JP6566755B2 - Analysis target speech extraction apparatus, analysis target speech extraction method, and computer program - Google Patents

Description

The present invention relates to an analysis target speech extraction apparatus , an analysis target speech extraction method, and a computer program.

2. Description of the Related Art In recent years, a technique for analyzing comment text posted by an SNS (Social Networking Service) user using SNS is known. For example, non-patent documents 1, 2, and 3 relating to Twitter (registered trademark) are known.
In Non-Patent Document 1, constrained clustering is applied to a relatively short comment sentence after preprocessing such as mitigation of word notation fluctuation is performed.
In Non-Patent Document 2, ranking of users who make influential statements is performed using the PageRank algorithm.
Non-Patent Document 3 proposes a method for determining whether a user is oriented to information transmission based on the user's follow relationship or a method for determining based on the frequency registered in the “favorites” of the user's comment text. .

  Moreover, in patent document 1, in social media, how much a user's utterance propagated is measured, and the user who made the utterance widely propagated is extracted as an influencer.

JP 2012-078933 A

Satoshi Aoshima, Naoki Fukuda, Shohei Yokoyama, Hiroshi Ishikawa, “Realization of constrained clustering for microblogging”, DEIM2010, 2010 J. J., Jianshu Weng, Ee-peng Lim, Q. He, “Twitter-rank: Finding topic-sensitive influential twitterers”, Web Information and Data Management, 2010. Atsushi Tanaka, Takashi Tajima, “Proposal of a Twitter Twitter Classification Method”, DEIM2010, 2010 J. Eisenstein, et.al., “A Latent Variable Model for Geographic Lexical Variation”, EMNLP2010, [online], [Search June 25, 2015], Internet <URL: http: //www.cs.cmu .edu / ~ nasmith / papers / eisenstein + oconnor + smith + xing.emnlp10.pdf> Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean, “Efficient Estimation of Word Representations in Vector Space”, In Proceedings of Workshop at ICLR, 2013.

  However, the above-described conventional technology cannot extract users whose ideas are biased.

The present invention has been made in consideration of such circumstances, and an analysis that can extract a group of users whose ideas are biased from users who participated in a conference such as an electronic conference room or SNS. It is an object to provide a target speech extraction device , an analysis target speech extraction method, and a computer program.

(1) According to one aspect of the present invention, a user comment extraction unit that groups comment sentences posted by a user for each user, and user feature information that represents a user's characteristics using a group of comment sentences for each user A user feature extraction unit that generates a user, a user classification unit that groups users based on user feature information of each user, a topic width evaluation unit that evaluates the topic width of each user group, and the topic of each user group A deflection group information output unit that outputs deflection group information, which is information of a user group having a width that is evaluated to be narrow among the widths, and excludes user comments included in the deflection group information from a plurality of user comments. An analysis target message extraction device including a message extraction unit that extracts an analysis target message .
(2) According to an aspect of the present invention, a user comment extraction unit that groups comment sentences posted by a user for each user, and user feature information that represents a user's characteristics using a group of comment sentences for each user A user feature extraction unit that generates a user, a user classification unit that groups users based on user feature information of each user, a topic width evaluation unit that evaluates the topic width of each user group, and the topic of each user group The topic center evaluation unit that evaluates the difference between the center and the center of the topic of the entire user, the user group having a width that is evaluated as narrow within the topic width of each user group, and the center of the topic is the center of the topic of the entire user a deflecting group information output unit for outputting a deflection group information based on the user group that is evaluated to be deviated from, the speech of a plurality of users, the polarized A speech extraction unit to the exclusion of the user's utterance to extract analyzed speech contained in the group information to be analyzed utterance extractor device comprising a.
(3) According to one aspect of the present invention, in the analysis target message extraction device according to (2), the deflection group information output unit is more than each user group evaluated as being deviated from the center of the topic of the entire user. This is an analysis target speech extraction device that weights a user group having a width that is evaluated to be narrower than a topic width of a user group and includes it in deflection group information.
(4) According to one aspect of the present invention, in the analysis target speech extraction device according to any one of (2) and (3), the topic center evaluation unit includes a centroid concerning comment texts of all users and comment texts of each user group It is an analysis object speech extraction apparatus which evaluates the shift | offset | difference of the center of the topic of each user group and the center of the topic of the whole user based on the similarity with the gravity center regarding.
(5) One aspect of the present invention is the analysis object utterance extraction apparatus of (4), the similarity is to be analyzed utterance extractor is cosine distance.
(6) According to one aspect of the present invention, in the analysis target speech extraction device according to any one of (1) to (5), the topic width evaluation unit includes a similarity in a semantic vector space of feature words of each user group It is an analysis object speech extraction apparatus which evaluates the breadth of the topic of each user group based on the average value.
(7) In one aspect of the present invention is the analysis object utterance extraction apparatus of (6), the similarity on the semantic vector space is to be analyzed utterance extractor is cosine distance.

(8) One aspect of the present invention, the analysis target utterance extractor device, and a user comment extracting step of grouping comment text posted by the user for each user, said analyzed utterance extractor, for each user, comment A user feature extraction step for generating user feature information representing a user feature using a group of sentences, and a user classification step in which the analysis target speech extraction device groups users based on the user feature information of each user; the analyzed speech extraction device, a topic width evaluation step of evaluating the width of the topic of each user group, wherein the analysis target utterance extractor device, a narrow and estimated width in the width of the topic of each user group a deflecting group information output step of outputting a deflection group information which is information of the user group, said analyzed utterance extractor From speech of a plurality of users, to be analyzed utterance extraction method comprising: a speech extracting the analyzed speech to the exclusion of the user's utterance included in the deflection group information.
(9) In one aspect of the present invention, the analysis target utterance extractor device, and a user comment extracting step of grouping comment text posted by the user for each user, said analyzed utterance extractor, for each user, comment A user feature extraction step for generating user feature information representing a user feature using a group of sentences, and a user classification step in which the analysis target speech extraction device groups users based on the user feature information of each user; , the deviation of the analyzed speech extraction device, a topic width evaluation step of evaluating the width of the topic of each user group, wherein the analysis target utterance extractor device, the center of the user overall topic topic of each user group commentary and topics central evaluation step of evaluating, wherein the analysis target utterance extractor device, a narrow in width topics for each user group A deflecting group information output step of the center of the user group and topic outputting deflection group information based on the user group that is evaluated to be offset from the center of the whole topic user width that is, from the remarks plurality of users, A speech extraction step for extracting a speech to be analyzed by excluding a user speech included in the deflection group information .

(10) According to one aspect of the present invention, a user comment extraction step of grouping comment sentences posted by a user on a computer for each user and a user's characteristics are expressed using a group of comment sentences for each user. User feature extraction step for generating user feature information, user classification step for grouping users based on the user feature information of each user, topic width evaluation step for evaluating the topic width of each user group, and each user group A deflection group information output step for outputting deflection group information, which is information on a user group having a width that is evaluated as being narrow within the topic width of the topic, and a user's remarks included in the deflection group information from a plurality of user remarks computer for executing a speech extracting the analyzed speech to the exclusion of Which is another program.

(11) According to one aspect of the present invention, a user comment extraction step for grouping comment sentences posted by a user on a computer for each user, and a user's characteristics are expressed using a group of comment sentences for each user. User feature extraction step for generating user feature information, user classification step for grouping users based on the user feature information of each user, topic width evaluation step for evaluating the topic width of each user group, and each user group The topic center evaluation step for evaluating the difference between the topic center of the topic and the center of the topic of the entire user, and the user group and the topic center of the width evaluated as narrow among the topic width of each user group A deflection group that outputs deflection group information based on a user group that is evaluated to be off the center of the topic. And up information output step, the statements of the plurality of users, a computer program for executing a speech extracting the analyzed speech to the exclusion of the user's utterance included in the deflection group information.

  According to the present invention, it is possible to extract a group of users whose ideas are biased from users who participated in a conference such as an electronic conference room or SNS.

It is a block diagram which shows the deflection group extraction apparatus 1 which concerns on one Embodiment of this invention. It is a figure which shows the structural example of the feature word set with respect to the user group. It is a figure for demonstrating the evaluation method of the width of a topic which concerns on one Embodiment of this invention. It is a figure for demonstrating the determination method of the deflection group which concerns on one Embodiment of this invention. It is a flowchart of the deflection | deviation group extraction method of one Embodiment of this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a configuration diagram showing a deflection group extracting apparatus 1 according to an embodiment of the present invention. In FIG. 1, the deflection group extraction apparatus 1 includes a user comment extraction unit 11, a user feature extraction unit 12, a user classification unit 13, a topic width evaluation unit 14, a topic center evaluation unit 15, and a deflection group information output unit 16.

  A plurality of SNS comments 101 are input to the deflection group extraction apparatus 1. The SNS comment 101 is text data of one comment sentence posted by the SNS user using the SNS. The SNS comment 101 includes user identification information (user ID) of the poster. Only one SNS comment 101 acquired from one SNS may be input to the deflection group extraction apparatus 1, or SNS comments 101 acquired from a plurality of SNSs may be input.

  The user comment extraction unit 11 groups a plurality of SNS comments 101 input to the deflection group extraction apparatus 1 for each user based on the user ID of each SNS comment 101. Thereby, the comment group 102 which collected only the SNS comment 101 of the same user ID is produced | generated for every user ID. The user comment extraction unit 11 outputs the comment group 102 of each user ID to the user feature extraction unit 12.

  Note that, among the SNS comments 101, the SNS comment 101 of the comment text posted only by quoting the comment text of another person may be deleted as not subject to grouping. The comment text posted only by quoting the comment text of another person does not necessarily reflect the characteristics of the contributor's own way of thinking. For this reason, since it may become a noise when investigating a user whose idea is biased, the SNS comment 101 of the comment text posted only by quoting the comment text of another person is deleted from the grouping target. Is preferred. As comment text posted only by quoting the comment text of another person, for example, only re-posting of the comment text of another person or link information to the comment text of another person (for example, URL (Uniform Resource Locator)) Comment texts that contain the

  The user feature extraction unit 12 uses the comment group 102 to generate user feature information representing the user features for each user ID. The user feature information is preferably information that represents a feature of the user's way of thinking. In the present embodiment, the user feature information is a feature vector Vi represented by the following equation.

Vi = k ₁ w ₁ + k ₂ w ₂ +...
However, Vi is a feature vector of the user whose user ID is ui. w _x is a word included in the comment group 102 whose user ID is ui. k _x is the appearance frequency of the word w _x in the comment group 102 whose user ID is ui. x is an integer of 1 or more.

  The feature vector Vi is based on the appearance frequency of words used in the comment text posted by the user whose user ID is ui. It can be considered that the comment text posted by the user reflects the way of thinking of the user. Therefore, the feature vector Vi is preferable as information representing the features of the concept of the user whose user ID is ui.

Note that the word w _x constituting the feature vector Vi is preferably one word representing synonyms and synonyms. Further, the total number of words w _x constituting the feature vector Vi (the maximum value of x) may be constant for all user IDs, or may be different for each user ID. For example, in each user ID, the feature vector Vi may be configured only from the word w _x whose appearance frequency is a certain value or more.

  The user feature extraction unit 12 outputs the feature vector Vi of each user ID (ui) to the user classification unit 13.

The user classifying unit 13 groups users based on the feature vector Vi of each user ID (ui). In the present embodiment, the user classifying unit 13 clusters the feature vectors Vi of each user ID (ui). The user classifying unit 13 generates a user group composed of users having a user ID (ui) determined to be preferably included in the same group based on the clusters generated by clustering. Among the unit vectors (words w _x ) of the feature vectors Vi of all user IDs (ui) belonging to the user group, the unit group topic is a unit that appears prominently in the user group without appearing in other user groups. A vector (word w _x ) is selected. For user grouping, for example, LDA (Latent Dirichlet Allocation) described in Non-Patent Document 4 may be used.

  Here, an example of a method for selecting a user group topic will be described. For convenience of explanation, a target user group for which a topic is selected is referred to as a target user group. In this example, based on Akaike's Information Criterion (AIC), a unit vector (word) that appears prominently in the target user group is selected.

First, an index Info (Gi, tj) is calculated. The index Info (Gi, tj) is an index representing how useful the word tj is for the discrimination of the user group Gi. The word tj is the word w _x included in the feature vector Vi of the user ID (ui) belonging to the user group Gi. The index Info (Gi, tj) is calculated by the following index calculation method (steps S101 to S103).

(Step S101) Using the word set U of the target user group Gi, the following four types of frequencies n11, n12, n21, n22 are calculated for each word tj included in the word set U. The word set U includes all words w _x included in the feature vector Vi of all user IDs (ui) belonging to the target user group Gi.

n11 = the number of SNS comments 101 including the word tj among the SNS comments 101 of the user IDs belonging to the target user group Gi n12 = of the SNS comments 101 not including the word tj among the SNS comments 101 of the user IDs belonging to the target user group Gi Number n21 = Number of SNS comments 101 including the word tj among SNS comments 101 of user IDs belonging to user groups other than the target user group Gi n22 = Of SNS comments 101 of user IDs belonging to user groups other than the target user group Gi Number of SNS comments 101 that do not include the word tj

  (Step S102) For the frequencies n11, n12, n21, and n22 calculated in Step S101, based on the Akaike information criterion, the value MLL_IM (Gi, tj) for the independent model and the value MLL_DM (Gi, tj for the dependent model) ) Is calculated by the following equation.

MLL_IM (Gi, tj) = (n11 + n12) log (n11 + n12)
+ (N11 + n21) log (n11 + n21)
+ (N21 + n22) log (n21 + n22)
+ (N12 + n22) log (n12 + n22)
-2N (log (N))

  MLL_DM (Gi, tj) = n11 (log (n11)) + n12 (log (n12)) + n21 (log (n21)) + n22 (log (n22)) − N (log (N))

  However, “N = n11 + n12 + n21 + n22”.

  (Step S103) Using the value MLL_IM (Gi, tj) and the value MLL_DM (Gi, tj) calculated in step S102, an index Info (Gi, tj) is calculated by the following equation. This index Info (Gi, tj) represents an unreasonable ratio in which the word tj appears biased to the target user group Gi.

AIC_IM (Gi, tj) = − 2 × MLL_IM (Gi, tj) + 2 × 2
AIC_DM (Gi, tj) = − 2 × MLL_DM (Gi, tj) + 2 × 3
Info (Gi, tj) = AIC_IM (Gi, tj) −AIC_DM (Gi, tj)

  Next, based on the index Info (Gi, tj), a unit vector (word) that appears prominently in the target user group Gi is selected. The index Info (Gi, tj) becomes a larger value as the degree that the word tj is useful for discrimination of the target user group Gi based on the Akaike information criterion. From this, from the word tj included in the word set U of the target user group Gi, the words to be the topic of the target user group Gi are selected in descending order of the value of the index Info (Gi, tj).

  The above is an explanation of an example of a method for selecting a user group topic.

  The user classifying unit 13 sets feature words for each user group. As the feature word of the user group, the word selected as the topic of the user group is set. The total number of feature words set for one user group may be constant for all user groups, or may be different for each user group. For example, in each user group, only a word tj whose index Info (Gi, tj) is a certain value or more may be set as a feature word.

  FIG. 2 is a diagram illustrating a configuration example of feature words set for a user group. In the example of FIG. 2, five feature words “story”, “sable”, “think”, “self” and “human” are assigned to the user group whose group identification information (group ID) is “G1”. Was set. In addition, five feature words “vehicle”, “express”, “train”, “departure”, and “formation” are set for the user group whose group ID is “G2”.

  The user classification unit 13 outputs the user group information 103 of each user group to the topic width evaluation unit 14 and the topic center evaluation unit 15. The user group information 103 includes all user IDs belonging to the user group and all feature words of the user group.

  The topic width evaluation unit 14 evaluates the topic width of each user group. Below, the example of the evaluation method of the topic width of a user group is demonstrated.

[Example of topic width evaluation method]
In general, in a user group with a narrow topic width, similar topics are often continuously spoken, so that feature words that tend to co-occur with each other tend to be extracted. For example, the user group with the group ID “G2” illustrated in FIG. 2 is an example of a user group with a narrow topic. In the user group with the group ID “G2”, five feature words “vehicle”, “limited express”, “train”, “departure”, and “formation” are extracted. , “Train”, “departure” and “organization” are considered to be characteristic words that tend to co-occur with each other. For this reason, when each feature word is projected onto the semantic vector space, as shown in FIG. 3A, the feature words of the user group with a narrow topic width are projected at positions close to each other. On the other hand, the user group with the group ID “G2” illustrated in FIG. 2 is an example of a user group with a wide range of topics. In the user group with the group ID “G1”, five feature words “story”, “sable”, “think”, “self”, and “human” were extracted. When these feature words “story”, “sable”, “think”, “self” and “human” are projected onto the semantic vector space, they are projected at positions far from each other as shown in FIG.

  As a method for converting a feature word into a semantic vector, for example, “word2vec” described in Non-Patent Document 5 may be used. For learning the basis of the semantic vector space, the SNS comments 101 of all users may be used, or a general large-scale corpus such as a newspaper may be used. For example, the basis of the semantic vector space may be learned using the word appearance frequency dictionary database 200 shown in FIG. The word appearance frequency dictionary database 200 is generated from a large number of comments posted on media such as newspapers, web sites, and SNSs. The word appearance frequency dictionary database 200 stores dictionary data that can search for the appearance frequency of each word.

  In the present embodiment, the similarity between the two feature words a and b is calculated by the following equation (1).

In the above formula (1), m _ab is the similarity between the feature word a and the feature word b, X _a is the semantic vector of the feature word a, and X _b is the semantic vector of the feature word b. In the formula (1), the molecule is the inner product of mean vector X _b and mean vector X _a, the denominator is the product of length and means the length of the vector X _b sense vector X _a. The similarity m _{ab in} Expression (1) is a cosine distance between the semantic vector X _a of the feature word _a and the semantic vector X _b of the feature word b.

In the present embodiment, the topic width V _j is calculated for the user group Gj by the following equation (2).

However, in the above equation (2), V _j is the topic width of the user group Gj, and m _ab is the above equation (1) of the feature word a and the feature word b included in the user group information 103 of the user group Gj. ), And N is the total number of feature words included in the user group information 103 of the user group Gj. The topic width V _{j in} Expression (2) is an average value of similarities for all pairs of feature words included in the user group information 103 of the user group Gj.

  The topic width evaluation unit 14 calculates the topic width of each user group by the above equation (2). The topic width evaluation unit 14 determines whether or not the calculated topic width of each user group is narrow. For example, the topic width evaluation unit 14 uses the user group information 103 of all user groups, and calculates the topic width combining all user groups by the above equation (2). The topic width evaluation unit 14 compares the topic width of each user group with the topic width of all user groups, and determines whether the topic width of each user group is narrow. The topic width evaluation unit 14 determines that a user group having a topic width satisfying a determination condition that can be said to be narrower than the topic width of all user groups is a user group having a narrow topic width.

  The user group may be a set of users whose topics are similar due to grouping by the user classification unit 13. For this reason, the topic width of each user group is considered to be narrower than the topic width of all user groups. However, since there are usually a user group with a relatively wide topic and a user group with a narrow topic, a user group with a relatively narrow topic may be determined to be a user group with a narrow topic. .

  The above is an explanation of an example of a method for evaluating the topic width of a user group.

  The topic width evaluation unit 14 outputs, to the deflection group information output unit 16, the group ID of the user group that has been evaluated that the topic width is narrow as a result of the topic width evaluation and all user IDs belonging to the user group.

  The topic center evaluation unit 15 evaluates the difference between the topic center of each user group and the topic center of the entire user. Below, the example of the evaluation method of the shift | offset | difference of the center of a topic is demonstrated.

[Example of evaluation method for misalignment of topic center]
In the present embodiment, the centroid vector g _j is calculated for the user group Gj by the following equation (3).

However, in the above equation (3), g _j is the centroid vector of the user group Gj, n is the total number of comments sentences user group Gj. x _i is a language vector of the i-th comment sentence among n comment sentences of the user group Gj. The language vector x _i is expressed by the following equation.
x _i = k ₁ w ₁ + k ₂ w ₂ +...
In the language vector x _i , w _y is a unit vector indicating one word W _y included in the i-th comment sentence. The unit vector w _y is determined for each word that is included in the i-th of the comment sentence. k _y is the number of times the word W _y appears in the i-th comment sentence. Language vector x _i is calculated for each comment sentence of user group Gj.

In the present embodiment, the similarity S _j between the centroid vector g _{all of all} users and the centroid vector g _j of the user group Gj is calculated by the following equation (4).

In the above equation (4), S _j is the similarity between the centroid vector g _{all of all} users and the centroid vector g _j of the user group Gj, and g _j is the centroid of the user group Gj according to the above equation (3). G _all is the centroid vector of all users. The center-of-gravity vector g _all of all users is calculated by the above equation (3), where x _i is the language vector of comment texts of all user groups and n is the total number of comment texts of all user groups. In the formula (4), molecules are inner product of the centroid vector _{g all} the user groups Gj centroid vector _{g j} for all users, the denominator of the centroid vector _{g j} of length and user groups Gj of centroid vectors _{g all} of all users It is a product of length. The similarity S _{j in} equation (4) is a cosine distance between the centroid vector g _{all of all} users and the centroid vector g _j of the user group Gj.

  Note that the SNS comments 101 of all users may be used for calculating the centroid vectors of all users, or a general large-scale corpus such as a newspaper may be used. For example, the centroid vectors of all users may be calculated using the word appearance frequency dictionary database 200 shown in FIG.

  The topic center evaluation unit 15 calculates, for each user group, the similarity between the centroid vector of all users and the centroid vector of the user group by the above equation (4). The topic center evaluation unit 15 selects a user group having a similarity that is equal to or less than a predetermined threshold among the similarities of each user group calculated by the above formula (4), so that the topic center is shifted from the topic center of the entire user. Is determined to be a user group.

  The above is an explanation of an example of a method for evaluating the deviation of the center of a topic.

  The topic center evaluation unit 15 deflects the group ID of the user group evaluated that the topic center is shifted from the center of the topic of the entire user and all user IDs belonging to the user group as a result of the evaluation of the topic center shift. The data is output to the group information output unit 16.

  The deflection group information output unit 16 outputs the deflection group information 104. The deflection group information 104 is information on a user group that is determined to be a group of users whose ideas are biased. A user group determined to be a group of users whose views are biased is referred to as a deflection group. The deflection group information 104 includes a group ID of the deflection group and information 105 of all user IDs belonging to the deflection group. Hereinafter, an example of a deflection group determination method will be described.

  As shown in FIG. 4, it is generally considered that the language space 30 created by the speech of all users and the language space 32 created by the speech of a certain user group are different in topic width and center. For this reason, in this embodiment, it is determined whether the user group is a deflection group by comparing topic widths and centers.

[Example 1 of deflection group determination method]
In Example 1 of the deflection group determination method, only a user group evaluated as having a narrow topic width is set as a deflection group. The group ID of the user group evaluated that the topic width is narrow is input from the topic width evaluation unit 14 to the deflection group information output unit 16. The deflection group information output unit 16 sets only the group ID of the user group input from the topic width evaluation unit 14 as the group ID of the deflection group.

  When the deflection group determination method example 1 is applied, the topic center evaluation result is not used, and therefore the deflection group extraction apparatus 1 may not include the topic center evaluation unit 15.

[Example 2 of deflection group determination method]
In example 2 of the deflection group determination method, only a user group that has been evaluated as having a narrow topic width and that has been evaluated as having a center of topic deviating from the center of the topic of the entire user is a deflection group. The group ID of the user group evaluated that the topic width is narrow is input from the topic width evaluation unit 14 to the deflection group information output unit 16. Further, the group ID of the user group evaluated that the topic center is deviated from the topic center of the entire user is input to the deflection group information output unit 16 from the topic center evaluation unit 15. The deflection group information output unit 16 sets only the group ID of the user group input from both the topic width evaluation unit 14 and the topic center evaluation unit 15 as the group ID of the deflection group.

[Example 3 of deflection group determination method]
In Example 3 of the deflection group determination method, at least only a user group that is evaluated as having a narrow topic width or evaluated that the topic center is deviated from the topic center of the entire user is set as a deflection group. . The group ID of the user group evaluated that the topic width is narrow is input from the topic width evaluation unit 14 to the deflection group information output unit 16. Further, the group ID of the user group evaluated that the topic center is deviated from the topic center of the entire user is input to the deflection group information output unit 16 from the topic center evaluation unit 15. The deflection group information output unit 16 sets only the group ID of the user group input from at least the topic width evaluation unit 14 or the topic center evaluation unit 15 as the group ID of the deflection group.

  When the deflection group determination method example 3 is applied, at least the group ID of the user group input from the topic width evaluation unit 14 is more than the group ID of the user group input from the topic center evaluation unit 15 alone. May be included in the deflection group information 104 with a large weight.

  Next, the deflection group extraction method of this embodiment will be described with reference to FIG. FIG. 5 is a flowchart of the deflection group extraction method of this embodiment.

  (Step S1) The user comment extraction unit 11 groups a plurality of SNS comments 101 input to the deflection group extraction apparatus 1 for each user.

  (Step S2) The user feature extraction unit 12 calculates the feature vector of each user using the grouping result of Step S1.

  (Step S3) The user classifying unit 13 groups users based on the feature vectors of each user.

  (Step S4) The topic width evaluation unit 14 evaluates the topic width for each user group as a result of the grouping in step S3.

  (Step S5) The topic center evaluation unit 15 evaluates the difference between the topic center of each user group and the topic center of the entire user for each user group as a result of the grouping in step S3.

  (Step S6) The deflection group information output unit 16 outputs the deflection group information 104.

  According to the embodiment described above, it is possible to extract a group of users whose ideas are biased from among users who participated in the SNS remarks. Thereby, for example, the following effects can be obtained.

  In recent years, market trends have been investigated by analyzing the content of SNS statements. Here, in the SNS remarks, there is a case where some users perform a large amount of negative opinions with respect to a specific product or company. In such a case, when statistical analysis of the content of a statement by SNS is performed, there is a possibility that a tendency different from the trend of the majority of users may appear. In response to such a problem, according to the present embodiment, based on deflection group information, it is possible to exclude a statement made by a user whose idea is biased from an analysis target. Thereby, the effect of improving the accuracy of the market trend survey based on the content remarked by the SNS can be obtained.

  As mentioned above, although embodiment of this invention was explained in full detail with reference to drawings, the specific structure is not restricted to this embodiment, The design change etc. of the range which does not deviate from the summary of this invention are included.

  In the above embodiment, the SNS is applied to the SNS as a user's utterance. However, the utterance can be applied to various utterances such as an SNS, an electronic conference room, and a blog comment field. In addition, the present invention is not limited to a place where a user speaks online via a communication network, and can also be applied to a place where a user speaks offline. As a place for the user to speak offline, for example, a meeting in which the user speaks orally, a posting column of a paper medium such as a newspaper, a questionnaire response such as a street questionnaire, and the like can be cited. As for comments made in the offline user's utterance or comments made in a document, the user ID is added to the comment text converted into text data and input to the deflection group extraction apparatus 1.

  In the above-described embodiment, the user feature vector is used as the user feature information. However, the user feature information may be information that represents the user's feature, and more preferably, information that represents the feature of the user's way of thinking. The feature vector is not limited to the user.

Further, a computer program for realizing the function of the deflection group extracting apparatus 1 described above is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into the computer system and executed. Also good. Here, the “computer system” may include an OS and hardware such as peripheral devices.
“Computer-readable recording medium” refers to a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a DVD (Digital Versatile Disk), and a built-in computer system. A storage device such as a hard disk.

Further, the “computer-readable recording medium” means a volatile memory (for example, DRAM (Dynamic DRAM) in a computer system that becomes a server or a client when a program is transmitted through a network such as the Internet or a communication line such as a telephone line. Random Access Memory)), etc., which hold programs for a certain period of time.
The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line.
The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

DESCRIPTION OF SYMBOLS 1 ... Deflection group extraction apparatus, 11 ... User comment extraction part, 12 ... User feature extraction part, 13 ... User classification part, 14 ... Topic width evaluation part, 15 ... Topic center evaluation part, 16 ... Deflection group information output part, 200 ... Word appearance frequency dictionary database

Claims (11)

A user comment extractor that groups comment text posted by the user for each user;
For each user, a user feature extraction unit that generates user feature information representing the user's features using a group of comment sentences;
A user classifying unit that groups users based on the user feature information of each user;
A topic width evaluation unit that evaluates the topic width of each user group;
A deflection group information output unit that outputs deflection group information, which is information of a user group having a width that is evaluated to be narrow among topics of each user group;
A utterance extraction unit that extracts utterances to be analyzed by excluding user utterances included in the deflection group information from utterances of a plurality of users;
An analysis object utterance extraction device comprising: A user comment extractor that groups comment text posted by the user for each user;
For each user, a user feature extraction unit that generates user feature information representing the user's features using a group of comment sentences;
A user classifying unit that groups users based on the user feature information of each user;
A topic width evaluation unit that evaluates the topic width of each user group;
A topic-centric evaluation unit that evaluates the difference between the topic center of each user group and the topic center of the entire user;
Deflection that outputs deflection group information based on a user group evaluated as being narrower than the topic center of the entire user and a user group evaluated that the topic center is deviated from the center of the topic of the entire user. A group information output section;
A utterance extraction unit that extracts utterances to be analyzed by excluding user utterances included in the deflection group information from utterances of a plurality of users;
An analysis object utterance extraction device comprising: The deflection group information output unit is configured to increase a user group having a width evaluated as being narrower in a topic width of each user group than a user group evaluated as being shifted from the center of the topic of the entire user. Weight and include in deflection group information,
The analysis object speech extraction device according to claim 2. The topic center evaluation unit evaluates a deviation between the topic center of each user group and the topic center of the entire user based on the similarity between the centroid of the comment text of all users and the centroid of the comment text of each user group. To
The analysis object utterance extraction apparatus of any one of Claim 2 or 3. The analysis target message extraction device according to claim 4, wherein the similarity is a cosine distance. The topic width evaluation unit evaluates the topic width of each user group based on the average value of the similarity on the semantic vector space of the feature words of each user group.
The analysis object speech extraction device according to any one of claims 1 to 5. The analysis target speech extraction device according to claim 6, wherein the similarity in the semantic vector space is a cosine distance. A user comment extraction step in which the analysis target speech extraction device divides the comment text posted by the user into groups for each user,
The analysis object utterance extraction device generates, for each user, user feature information that represents user features using a group of comment sentences, and a user feature extraction step;
A user classification step in which the analysis target message extraction device groups users based on user feature information of each user;
The analysis target utterance extraction device evaluates the topic width of each user group, a topic width evaluation step,
A deflection group information output step in which the analysis target message extraction device outputs deflection group information that is information of a user group having a width that is evaluated as being narrow in the topic width of each user group;
The analysis object utterance extraction device extracts the analysis object utterance by excluding the user utterances included in the deflection group information from a plurality of user utterances; and
Analysis target speech extraction method including A user comment extraction step in which the analysis target speech extraction device divides the comment text posted by the user into groups for each user,
The analysis object utterance extraction device generates, for each user, user feature information that represents user features using a group of comment sentences, and a user feature extraction step;
A user classification step in which the analysis target message extraction device groups users based on user feature information of each user;
The analysis target utterance extraction device evaluates the topic width of each user group, a topic width evaluation step,
The analysis target utterance extraction device evaluates the deviation between the topic center of each user group and the topic center of the entire user,
The analysis target speech extraction device is based on a user group having a width that is evaluated as being narrow in the topic width of each user group and a user group in which the topic center is evaluated to be deviated from the topic center of the entire user. Deflection group information output step for outputting deflection group information
The analysis object utterance extraction device extracts the analysis object utterance by excluding the user utterances included in the deflection group information from a plurality of user utterances; and
Analysis target speech extraction method including On the computer,
User comment extraction step for grouping comment text posted by the user for each user,
For each user, a user feature extraction step for generating user feature information representing a user feature using a group of comment sentences;
A user classification step of grouping users based on the user feature information of each user;
A topic width evaluation step for evaluating the topic width of each user group;
A deflection group information output step for outputting deflection group information, which is information of a user group having a width that is evaluated to be narrow among the topic widths of each user group;
A utterance extraction step of extracting analysis target utterances by excluding user utterances included in the deflection group information from a plurality of user utterances;
A computer program for running. On the computer,
User comment extraction step for grouping comment text posted by the user for each user,
For each user, a user feature extraction step for generating user feature information representing a user feature using a group of comment sentences;
A user classification step of grouping users based on the user feature information of each user;
A topic width evaluation step for evaluating the topic width of each user group;
A topic-centric evaluation step for evaluating the difference between the topic center of each user group and the topic center of the entire user;
Deflection that outputs deflection group information based on a user group evaluated as being narrower than the topic center of the entire user and a user group evaluated that the topic center is deviated from the center of the topic of the entire user. A group information output step;
A utterance extraction step of extracting analysis target utterances by excluding user utterances included in the deflection group information from a plurality of user utterances;
A computer program for running.

Images