents the feature of this article. The TF value of a certain word t in a certain article D is expressed by the following expression, normalized by the number of types m and logarithms of the word in the article D, for example, where the appearance frequency of the word t in the article D is f.

  Even if a word has a large TF value, a word that frequently appears in many articles is often a general word that is not a word representing the characteristics of a specific article. Here, IDF (Inverse Document Frequency) is the reciprocal of the number of articles in which the target word appears, and the larger this value, the smaller the number of articles in which this word appears. This word well represents the characteristics of a specific article. It is thought that. The IDF value of a certain word t is expressed by the following expression normalized by the number of all articles N, for example, where Df is the number of documents in which the word t appears in all articles.

  Since the word t having a large value of both TF and IDF is considered to represent the feature of the document D truly well, the TF-IDF value of the word t is obtained by converting the product of TF and IDF to an integer. It is expressed by the following formula.

  This TF-IDF value is calculated for each feature word (compound noun) for the target article. If the target article is different, the feature words included in the article are different, and even if the same feature word is used, the appearance frequency in the article is different. Therefore, the TF-IDF values are different from each other. In Equation 2, the total number of articles N is the total number of articles 202 including subscriptions of other users because the accuracy increases as the number of target articles increases. Further, the number of documents Df in which the word t appears can be obtained by performing a full text search on the entire article 202 using the word t as a key.

  In this embodiment, even if the article is a short sentence, the content is reflected in the feature word data, and the feature word is extracted from the article so that the similarity can be properly compared. Although all the combined nouns are targeted for feature words, those having a small TF-IDF value may be excluded from the feature words based on a predetermined condition. Various formulas for calculating the TF value and the IDF value have been proposed, and not only the above formula but also an appropriate formula can be used according to accuracy, processing time, and the like.

[Calculation of similarity between original article and subscription article]
FIG. 5 is a diagram for explaining an example of processing (step S103) for calculating the similarity between the feature word data of the original article 401 calculated in step S101 and the feature word data of each subscribed article 204 calculated in step S102. is there. When calculating the similarity, the similarity calculation unit 123 uses the feature word data (TF-IDF value) of the original article 401 calculated in steps S101 and S102 by the processing in the feature word data calculation unit 121 described above and each Each feature word data (TF-IDF value) of the subscribed article 204 is converted into a vector space. Further, a vector space method is used in which similarity is calculated by using the inner product of the feature word data vector of the original article 401 and the feature word data vector of each subscribed article 204.

If the number of merging all feature words contained in the feature word data of the feature word data and the subscription article 204 of the original article 401 was of n, the feature word data d _o of the original article 401 vector V (d _o) Further, the vector V (d ₁ ), V (d ₂ ),... Of the feature word data d ₁ , d ₂ ,... Of each subscribed article 204 has TF-IDF values of n feature words as elements. Expressed as an n-dimensional vector. At this time, for feature word data that does not include the target feature word, the value of the element of the vector corresponding to the feature word is zero.

Here, the direction of the vector is considered to represent the feature of the target article. Therefore, the similarity between the feature word data of the original article 401 and the feature word data of each subscribed article 204 is the angle between the vector V (d _o ) and the vectors V (d ₁ ), V (d ₂ ),. It can be expressed in smallness. That is, when the angle between two vectors is θ, it can be said that the closer the cos θ is to 1, the higher the similarity. cos θ is obtained by dividing the inner product of two vectors by the absolute value of each vector. Accordingly, the similarity sim (d _a , d _b ) of the feature word data d _a and d _b is expressed by the following equation using the two vectors V (d _a ) and V (d _b ).

Using the formula 4, the similarity between the feature word data of the original article 401 and the feature word data of each subscribed article 204 is calculated. That is, sim (d _o , d ₁ ), sim (d _o , d ₂ ),. When the similarity with the feature word data of the original article 401 is calculated for all the subscribed articles 204, the subscribed articles 204 are rearranged in descending order of the calculated similarity. In the example of FIG. 5, it is shown that the similarities are rearranged in the order of “subscription article 3”, “subscription article 1”, and “subscription article 2”. In the present embodiment, the similarity between the original article 401 and each subscribed article 204 is calculated using the vector space method. However, the present invention is not limited to this, and the similarity may be calculated using another calculation method. .

[Classification of related articles and duplicate articles]
FIG. 6 is a process for classifying the related articles 402 and the duplicate articles 403 from the subscribed articles 204 based on the similarity between the feature word data of the original article 401 calculated in step S103 and the feature word data of each subscribed article 204 ( It is a figure explaining the example of step S104).

  First, the related article classifying unit 122 extracts only those articles whose similarity is higher than a predetermined threshold among the subscribed articles 204 rearranged in order of higher similarity in step S103. Next, for each of the extracted subscription articles 204, it is determined whether or not the difference in issue date and time from the original article 401 is equal to or greater than a predetermined time interval, and the subscribed articles 204 that are equal to or greater than the predetermined time interval are set as related articles 402. Classify and classify subscription articles 204 that are smaller than a predetermined time interval as duplicate articles 403. As described in Patent Document 2 and the like, it is highly possible that articles having similar contents and similar publication dates are substantially the same. Therefore, these articles are classified as duplicate articles 403 and excluded from related articles 402.

  The predetermined threshold and time interval are preferably held in the information collection management server 100 by an appropriate means such as the database 200 or a definition file so that the setting can be changed as appropriate according to the situation. In the present embodiment, the predetermined threshold is set to 0.2 as shown in FIG. In the case of the similarity calculation method using the vector space method in step S103 described above, it was experimentally confirmed that the related article 402 can be extracted with a practically satisfactory accuracy by setting the predetermined threshold value to 0.2. ing.

[Calculation of feature word data representing user preferences]
FIG. 7 is a diagram illustrating an example of processing (step S105) for calculating feature word data representing user preferences from the read article 207. First, read articles 207 of the target user are acquired based on the subscribed articles 204 and the reference history 205 of the target user, and all the read articles 204 are characterized by the processing in the feature word data calculation unit 121 described above. Each word data is calculated.

  Next, the average value of the feature word data (TF-IDF value) of each feature word in the entire read article 207 is calculated. Here, each feature word is obtained by merging feature words extracted from all the read articles 207. At this time, for a read article 207 that does not include a certain feature word, the TF-IDF value of the feature word in the read article 207 is zero. In this embodiment, the average value of the TF-IDF values is simply obtained by dividing the total value of the TF-IDF values of the feature words in all the read articles 207 by the number of the read articles 207. However, the method of calculating the average value is not limited to this, and for example, a weighted average may be calculated according to a predetermined condition.

  The average value of the TF-IDF values calculated here is data representing the characteristics of the content of the entire read article 207, that is, the preference of the target user, and this is feature word data representing the user preference. . Thus, by storing the reference history 205 for each user and using it, information representing the user's preference / interest can be acquired.

[Calculation of similarity between user preferences and related articles]
FIG. 8 is a diagram for explaining an example of processing (step S106) for calculating the similarity between the feature word data representing the user preference calculated in step S105 and the feature word data of each related article 402 classified in step S104. is there. The similarity calculation method is the same as the calculation method in the similarity calculation processing in step S103 shown in FIG.

  That is, the feature word data representing the user preference calculated in step S105 by the similarity calculation unit 123 and the feature word data of each related article 402 classified in step S104 (feature of each target subscribed article 204 calculated in step S102). Word data) and vector spaces. Further, a vector space method is used in which the similarity is calculated by using the inner product of the feature word data vector representing the user preference and the feature word data vector of each related article 402.

  When the similarity between the feature word data representing the user preference and the feature word data of the related article 402 is calculated for all related articles 402 in the same procedure as the calculation method in step S103, the calculated similarities are related in descending order. The articles 402 are rearranged. In the example of FIG. 8, it is shown that the similarities are rearranged in the order of “related article 3”, “related article 1”, and “related article 2”. By presenting the related articles 402 thus rearranged to the user, the related articles 402 that match the user's interests and preferences can be recommended to the user.

  In this embodiment, the related articles 402 are preferentially presented and recommended by rearranging and displaying the related articles 402 in order from the highest similarity. However, the presenting method is not limited to this, and various methods such as highlighting by changing the character color or font can be employed. In this embodiment, the duplicate article 403 is excluded from the related articles 402. However, when the related articles 402 are presented to the user, for example, the duplicate articles 403 are collectively displayed as a list only. Then, the user may be able to recognize.

  As described above, according to the related article recommendation method of the present embodiment, the user is interested and referred to by using the reference history 205 of the subscribed article 204 referred to in the past held for each user. It is possible to recommend a related article 402 that matches the preference and interest of each user with respect to the original article that is being processed. Accordingly, it is possible to improve the possibility of finding an unexpected related article 402 that matches the user's preference, and to support the spread of the user's interest and interest.

  In addition, when the related article 402 is presented, the user can efficiently perform information utilization activities by excluding the duplicate articles 403 whose contents substantially overlap from the related articles 402. Furthermore, as the number of already read articles 207 (reference history 205) for each user is accumulated by continued use, the accuracy of the recommendation of the related article 402 that matches the user's preference improves, and the user uses information more effectively. It becomes possible to carry out activities.

  As mentioned above, the invention made by the present inventor has been specifically described based on the embodiment. However, the present invention is not limited to the embodiment, and various modifications can be made without departing from the scope of the invention. Needless to say.

  The present invention can be used for a related article recommendation method and a related article recommendation program that automatically recommend other articles having contents related to the contents of a specific article.

DESCRIPTION OF SYMBOLS 100 ... Information collection management server, 110 ... New article collection part, 120 ... Related article extraction part, 121 ... Feature word data calculation part, 122 ... Related article classification | category part, 123 ... Similarity degree calculation part,
200 ... Database, 201 ... Article group, 202 ... Article, 203 ... User subscription information, 204 ... Subscription article, 205 ... Reference history, 206 ... Subscription designation, 207 ... Read article, 210 ... User information, 220 ... Subscription information, 230 ... feed list, 240 ... reaction information, 250 ... article data,
310 ... Web server, 320 ... Document server,
400: Client terminal, 401: Original article, 402: Related article, 403: Duplicate article.

Claims (8)

A related article recommendation method for recommending to a user a related article whose contents are related to an original article consisting of text data that is referred to by a user by a computer system,
The computer system holds, for each user, a plurality of subscribed articles that are referred to by the user, and a reference history for each subscribed article,
One or more words are extracted as feature words from the original article based on a predetermined extraction condition, and the importance level of the feature word in the original article is indicated based on a predetermined calculation condition for each extracted feature word A first step of calculating a weight value to be feature word data of the original article;
One or more words are extracted as feature words from each subscribed article based on the predetermined extraction condition, and for each extracted feature word, the feature word in each subscribed article is extracted based on the predetermined calculation condition. A second step of calculating the weighting value to be feature word data of each subscribed article;
A third step of calculating a similarity between the feature word data of the original article calculated in the first step and the feature word data of each subscribed article calculated in the second step based on a predetermined comparison condition; ,
The subscribed articles in which the similarity calculated in the third step is higher than a predetermined threshold and the difference in issue date and time from the original article is equal to or greater than a predetermined time interval are classified as the related articles, and in the second step A fourth step in which the feature word data of each subscribed article to be calculated is the feature word data of each related article;
The user's read articles are acquired based on the respective subscribed articles and the reference history, and one or more words are extracted as feature words from all the read articles based on the predetermined extraction condition, and extracted. For each of the feature words, the weight value of the feature word in each read article is calculated based on the predetermined calculation condition, and the average value of all the read articles is calculated to obtain the user's preference. A fifth step for representing feature word data;
Based on the predetermined comparison condition, a similarity level between the feature word data representing the user's preference calculated in the fifth step and the feature word data of each related article classified in the fourth step is calculated. 6 steps,
And a seventh step of preferentially presenting the related article having a higher similarity calculated in the sixth step to the user. In the related article recommendation method according to claim 1,
In the fourth step, the subscribed article, in which the similarity calculated in the third step is higher than a predetermined threshold and a difference in issue date and time from the original article is smaller than a predetermined time interval, is added to the original article. Categorized as a duplicate article with substantially overlapping content,
In the seventh step, the related article recommendation method, wherein the duplicate article is presented to the user so that the user can recognize the duplicate article. In the related article recommendation method according to claim 1 or 2,
The weighting value is a TF-IDF value calculated from the TF value calculated for the feature word for the article including the feature word and the IDF value calculated for the subscribed articles of all the users. Related article recommendation method characterized by being. In the related article recommendation method according to any one of claims 1 to 3,
The predetermined comparison condition in the third step and the sixth step is that the feature word data for which the similarity is to be calculated is converted into a vector space, and the similarity is calculated based on an angle formed by both vectors. Related article recommendation method characterized by being. A related article recommendation program for causing a computer system to function to present a related article whose contents are related to an original article consisting of text data referred to by a user and recommend the user,
The computer system holds, for each user, a plurality of subscribed articles that are referred to by the user, and a reference history for each subscribed article,
The related article recommendation program extracts one or more words as feature words from the original article based on a predetermined extraction condition, and for each extracted feature word, the original article in the original article based on a predetermined calculation condition An eighth step of calculating a weighting value indicating the importance of the feature word and using it as the feature word data of the original article;
One or more words are extracted as feature words from each subscribed article based on the predetermined extraction condition, and for each extracted feature word, the feature word in each subscribed article is extracted based on the predetermined calculation condition. A ninth step of calculating the weighting value to be feature word data of each subscribed article;
A tenth step of calculating a similarity between the feature word data of the original article calculated in the eighth step and the feature word data of each subscribed article calculated in the ninth step based on a predetermined comparison condition; ,
Classifying the subscribed articles in which the similarity calculated in the tenth step is higher than a predetermined threshold and the difference between the issue date and the original article is a predetermined time interval or more as the related articles, and in the ninth step An eleventh step in which the feature word data of each subscribed article to be calculated is set as the feature word data of each related article;
The user's read articles are acquired based on the respective subscribed articles and the reference history, and one or more words are extracted as feature words from all the read articles based on the predetermined extraction condition, and extracted. For each of the feature words, the weight value of the feature word in each read article is calculated based on the predetermined calculation condition, and the average value of all the read articles is calculated to obtain the user's preference. A twelfth step for representing feature word data;
Based on the predetermined comparison condition, a similarity level between the feature word data representing the user's preference calculated in the twelfth step and the feature word data of each related article classified in the eleventh step is calculated. 13 steps,
And a 14th step of preferentially presenting the related article having a higher similarity calculated in the 13th step to the user. In the related article recommendation program according to claim 5,
In the eleventh step, the subscribed article, in which the similarity calculated in the tenth step is higher than a predetermined threshold and a difference in issue date and time from the original article is smaller than a predetermined time interval, is added to the original article. Categorized as a duplicate article with substantially overlapping content,
In the fourteenth step, the related article recommendation program, wherein the duplicate article is presented to the user so that the user can recognize the duplicate article. In the related article recommendation program according to claim 5 or 6,
The weighting value is a TF-IDF value calculated from the TF value calculated for the feature word for the article including the feature word and the IDF value calculated for the subscribed articles of all the users. Related article recommendation program characterized by being. In the related article recommendation program according to any one of claims 5 to 7,
The predetermined comparison condition in the tenth step and the thirteenth step is to calculate the similarity based on an angle formed by the vector of the feature word data for which the similarity is to be calculated, respectively. Related article recommendation program characterized by being.

Images