JP2013077044A - Credibility analysis device and credibility analysis method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To appropriately determine the credibility of a document.SOLUTION: A credibility analysis device 1 configured as a device for analyzing the credibility of a document includes: a document storage part 20 for storing a plurality of documents; a classification part 11 for specifying a reference relation showing that a document of a reference source has been generated by referring to a document of a reference destination in a plurality of documents, and for detecting that a specific operation relating to the credibility of the document of the reference destination is being performed with respect to the document of the reference source shown by the reference relation; a first number calculation part 14 for calculating the first number as the number of the documents of the reference source to which the specific operation is being performed for each document of the reference destination; a credibility determination part 15 for determining the credibility of the document of the reference destination from the first number; and an output part 16 for outputting the information showing the result of determination.

Description

  The present invention relates to a credibility analysis apparatus and a credibility analysis method for analyzing the credibility of a document.

  In recent years, communication using miniblogs or microblogs has become common. Microblogging is characterized by posting short sentences consisting of tens to hundreds of characters, and users can use it as a simple blog to post casual words and share their situation with other users stand out. You can also share information with other users by posting comments with URLs (Uniform Resource Locators) such as news articles that interest you, so miniblogs are not just a tool to get the status of friends. It is also widely used as an information collection tool. Recently, it is not uncommon for many media to use a Twitter account to distribute information, and it is possible to collect information that interests you by following your favorite media (registering in your browsing list). Become. Such information is automatically pushed (transmitted) from the information posting side to itself.

  Twitter, which is famous for microblogging, has the property that information can be easily propagated and diffused through users via retweets, informal retweets, and posting tweets with hashtags.

  Retweeting is a method often used when sharing information with your followers (users who are registered in the browsing list). According to a general trend, tweets containing more important information tend to be retweeted by many users. The retweeted information is passed to the follower, and the follower retweets further, so that the information propagates through the user. In general, this retweet is often called official retweet.

  Informal retweets are not simply retweeting a user's comments, but retweeting with their own comments. In this case, since it is possible to retweet the original information with its own opinion, it is possible to give additional information or to correct if the information is incorrect. Like official retweets, informal retweets are a common means of spreading information.

  A hash tag is a function for sharing information on a specific topic by tagging a tweet. When the user searches for a hash tag using a search provided by Twitter or a third party, it is possible to easily view the posted content related to the topic. When a user browses a specific hash tag in addition to following the user, information about a specific topic can be collected.

  In the Great East Japan Earthquake, major communication means such as mobile phones and landlines were paralyzed due to the earthquake. While the main communication infrastructure is paralyzed, SNS (Social Networking Services) such as Twitter and Facebook have attracted a great deal of attention for their safety and information sharing. In Facebook and Twitter, your friends and family are often registered in your contact list (equivalent to Friends in Twitter and Facebook), so it is easy to check the safety of people around you through SNS communication. I was able to do it. From the viewpoint of information sharing, people who report the status of the disaster-stricken area on microblogs, such as posting photos of the disaster-stricken area at the time of the earthquake, are conspicuous. It was not uncommon to enter. Communication using Twitter at the time of the earthquake is not limited to the Great East Japan Earthquake. In Eileen, a large hurricane that occurred in August 2011, sharing of disaster information by Twitter was conspicuous. At the time of the earthquake disaster, Internet SNS tools such as Twitter and Facebook are now becoming one of the common communication tools.

  While there is a good aspect in terms of communication at the time of an earthquake disaster utilizing Twitter, there is also a bad aspect that causes inconvenience by flowing uncertain or false information using the characteristics that information is easy to propagate. For example, immediately after the Great East Japan Earthquake, oil factories exploded and radioactive rain fell, celebrity deaths due to the earthquake, criticism of policies, political parties, administrations, etc. . Such low-credibility, unfounded information can be particularly confusing during an earthquake, so a technique to identify low-credibility information is required.

  In Non-Patent Document 1, from the information such as the characteristics of the user who posted the tweet to the tweet (the number of tweets posted, the number of friends, the number of followers), the characteristics of the posted tweet (including the number of characters, URL, or hash tag) A feature was generated as an input to machine learning, and a model was generated by machine learning. Using the created model, the feature amount of the tweet is extracted and input to the model, so that the credibility can be scored for each tweet.

  Further, in Patent Document 1, one or more comments included in a Web page are divided into single comments, and a reputation expression is performed on a keyword (for example, a product name of a company) in which each comment is designated. It describes the technology for extracting comments. The presence or absence of reputation expression is specified by text matching the character contained in the comment with the reputation expression dictionary. By generating a list of comments that express reputation, it is possible to view what comments are commented on that keyword. In Patent Document 2, the same processing as Patent Document 1 is performed, but by collecting the number of comments in which reputation expression is performed in units of time, how much reputation expression is performed per hour for the specified keyword. It is possible to browse.

JP 2004-70405 A JP 2005-63242 A

C. Carlos, M. Marcelo, P.Barbara, "Information Credibility on Twitter", WWW 2011

  However, the techniques described in Non-Patent Document 1, Patent Document 1, and Patent Document 2 have the following problems.

  Since the processing in these techniques is mainly performed for each single comment, the extraction accuracy is low. In Patent Document 1 and Patent Document 2, since processing is performed by text matching with a comment expression dictionary in comment units, there is a problem in the information extraction accuracy of comments expressed in comments. For example, it is not possible to extract a comment expressing a reputation expression using a word not registered in the reputation expression dictionary. There is a limit in extracting reputational expressions from text matches. In the microblog, when applying the technique of Patent Document 1 to a short unit of text such as several tens to hundreds of characters, the amount of information is too small in order to determine whether or not the reputation is performed. In Non-Patent Document 1, statistical user information such as the number of tweets posted by past users, the number of friends, and the number of followers is also used. Not. Therefore, the analysis of credibility is mainly a single tweet unit. Since the user's reaction is not used, it is not possible to accurately specify a comment including a reputation expression by processing using a single comment.

  The present invention has been made in view of the above problems, and the sentence size is not limited. In particular, the credibility of documents (including the comments and tweets described above) posted on the miniblog described above is more appropriate. It is an object of the present invention to provide a credibility analysis apparatus and a credibility analysis method that can be determined in a simple manner.

  In order to achieve the above object, a credibility analysis apparatus according to the present invention is a credibility analysis apparatus that analyzes the credibility of a document, and is stored by a document storage unit that stores a plurality of documents and a document storage unit. A reference relationship specifying unit for specifying a reference relationship indicating that the reference source document is generated by referring to the reference destination document, and the reference relationship specified by the reference relationship specifying unit Specific operation detecting means for detecting that a specific operation related to the authenticity of the referenced document is performed, and a specific operation detected by the specific operation detecting means. Calculating means for calculating a first number, which is the number of reference source documents, for each reference destination document; It comprises a credibility determination means for determining the authenticity of the referenced document from a first number which is calculated by the means and output means for outputting information indicating a result of determination by the authenticity determination means, a.

  In the credibility analysis apparatus according to the present invention, the credibility is determined based on the reference relationship between a plurality of documents and the number of specific operations related to the credibility of the reference destination document with respect to the reference source document. Therefore, the credibility of the reference destination document is determined based on how the reference source document was created or handled according to the reference destination document. That is, the credibility of the reference destination document is determined by reflecting the reaction of the user who created the reference source document to the reference destination document. Therefore, according to the credibility analysis apparatus according to the present invention, it is possible to more appropriately determine the credibility of the document.

  The calculation means calculates a second number, which is the number of reference source documents with respect to the reference destination document, for each reference destination document, and the credibility determination means calculates the reference destination from the first number and the second number. Judging the authenticity of the document. According to this configuration, since the credibility can be determined in consideration of the number of reference source documents for each reference document, it is possible to more appropriately determine the credibility of the document.

  The reference relationship may be at least one of a relationship in which the reference source document is a citation or a reply of the reference destination document. Further, the reference source document is at least one of a reference document that is cited without processing, and a reference document other than the reference document that is cited without processing. Can be. According to these configurations, the reference relationship between documents can be grasped appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  The specific operation related to the authenticity of the reference destination document may include deleting the reference source document. According to these configurations, it is possible to grasp a specific operation for a reference source document appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  The specific operation related to the authenticity of the reference destination document may include that the reference source document is generated as a correction of the content of the reference destination document. More specifically, the specific operation detection unit stores information indicating the corrected expression in advance, and determines whether the corrected expression is included in the reference source document, thereby determining the reference source document. Can be detected as correcting the content of the document to be referred to. According to these configurations, it is possible to grasp the specific operation for the reference source document appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  The specific operation detection means determines whether or not the reference document includes a corrected expression. If the specific operation detection unit determines that the corrected expression is included, the reference source document is the reference document. It is possible not to detect that the content is generated as a correction of the content of. According to this configuration, it is possible to appropriately determine whether the reference source document corrects the content of the reference destination document, and the present invention can be more appropriately implemented.

  The credibility determination means extracts link information indicating a link to another data included in the document determined to have low credibility, and determines whether or not the document includes at least a part of the link information. It is also possible to determine the authenticity of the document. Further, the credibility determination means can determine the credibility of the document by determining whether or not the document includes the domain of the extracted link information. According to these configurations, it is possible to efficiently determine the credibility of a document using the information of the document once determined to have low credibility.

  The reference relationship specifying means can compare the texts of the documents and specify the reference relationship based on the match between the texts. According to this configuration, the reference relationship between documents can be grasped appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  The reference relationship specifying unit specifies the reference relationship by determining whether or not the reference relationship information indicating the reference destination document is added to the reference source document, and the reference relationship information is specified as having the reference relationship. It is possible to compare texts with other documents only for the referenced document. According to this configuration, documents for which credibility is determined can be narrowed down, and efficient processing becomes possible.

  The document is associated with a time, and the reference relationship specifying unit can specify the reference relationship according to the time associated with the document. According to this configuration, it is possible to determine the authenticity of the document for each time zone, and it is possible to perform efficient processing.

  By the way, the present invention can be described as an invention of a credibility analysis apparatus as described above, and can also be described as an invention of a credibility analysis method as follows. This is substantially the same invention only in different categories and the like, and has the same operations and effects.

  In other words, a credibility analysis method according to the present invention is a credibility analysis method by a credibility analysis apparatus that includes a document storage unit that stores a plurality of documents, and that is stored by a document storage unit. In a plurality of documents, a reference relationship specifying step for specifying a reference relationship indicating that a reference source document is generated by referring to a reference destination document, and a reference relationship specified in the reference relationship specifying step A specific operation detection step for detecting that a specific operation related to the authenticity of the reference destination document is performed on the indicated reference source document, and the specific operation detected in the specific operation detection step are performed. Calculate the first number, which is the number of referenced documents, for each referenced document A calculating step for determining, a credibility determining step for determining the credibility of the reference document from the first number calculated in the calculating step, and an output step for outputting information indicating a determination result in the credibility determining step. Including.

  According to the present invention, since the credibility of the reference destination document is determined by reflecting the trust of the user who created the reference source document with respect to the reference destination document, the credibility of the document is more appropriately determined. be able to.

It is a figure which shows the function structure of the reliability analysis apparatus which concerns on embodiment of this invention. It is a figure which shows the data stored in the deletion data table of a classification data storage part. It is a figure which shows the data stored in the spreading | diffusion data table of a classification data storage part. It is a figure which shows the data stored in the document data table of a classification data storage part. It is a figure which shows the data stored in the correction expression storage part. It is a figure which shows the data stored in a black list URL storage part. It is a figure which shows the data stored in a hoax storage part. It is a figure which shows the hardware constitutions of the reliability analysis apparatus which concerns on embodiment of this invention. It is a flowchart which shows the process (credibility analysis method) performed with the credibility analysis apparatus which concerns on embodiment of this invention. It is a flowchart which shows a document classification | category process among the processes (credibility analysis method) performed with the credibility analysis apparatus which concerns on embodiment of this invention.

  Hereinafter, embodiments of a credibility analysis apparatus and a credibility analysis method according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 shows a credibility analyzer 1 according to the present embodiment. The authenticity analysis device 1 is a device that analyzes the authenticity of a document. Specifically, it is an apparatus that determines whether or not the credibility of one or more documents to be analyzed is low (whether or not the content of the document is not trustworthy). The document to be analyzed is, for example, a document posted by a user and published on a miniblog or microblog posted on the Web. In this embodiment, in order to simplify the description, Twitter, which is a representative of a microblog, is used as a specific example. In this embodiment, the analysis target is referred to as a document, but depending on the miniblog or microblog service, it is also referred to as a tweet or a comment. The analysis target document does not necessarily need to be a document published on the Web.

  In the case of a microblog such as Twitter, a referential relationship in which one document is generated by referring to another document is formed by retweet or informal retweet for quoting or replying to a document. In the present embodiment, the authenticity of the document is analyzed using this reference relationship.

  In microblogs like Twitter, it is possible to spread information by retweets and informal retweets. However, due to the easy diffusion of information, tweets with low credibility such as hoax tweets could be easily spread, and such tweets were actually spread immediately after the Great East Japan Earthquake. There was a tendency for unofficial retweets calling such tweets to be hoaxes, and for their retweets by users who retweeted to include hoax information. In the present embodiment, the document that corrects the content of such a document or the deletion of the document is assumed to be that the specific operation related to the authenticity of the reference destination document is performed on the reference source document. Used for analysis. Note that only one of the correction and deletion described above may be used as the specific operation.

  In the present embodiment, the reference relationship is at least one of a relationship in which a reference source document (a document posted later) is a citation or a reply of a reference destination document (a document posted earlier). In addition, the reference document is a reference document that is quoted without being processed (equivalent to the official retweet in Twitter, or equivalent to Facebook in Facebook), and other than the reference document and the reference relationship It is at least any one of having.

  Documents that are quoted without processing, that is, spread documents, are considered to be documents that the user finds more useful as the number of spreads increases. In addition, a document on which a comment is posted by the user such as his / her claim, that is, a shared document is considered to be a document that the user thinks is more useful as the number of shared documents increases.

  As shown in FIG. 1, the credibility analysis apparatus 1 includes an analysis server 10 that performs document credibility analysis processing, a document storage unit 20 that is a database that holds data used for analysis, and a classification data storage unit 30. A corrected expression storage unit 40, a blacklist URL storage unit 50, and a hoax storage unit 60. The credibility analysis apparatus 1 is connected to a device (for example, a server providing a microblog service) that outputs a document to be analyzed (acquired) so as to acquire (receive) the document to be analyzed via a network such as the Internet. . Below, the data stored in each data storage part 20-60 are demonstrated first.

  The document storage unit 20 stores a plurality of documents to be analyzed for credibility in advance. Acquisition of a document stored in the document storage unit 20 is acquired (reception), for example, when the analysis server 10 provides a microblog service via the Internet and requests a server that stores the document to acquire the document. It is also possible to receive document data by streaming from the server. The document is, for example, a comment unit generated by the user. Each document in Twitter corresponds to each tweet data.

  The data of each document stored in the document storage unit 20 includes text data indicating the content of the document. This text indicates, for example, the content of a post created by the user. In addition, an ID that uniquely identifies the document data is assigned. This ID may be given in advance when the document is acquired, or a unique ID may be given by the analysis server 10 when the document is acquired. Further, data indicating the time may be added to the document data. This time is, for example, the time when a document is posted (or generated) to a server such as a miniblog by a user.

  The document data may be provided with reference relationship information indicating the reference destination document, for example, information indicating the ID of the reference destination document. As described above, for example, when a document relates to retweets, information specifying a tweet that is a retweet source (reference destination of retweets) may be given. Further, the information indicating the reference document indicates information indicating how the reference is performed, for example, the spread data generated by quoting the reference document without being processed. A flag may be given. The information indicating the ID of the reference destination document and the flag indicating the spread data are given to the document when the reference source document is created or posted, for example.

  The deleted document data may be provided with a deletion flag indicating that the document has been deleted so that it can be determined that the document has been deleted. This deletion is to delete what the user has once posted in a microblog service or the like. Moreover, when the deletion flag is given, the information which shows the time deleted further may be given. The deleted document data can also be acquired from a server that provides a microblog service, for example, via Twitter's API. The deletion flag is given when a document is deleted. Note that the data indicating that the document has been deleted (deletion flag data) may be acquired independently of the text of the deleted document.

  The classification data storage unit 30 classifies and stores the documents stored in the document storage unit 20. As will be described later, this classification is performed by the analysis server 10, and each data is stored in the classification data storage unit 30 by the analysis server 10. The classification data storage unit 30 includes a deletion data table 31, a diffusion data table 32, and a document data table 33.

  The deleted data table 31 stores deleted document data among the document data stored in the document storage unit 20. As shown in FIG. 2, the deletion data table 31 stores the ID of the deleted document and the deletion date and time corresponding to the deletion time in association with each other.

  The diffusion data table 32 stores data of a document obtained by diffusing another document (quoted without processing another document) among the document data stored in the document storage unit 20. Such a document (diffusion data) is equivalent to official retweet in Twitter, and is equivalent to Like in Facebook. As shown in FIG. 3, in the spread data table 32, the ID of the spread document (reference source document), the spread date and time that is the posting time, and the ID of the reference destination document correspond to each other. Attached and stored.

  The document data table 33 stores document data other than the spread data (reference source document) among the document data stored in the document storage unit 20. The document data table 33 is given a corrected expression flag that is information indicating whether or not each document includes a corrected expression. As shown in FIG. 4, the document data table 33 is associated with an ID of a posted document, a posting date and time that is the posting time, a posted text that is the content of the document, and a correction expression flag. Stored. The corrected expression flag indicates that the document includes a corrected expression when the value is “1”, and indicates that the document does not include a corrected expression when the value is “0”. However, the correction expression flag is not necessarily in the above format. The various tables 31 to 33 described above may include data of items other than those described above. In addition, data may be stored in a format other than a format called a table.

  The above information may be obtained by subdividing data by time. As a result, it is possible to refer to only the data necessary when processing the data for each time period, and it is possible to reduce the cost for the processing time.

  The corrected expression storage unit 40 stores (stores) information indicating the corrected expression in advance. The corrected expression is an expression indicating that the content of the document is incorrect. Specifically, as shown in FIG. 5, keywords (text) such as “doma” and “don't be fooled” are stored. Information indicating the corrected expression stored in the corrected expression storage unit 40 is input in advance by an administrator of the credibility analysis apparatus 1 or the like. Information indicating the corrected expression is used for determination of authenticity.

  The blacklist URL storage unit 50 stores a URL that is link information indicating a link to another data included in a document that is a document with low credibility. This URL is stored in the blacklist URL storage unit 50 by the analysis server 10 as will be described later. This URL is used for determination (pre-processing) of authenticity.

  When the URL included in the document is a shortened URL, the URL may be converted into an expanded URL and stored. Specifically, a URL as shown in FIG. 6 is stored. The URL to be stored may be a URL with parameters, a URL without parameters, a domain, or a part of a URL. Since the posting is often performed with a shortened URL on a microblog, the stored URL may be a shortened URL or an expanded URL.

  As described later, the hoax storage unit 60 stores information indicating a document that has been determined to be a document with low reliability by the analysis server 10. Information on a document with low credibility is used for determination (pre-processing) of credibility. As illustrated in FIG. 7, the hoax storage unit 60 includes an ID of a document whose reliability is low, a posted text that is the content of the document, and a creation date that is the time when the document is created (posted). Are stored in association with each other. The hoax storage unit 60 described above may include data of items other than those described above.

  Subsequently, the configuration of the analysis server 10 will be described. The analysis server 10 includes a classification unit 11, a preprocessing unit 12, a second number calculation unit 13, a first number calculation unit 14, a credibility determination unit 15, and an output unit 16. The The analysis server 10 may include means for acquiring document data and storing it in the document storage unit 20 in addition to the above-described components.

  The classification unit 11 is means for reading out the document to be analyzed stored in the document storage unit 20, classifying it, and storing it in the respective tables 31 to 33 of the classification data storage unit 30. The classification unit 11 performs a classification process for each read document. First, the classification unit 11 confirms whether or not a document has been deleted. Specifically, for example, this confirmation is performed by determining whether or not a deletion flag is given to document data. If the classification unit 11 determines that the document has been deleted, the classification unit 11 acquires the document ID and the deletion date and time that is the deletion time from the document data, and associates them with the classification data storage unit. 30 stored in the deletion data table 31.

  Further, the classification unit 11 confirms whether or not the document is spread data obtained by spreading the reference destination document. Specifically, for example, this confirmation is performed by determining whether a flag indicating that the data is spread data is added to the document data. If the classification unit 11 determines that the document is spread data, the classification unit 11 acquires the ID of the document, the spread date and time that is the posting time, and the ID of the reference document from the document data, and Are stored in the diffusion data table 32 of the classification data storage unit 30 in association with each other.

  Note that the determination of whether a document is spread data may be made by comparing the texts of the documents and determining whether or not they completely match (the document with the earlier submission time is the reference document, The document with the later time is the source document).

  For a document that is determined not to be spread data, the classification unit 11 acquires information indicating the corrected expression stored in the corrected expression storage unit 40, and determines whether or not those corrected expressions are included in the document. to decide. Specifically, a text match is performed between the text of the document and each corrected expression related to information stored in the corrected expression storage unit 40.

  The classification unit 11 acquires the ID of the document that is determined not to be spread data, the posting date and time that is the posting time, and the posting text that is the content of the document from the data included in the document, and associates them with each other The data is stored in the document data table 33 of the classification data storage unit 30. Further, a correction expression flag (“1” if the document includes the corrected expression, “0” if the document does not include the correction expression) corresponding to the determination result of whether or not the correction expression is included is also stored in the document data table 33 in association with it.

  As described above, the classification unit 11 specifies that the document is spread data, and the reference source document is generated by referring to the reference destination document among the plurality of documents stored by the document storage unit 20. This is a reference relationship specifying means (one function) for specifying a reference relationship indicating that the information has been processed.

  Furthermore, the classification unit 11 detects that the document may be deleted or include a corrected expression, that is, may be generated as a document that corrects the content of another document (reference destination document). In other words, the classification unit 11 is a specific operation detection unit (one function) that detects that a specific operation related to the credibility of the reference document is performed (possibly) on the document.

  The classification of the document by the classification unit 11 is performed, for example, at the timing when the document authenticity is analyzed. This timing is, for example, a timing triggered by the operation of the administrator of the credibility analyzer 1 or a preset timing (for example, every predetermined time).

  The classification unit 11 may perform classification for analyzing document credibility (that is, specifying a reference relationship between documents) according to the time associated with the documents. For example, classification may be performed using only documents for each preset time period. That is, the document credibility analysis may be performed using only a document for each preset time period. As a result, processing such as text matching, which will be described later, is not required for old data that is not required for credibility analysis.

  The pre-processing unit 12 is a means for filtering a document that does not require credibility analysis. The unnecessary document referred to here is a document that does not require credibility determination by the credibility determination unit 15. The pre-processing unit 12 performs the filtering described above using document data that has been determined to be low in credibility by the credibility determination unit 15 in the past. A document whose credibility analysis is not required by the above filtering is determined to have a low credibility in the same way as a document whose reliability is determined to be low by the credibility determination unit 15 in the past. It is excluded from the subsequent analysis. That is, the pre-processing unit 12 is a credibility determination unit (one function) that determines the credibility of a document using the credibility determination by the past credibility determination unit 15.

  Specifically, the pre-processing unit 12 acquires document data from the document data table 33 of the classification data storage unit 30 as document data that can be analyzed. The pre-processing unit 12 acquires the document data stored in the hoax storage unit 60 and collates it with the data of the document that can be analyzed, and the document that can be analyzed is stored in the hoax storage unit 60. Check whether it matches the document data. This collation is performed between document IDs or between texts. The comparison between texts may be a partial match instead of a complete match. Alternatively, matching at the morpheme level may be performed using morpheme analysis. If it is determined by the above collation that the document that can be analyzed matches the data of the document stored in the hoax storage unit 60, the preprocessing unit 12 determines that the document that can be analyzed has low credibility. Thus, the credibility determination unit 15 does not determine the credibility.

  Further, the pre-processing unit 12 acquires the URL data stored in the blacklist URL storage unit 50 and collates it with the data of the document that can be analyzed, and the document that can be analyzed is at least one of the URLs. It is judged whether the part is included. If it is determined in the above determination that the document that can be analyzed includes the URL stored in the blacklist URL storage unit 50, the preprocessing unit 12 determines that the document that can be analyzed has low credibility. Thus, the credibility determination unit 15 does not determine the credibility.

  Note that if the document that can be analyzed is not at least a part of the URL but includes the entire URL, the document may be determined to have low credibility. Further, when a document that can be analyzed includes a URL domain, it may be determined that the document has low credibility. Further, as described above, the URL to be compared may be either or both of the shortened URL and the decompressed URL.

  The pre-processing unit 12 notifies the second number calculation unit 13 and the first number calculation unit 14 of information other than the document for which the determination of credibility is not performed by the above-described processing. The credibility is determined for the notified document. As will be described later, the determination of authenticity is performed on a reference document (a document referenced from another document) among the documents to be analyzed that are not excluded from the determination of authenticity. This is performed by using information of a reference source document having a reference relationship with the previous document.

  The second number calculation unit 13 is a calculation unit (one function) that calculates, for each reference destination document, a second number that is the number of reference source documents with respect to the reference destination document to be analyzed. As shown in FIG. 1, the second number calculating unit 13 includes a diffusion number calculating unit 13a and a shared number calculating unit 13b.

  The diffusion number calculation unit 13a calculates, as the second number, the number of reference sources (the number of diffusions) that is cited without processing the reference destination document to be analyzed (the reference destination document is diffused). . This calculation can be performed by referring to the diffusion data table 32 of the classification data storage unit 30 and counting the number of IDs of the reference source document in units of IDs of the reference destination documents to be analyzed. The diffusion number calculation unit 13a outputs the calculated diffusion number to the credibility determination unit 15 in association with the ID of the reference destination document to be analyzed.

  The number-of-shares calculation unit 13b uses, as the second number, a document that has a reference relationship other than what is cited without processing the reference destination document to be analyzed, specifically, for example, the reference destination to be analyzed. The number of documents quoted (or replied) with the comments of the referrer's own user attached to the document is calculated (the number of shared documents). This is the number of documents quoting the contents of the referenced document and posting their own comments. It is equivalent to informal retweets in Twitter.

  For such a reference relationship, information indicating the reference relationship is not stored in the classification data storage unit 30 as in the case of a spread document. Therefore, the shared number calculation unit 13b extracts a reference relationship (connection) between documents using data in the document data table 33 of the classification data storage unit 30 illustrated in FIG. The number-of-shares calculation unit 13b compares (collates) the text of the reference destination document to be analyzed included in the document data table 33 with the text of another document posted after the reference destination document.

  For example, in a text of a reference destination document, a predetermined number of characters (for example, ten characters) from the first character is included in the text of another document or collated by text matching. In the case of the data shown in FIG. 4, since the text “Today is sunny” of the document with ID “23450” and the text of the document with ID “89012” are included, the document with ID “89012” includes the document with ID “23450”. It is determined that the document is cited (unofficially retweeted). It should be noted that the text collation location does not necessarily have to be from the first character, and it is only necessary to determine the matching of the character strings (the matching of at least some of the texts) from which the reference relationship can be determined.

  In addition to documents that have been quoted and diffused without processing, reference relationship information indicating the reference destination document is added to the data of the reference source document having a reference relationship such as a citation relationship, as with the diffused document. If it is, the reference relationship may be determined using the reference relationship information.

  The number-of-shares calculation unit 13b determines the above reference relationship for each reference document to be analyzed as described above, and totals the number of reference source documents determined to have a reference relationship, To do. As described above, in the calculation of the number of sharing cases, the sharing number calculation unit 13b also specifies the reference relationship between documents, and the reference for specifying the reference relationship between the reference source document and the reference destination document. It is also a relationship specifying means (one function). The number-of-shares calculation unit 13b outputs the calculated number of shares to the credibility determination unit 15 in association with the ID of the reference destination document to be analyzed.

  As described above, when a reference relationship is specified by comparing texts for all documents, it takes a very long processing time. Therefore, the processing time can be shortened by narrowing down the reference documents to be analyzed.

  Specifically, the sharing number calculation unit 13b sets only the document whose ID is stored as the ID of the reference destination document in the diffusion data table 32 of the classification data storage unit 30 as the reference destination document to be analyzed. That is, only a document spread by another user, which is a document in which an ID is added as reference relation information indicating a reference destination document to another document, is analyzed (processed).

  Alternatively, only documents with a certain number of reference destination documents from another document indicated by the reference relationship information may be analyzed. As described above, if only the diffused document is set as the analysis target, the number of documents to be processed can be reduced, the processing cost can be reduced, and the reliability analysis is performed by narrowing down to a document having a high degree of attention from the user. be able to. Thereby, real-time processing becomes possible.

  The first number calculation unit 14 is a calculation unit (one function) that calculates a first number that is the number of reference source documents for which a specific operation is performed for each document to be analyzed (reference destination). ). As shown in FIG. 1, the first number calculation unit 14 includes a correction number calculation unit 14a and a deletion number calculation unit 14b.

  The correction number calculation unit 14a calculates, as the first number, the number of reference source documents (correction number) for correcting the contents of the reference target document to be analyzed. The correction number calculation unit 14a specifies the reference relationship between the reference destination document and the reference source document in the same manner as the second number calculation unit 13 (this reference relationship is processed by the second number calculation unit 13). The result may be used). The number-of-corrections calculation unit 14a refers to the correction expression flag of the document data table 33 of the classification data storage unit 30 from the reference source document that has a reference relationship for each reference destination document, and determines the reference source document including the correction expression. The number is counted as the number of corrections. The number-of-corrections calculation unit 14 a outputs the calculated number of corrections to the credibility determination unit 15 in association with the ID of the reference destination document to be analyzed.

  The correction number calculation unit 14a may perform the following processing. The number-of-corrections calculation unit 14a refers to the correction expression flag in the document data table 33 of the classification data storage unit 30, and determines whether or not the reference document to be analyzed includes a correction expression. If the reference document includes a corrected expression, the number of corrections is equivalent to the number of sharing cases. In that case, credibility analysis using the number of corrections is not appropriate. For this reason, when the correction number calculation unit 14a determines that the reference target document to be analyzed includes a corrected expression, the reference source document is generated as correcting the contents of the reference target document. It is not detected as a thing, and the number of corrections is not calculated.

  The number-of-deletions-calculation section 14b has, as a first number, the number of reference-source documents that have a reference relationship with the reference-target document to be analyzed and whose contents are corrected (the number of deletions). Is calculated. The deletion number calculation unit 14b specifies the reference relationship between the reference destination document and the reference source document in the same manner as the second number calculation unit 13 (this reference relationship is processed by the second number calculation unit 13). The result may be used). The number-of-deletions-calculation unit 14b uses the number of reference-source documents deleted by referring to the deletion data table 31 of the classification data storage unit 30 from the reference-source documents having a reference relationship for each reference-destination document as the number of deletions Tally. The number of deletions is the number of reference documents (official retweets) that diffused the referenced document (the number of deletions for the number of diffusions), the number of other reference documents (unofficial retweets) (the number of deletions for the number of shares), It is also possible to calculate the sum of them. The deletion number calculation unit 14b outputs the calculated deletion number to the credibility determination unit 15 in association with the ID of the reference destination document to be analyzed.

  As described above, in calculating the number of corrections or the number of deletions, the correction number calculation unit 14a and the deletion number calculation unit 14b also specify the reference relationship between documents, and the reference source document and the reference destination document It is also a reference relationship specifying means (one function) for specifying the reference relationship between the two.

  Note that the reference relationship used by the sharing number calculation unit 13b, the correction number calculation unit 14a, and the deletion number calculation unit 14b is referred to in advance by the document data table 33 before the sharing number, correction number, and deletion number are calculated. May be specified and stored (for example, in a shared document storage unit (not shown)). As a result, the subsequent processing can be performed smoothly. That is, retweet-related information between tweets, which is referred to as Twitter, is extracted in advance and applied to subsequent processing. In addition, this data may be data for each time zone, similar to the data stored in the classification data storage unit 30.

  The credibility determination unit 15 uses the first number calculated by the first number calculation unit 14 and the second number calculated by the second number calculation unit 13 as a reference destination document to be analyzed. It is a credibility judgment means for judging the credibility of the. Specifically, the credibility determination unit 15 stores in advance an expression and a threshold value for determining credibility, and makes a determination based on the formula.

Specifically, if the corrected document is not included in the reference destination document, the determination is made using any or all of the following four expressions. The authenticity determination unit 15 determines that a document satisfying the relationship of the following expression is a document with low authenticity (the same applies to the following).
Corrections / Diffusions> α
Number of deletions / spreads> β
Number of corrections / shares> θ
Number of deleted / shared items> π
Here, α, β, θ, and π are preset threshold values. As the number of deletions, either the number of deletions with respect to the number of diffusions or the number of deletions with respect to the number of sharing may be used according to the formula (the same applies to the following).

Further, as described above, when the corrected document is included in the reference destination document, the number of corrections is not calculated, and therefore determination is made using either or all of the following two expressions.
Number of deletions / spreads> γ
Number of deleted / shared items> χ
Here, γ and χ are preset threshold values.

Although the above-mentioned judgment criteria are those using the number of diffusion cases and the number of sharing cases, they may be judgment criteria in which they are not used (both may be used). Specifically, if the corrected document is not included in the reference destination document, the determination is made using either or all of the following two expressions.
Number of corrections> α '
Number of deletions> β '
When the corrected document is included in the reference document, it is determined using the following formula.
Number of deletions> γ '
Here, α ′, β ′, and γ ′ are preset threshold values.

  The credibility determination unit 15 outputs the determination result of the reference document to be analyzed to the output unit 16.

  The output unit 16 is an output unit that outputs information indicating the determination result by the credibility determination unit 15. Specifically, the output unit 16 stores in the hoax storage unit 60 information on a document determined to have low credibility. Specifically, as illustrated in FIG. 7, the hoax storage unit 60 associates the ID of a document determined to have low credibility, the text of the document, and the creation date (posted time) of the document. To store.

  Further, the output unit 16 determines whether or not the document that is determined to have low credibility includes a URL. If the URL is included, the output unit 16 displays the URL in a blacklist as illustrated in FIG. It is stored in the URL storage unit 50. Note that when the URL is included in a document determined to have low credibility for a certain number or more, that is, when it frequently appears in a document determined to have low credibility, the blacklist URL storage unit 50 It is good also as storing in. Further, as described above, URL domains, URLs without parameters, shortened URLs, decompressed URLs, and the like can be stored in the blacklist URL storage unit 50.

  As a site having such a URL, there is a news site based on false information such as a fictional newspaper (http://kyoko-np.net/), and the information on this site is frequently used on Twitter. May flow as. At first glance, it seems to be true because the information is shared by the tweet with the URL, but it is important to identify such a site because the contents of the actual site are gathered with lie information.

  Alternatively, the output unit 16 outputs the display so that the administrator of the credibility analysis device 1 can confirm by displaying on the display included in the credibility analysis device 1. The functional configuration of the credibility analysis apparatus 1 has been described above.

  FIG. 8 shows a hardware configuration of the credibility analysis apparatus 1. As shown in FIG. 8, the reliability analysis apparatus 1 includes a CPU (Central Processing Unit) 101, a RAM (Random Access Memory) 102 and a ROM (Read Only Memory) 103, which are main storage devices, a communication module 104 for performing communication, In addition, the computer includes a computer including hardware such as an auxiliary storage device 105 such as a hard disk. The functions of the credibility analysis apparatus 1 described above are exhibited when these components are operated by a program or the like. The above is the configuration of the credibility analysis apparatus 1.

  Subsequently, a credibility analysis method that is a process executed by the credibility analysis apparatus 1 according to the present embodiment will be described with reference to flowcharts of FIGS. 9 and 10. This process is performed, for example, as an operation and a trigger by the administrator of the credibility analysis apparatus 1 or as a regular process at regular intervals.

  First, the analysis server 10 acquires a plurality of documents to be analyzed and stores them in the document storage unit 20 (S01). Subsequently, the document stored in the document storage unit 20 is classified by the classification unit 11 and stored in the classification data storage unit 30. Document classification will be described with reference to the flowchart of FIG.

  First, it is confirmed whether or not a deletion flag is given to document data (S201, specific operation detection step). If it is determined that the document has been deleted, information relating to the document is stored in the deletion data table 31 of the classification data storage unit 30 (S202, specific operation detection step). In this case, the document classification ends here. The above processing is based on the assumption that the document data related to the deletion flag is acquired independently of the deleted document text. In this case, information related to the text of the deleted document is already stored in the diffusion data table 32 or the document data table 33. When acquired as one data including the text of the deleted document and the deletion flag, the processing after S203 may be performed after the processing of S202. If it is not determined in step S201 that the document has been deleted, it is subsequently checked whether the document data has been given a flag indicating that it is spread data (step S203, reference relationship specifying step). . When it is determined that the document is spread data, information related to the document is stored in the spread data table 32 of the classification data storage unit 30 (S204, reference relationship specifying step). In this case, the classification process ends here.

  If it is not determined in S203 that the document is spread data, it is determined whether or not the document includes a corrected expression (S205, specific operation detection step). Subsequently, the document data is stored in the document data table 33 of the classification data storage unit 30 together with a flag indicating whether or not the corrected expression is included (S206, reference relationship specifying step). This completes the classification process. The above classification is performed for all documents used for authenticity analysis.

  When the above classification is completed, the pre-processing unit 12 narrows down the documents to be analyzed (S03 in FIG. 9, credibility determination step). Specifically, the document data stored in the document data table 33 of the classification data storage unit 30 is stored in the document data stored in the hoax storage unit 60 and the black list URL storage unit 50. By comparing with the URL data, it is determined whether or not the document stored in the document data table 33 has low credibility. If it is determined that the document has low credibility, the document is not used for further processing. In the subsequent processing, a document for which it has not been determined that the document has low credibility as a result of the above determination is taken as an analysis target.

  Subsequently, for the document that is the reference destination document, the second number calculation unit 13 calculates the number of diffusion cases and the number of sharing cases as the second number for each document (S04, second number calculation). Step, reference relationship identification step). The first number calculation unit 14 calculates the number of corrections and the number of deletions as the first number for each document as a reference destination document (S05, first number calculation step). , Reference relationship identification step). Note that, as described above, the document to be referred to here may be only the spread document, and the processing for specifying the reference relationship by comparing the text may be reduced.

  Subsequently, the credibility determination unit 15 determines the credibility of the referenced document using the first number and the second number calculated as described above (S06, credibility determination step). . The determination result of credibility is output by the output unit 16 (S07, output step). Specifically, data of a document that is determined to have low credibility is used, information on the document is accumulated in the hoax storage unit 60, and information in the blacklist URL storage unit 50 is updated. The above is the credibility analysis method that is the process executed by the credibility analysis apparatus 1 according to the present embodiment.

  As described above, according to the present embodiment, the credibility is determined based on the reference relationship between a plurality of documents and the number of specific operations related to the credibility of the reference destination document with respect to the reference source document. Therefore, the credibility of the reference destination document is determined based on how the reference source document was created or handled according to the reference destination document.

  Specifically, when a user who creates (posts) a reference source document thinks that the reference destination document has low credibility, such as a hoax, other users can use the hoax by the reference destination document. The reference document is corrected in the reference source document or the reference source document is deleted so as not to be confused. As described above, according to the present embodiment, the response of the user who created the reference source document to the reference destination document is reflected, and the credibility of the reference destination document is determined. Therefore, according to the present embodiment, it is possible to more appropriately determine the authenticity of the document.

  The determination of credibility may be performed only from the number of corrections and deletions that are the first number, but may be performed using the number of diffusions and the number of sharings that is the second number. According to this configuration, since the credibility can be determined in consideration of the number of reference source documents for each reference document, it is possible to more appropriately determine the credibility of the document. More specifically, the credibility can be more appropriately determined based on the ratio of correction or deletion.

  In addition, as in this embodiment, the reference relationship is quoted or replied, quoted without processing (diffusion, equivalent to official retweet in Twitter), or other (equivalent to informal retweet in Twitter), etc. Also good. According to these configurations, the reference relationship between documents can be grasped appropriately and reliably, and the present invention can be implemented appropriately and reliably. However, as long as there is a reference relationship that can be grasped other than the above, an arbitrary one may be used.

  Further, the specific operation related to the authenticity of the reference document can be corrected or deleted as in the present embodiment. According to these configurations, it is possible to grasp a specific operation for a reference source document appropriately and reliably, and the present invention can be implemented appropriately and reliably. However, as long as there is a specific operation related to the credibility of the document that can be grasped other than the above, an arbitrary one may be used.

  Further, in the case of correction, considering the case where the reference document itself includes a correction expression as in this embodiment, it is appropriately determined whether the reference source document corrects the content of the reference destination document. This makes it possible to judge more appropriate credibility.

  Further, if the document to be analyzed is narrowed down using the information of the document once determined to be low as in the processing of the pre-processing unit 12 in the present embodiment, the reliability of the document is efficiently improved. Judgment can be made.

  Also, by determining the document processing target according to the time associated with the document, it is possible to determine the credibility of the document for each time zone, and for efficient processing. Real time credibility analysis becomes possible.

DESCRIPTION OF SYMBOLS 1 ... Reliability analysis apparatus, 10 ... Analysis server, 11 ... Classification part, 12 ... Pre-processing part, 13 ... 2nd number calculation part, 13a ... Diffusion number calculation part, 13b ... Shared number calculation part, 14 1st Number calculation unit, 14a ... Correction number calculation unit, 14b ... Deletion number calculation unit, 15 ... Reliability determination unit, 16 ... Output unit, 20 ... Document storage unit, 30 ... Classification data storage unit, 31 ... Delete data table, 32 ... diffusion data table, 33 ... document data table, 40 ... correction expression storage section, 50 ... blacklist URL storage section, 60 ... hoax storage section, 101 ... CPU, 102 ... RAM, 103 ... ROM, 104 ... communication module, 105 ... Auxiliary storage device.

Claims (14)

A credibility analyzer that analyzes the authenticity of a document,
A document storage unit for storing a plurality of documents;
A reference relationship specifying means for specifying a reference relationship indicating that a reference source document is generated by referring to a reference destination document in the plurality of documents stored by the document storage unit;
Specific operation detecting means for detecting that a specific operation related to the credibility of the reference destination document is performed on the reference source document indicated by the reference relation specified by the reference relation specifying means;
Calculating means for calculating, for each reference destination document, a first number that is the number of the reference source documents on which the specific operation detected by the specific operation detection means is performed;
Credibility determination means for determining the credibility of the referenced document from the first number calculated by the calculation means;
Output means for outputting information indicating a determination result by the credibility determination means;
A credibility analyzer with The calculating means calculates a second number, which is the number of the reference source documents with respect to the reference destination document, for each reference destination document,
The authenticity determining means determines the authenticity of the document referred to from the first number and the second number;
The credibility analyzer according to claim 1.   The credibility analysis apparatus according to claim 1, wherein the reference relationship is at least one of a relationship in which a reference source document is a citation or a reply of a reference destination document.   The reference source document is at least one of those cited without processing the reference destination document and those having a reference relationship with the reference destination document other than quoting the reference destination document without processing. The credibility analyzer according to claim 3.   The credibility analysis apparatus according to claim 1, wherein the specific operation related to the credibility of the reference destination document includes deleting the reference source document.   The specific operation related to the authenticity of the reference destination document includes that the reference source document is generated as correcting the content of the reference destination document. The credibility analyzer described.   The specific operation detection unit stores information indicating a corrected expression in advance, and determines whether the reference source document includes the corrected expression by determining whether the reference source document includes the corrected expression. The credibility analysis apparatus according to claim 6, wherein the credibility analysis apparatus detects that the document is generated to correct the content of the document.   The specific operation detecting means determines whether or not the corrected expression is included in the reference destination document, and when determining that the corrected expression is included, the reference source document is The credibility analysis apparatus according to claim 7, wherein the credibility analysis apparatus does not detect that the content of the reference destination document is generated as being corrected.   The credibility determination means extracts link information indicating a link to another data included in the document determined to have low credibility, and determines whether or not the document includes at least a part of the link information. The credibility analysis apparatus according to any one of claims 1 to 8, wherein the credibility of the document is also determined by performing the process.   The credibility analysis apparatus according to claim 9, wherein the credibility determination unit determines the credibility of a document also by determining whether or not the document includes a domain of the extracted link information.   The credibility analysis apparatus according to any one of claims 1 to 10, wherein the reference relationship specifying unit compares the texts of the documents and specifies the reference relationship based on a match between the texts.   The reference relationship specifying means specifies the reference relationship by determining whether or not the reference relationship information indicating the reference destination document is added to the reference source document, and there is a reference relationship based on the reference relationship information. The credibility analysis apparatus according to claim 11, wherein the texts are compared with other documents only for the reference destination document identified as. The document is associated with a time,
The reference relationship specifying means specifies the reference relationship according to the time associated with the document;
The credibility analyzer according to any one of claims 1 to 12. A credibility analysis method using a credibility analysis device including a document storage unit for storing a plurality of documents for analyzing the credibility of a document,
A reference relationship specifying step for specifying a reference relationship indicating that a reference source document is generated by referring to a reference destination document in the plurality of documents stored by the document storage unit;
A specific operation detecting step for detecting that a specific operation related to the credibility of the reference destination document is performed on the reference source document indicated by the reference relationship specified in the reference relationship specifying step;
A calculation step of calculating, for each reference destination document, a first number that is the number of the reference source documents on which the specific operation detected in the specific operation detection step is performed;
A credibility determination step of determining the credibility of the referenced document from the first number calculated in the calculation step;
An output step of outputting information indicating a determination result in the credibility determination step;
A credibility analysis method comprising:

Images