JP6895167B2 - Utility value estimator and program - Google Patents

Description

The present invention relates to a utility value estimated instrumentation置及beauty program, spoken text (e.g., post text) about the utility value estimated instrumentation置及beauty program for estimating the utility value for the user for.

As a prior art, in Non-Patent Document 1, utility information is extracted from submitted opinions by the following procedure.

First, a noun and a noun phrase (a phrase in which nouns are continuous) are regarded as issues, and the issues are extracted using MeCab, a morphological analysis engine. Next, in order to cluster the extracted issues, let word2vec learn the word vector from the Japanese corpus of Wikipedia (R), obtain the word vector for the issue, perform clustering using k-means, and set the issues. Is extracted. Calculate the importance to the issue set for each user. The method of extracting the importance to the issue set creates a co-occurrence network of the issues and calculates the importance of each issue by the PageRank algorithm. After that, the total importance of the issues belonging to the issue set is calculated, and the calculated utility value is normalized to 0.0 to 1.0.

Motoki Sato, Akihisa Sengoku, Yuma Imi, Shun Shiramatsu, Keihide Fujita, Takayuki Ito. “Calculation of Pareto Front by Text Opinion Analysis in Online Discussions”. Proceedings of the Japanese Society for Artificial Intelligence National Convention vol.29, pp.1-4 , 2015.

However, the technique described in Non-Patent Document 1 has the following problems with respect to the extraction of the issue set and the estimation of the utility value of the user.

First, when the issue set is automatically extracted, the issue is expressed by a word vector and k-means clustering is used. Therefore, the obtained issue set is a set of issues having similar vectors. Therefore, even words that refer to the same topic are likely to be classified into different issues.

Second, the user's emotions are not taken into consideration when estimating the user's utility value from the posted opinion. Therefore, an opinion having a content that is negative to the user may have a high utility value.

The present invention has been made to solve the above problems, and an object of the present invention is to provide a utility value estimation device and a program capable of accurately estimating a utility value for a user.

Another object of the present invention is to provide an issue set importance calculation device and a program capable of calculating the importance of an issue set for accurately estimating a utility value for a user.

In order to achieve the above object, the utility value estimation device of the present invention is a utility value estimation device that estimates the utility value indicating the degree of preference for the target user with respect to the utterance text to be evaluated, and is the utterance text to be evaluated. From the above, whether the issue extraction unit that extracts a nomenclature satisfying a predetermined condition as an issue or the issue extracted by the issue extraction unit belongs to the issue set classified according to the latent topic model is described above. A utility value estimation unit that estimates the utility value of the target user for the spoken text to be evaluated based on the importance of the issue set to the target user, which is determined for each classified issue set. Is configured to include.

The issue set importance calculation device of the present invention includes an issue extraction unit that extracts a plurality of nomenclature satisfying predetermined conditions from a plurality of utterance texts by a plurality of users as issues, and a plurality of issues extracted by the issue extraction unit. Based on the issues of, the issue set classification unit that classifies into a plurality of issue sets according to the latent topic model, and for each of the users, each of the issues extracted from the utterance text by the user is the classified issue. For each of the classified issue sets based on which of the sets it belongs to, an importance calculation unit for calculating the importance of the issue set to the user is included.

Further, the program of the present invention is a program for making a computer function as each part constituting the above-mentioned utility value estimation device or issue set importance calculation device.

As described above, according to the utility value estimation device and the program of the present invention, which of the issue sets classified according to the latent topic model belongs to the issues extracted from the utterance text to be evaluated, as described above. By estimating the utility value of the target user for the utterance text to be evaluated based on the importance of the issue set to the target user, which is determined for each classified issue set, the user can use it. The utility value can be estimated accurately.

Further, according to the issue set importance calculation device and the program of the present invention, it is classified into a plurality of issue sets according to the latent topic model based on a plurality of issues extracted from a plurality of utterance texts by a plurality of users. For each of the users, for each of the classified issue sets, for each of the classified issue sets, for the user, based on which of the classified issue sets each of the issues extracted from the utterance text by the user belongs. By calculating the importance of the issue set, it is possible to calculate the importance of the issue set for accurately estimating the utility value for the user.

It is a block diagram which shows the functional structure in the utility value estimation apparatus which concerns on 1st Embodiment. It is a figure which shows the example of the co-occurrence network. It is a figure for demonstrating the method of calculating the importance of each issue. It is a flowchart which shows the issue set importance calculation processing routine in the utility value estimation apparatus which concerns on 1st Embodiment. It is a flowchart which shows the utility value estimation processing routine in the utility value estimation apparatus which concerns on 1st Embodiment. It is a block diagram which shows the functional structure in the utility value estimation apparatus which concerns on 2nd Embodiment. It is a figure which shows an example of a dependency relation and a tree structure. It is a figure for demonstrating the method of the polarity determination when a child clause is a polar word. It is a flowchart which shows the issue set importance calculation processing routine in the utility value estimation apparatus which concerns on 2nd Embodiment. It is a block diagram which shows the functional structure in the utility value estimation apparatus which concerns on 3rd Embodiment. It is a figure for demonstrating the method of extracting the Pareto optimal utterance text. It is a figure which shows the example of visualization of the agreement proposal candidate. It is a flowchart which shows the agreement draft visualization processing routine in the utility value estimation apparatus which concerns on 3rd Embodiment. It is a block diagram which shows the functional structure in the utility value estimation apparatus which concerns on 4th Embodiment. It is a flowchart which shows the utility value estimation processing routine in the utility value estimation apparatus which concerns on 4th Embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
<Structure of utility value estimation device according to the first embodiment>
Next, the configuration of the utility value estimation device according to the first embodiment will be described. As shown in FIG. 1, the utility value estimation device 100 according to the present embodiment includes a CPU, a RAM, a program for executing an issue set importance calculation processing routine and a utility value estimation processing routine described later, and various data. It can be configured with a stored ROM and a computer including. The utility value estimation device 100 is functionally configured to include an input unit 10, a calculation unit 20, and an output unit 50, as shown in FIG.

The input unit 10 receives a plurality of utterance texts (for example, posted texts) expressing opinions about each of the plurality of users, and stores them in the opinion database 22 described later.

Further, the input unit 10 receives an utterance text representing an opinion for which the utility value is to be calculated.

The calculation unit 20 includes an opinion database 22, an issue extraction unit 24, an issue set classification unit 26, a topic model storage unit 28, a classification result storage unit 30, an importance calculation unit 32, and a utility value parameter storage unit 34. , The issue point extraction unit 36, the issue point classification unit 38, and the utility value calculation unit 40.

The opinion database 22 stores a plurality of utterance texts expressing opinions about each of the plurality of users.

The issue point extraction unit 24 extracts a plurality of nouns satisfying predetermined conditions as issues from a plurality of utterance texts stored in the opinion database 22 by a plurality of users.

Specifically, the utterance text is extracted by morphological analysis as a noun that satisfies a predetermined condition. Morphological analysis is one of the techniques of natural language processing in which text is separated for each word (morpheme) and part of speech is specified. For example, it may be realized by using MeCab (Non-Patent Document 2). In the present embodiment, nouns satisfying the following conditions (1) to (4) are extracted as issues.

(1) Numerals are not extracted.
(2) Nouns with only hiragana are not extracted.
(3) Prefixes are treated as nouns.
(4) Nouns with only one letter are not extracted.

The above process is performed for each user and summarized as an issue for each user.

[Non-Patent Document 2] Taku Kudo, Kaoru Yamamoto, Yuji Matsumoto. “Japanese Morphological Analysis Using Conditional Random Fields.” Research Report of the Information Processing Society Natural Language Processing (NL), 2004.

The issue set classification unit 26 sets a plurality of issue sets according to the latent topic model based on the plurality of issues extracted by the issue extraction unit 36 and the topic model stored in the topic model storage unit 28. It is generated as a classification of issues and stored in the classification result storage unit 30.

Specifically, a plurality of extracted issues are classified into each topic by a topic model, and a set of issues consisting of the issues classified into the topic is extracted for each topic. The topic model is a model that statistically estimates the topic of a word in a document from a plurality of documents. Words in a document do not appear independently, but have potential topics, and are proposed under the assumption that the same topic is likely to appear in the same document. As a topic model method, for example, LDA (Latent Dirichlet Allocation) (Non-Patent Document 3) may be used.

[Non-Patent Document 3] David M. Blei, Andrew Y. Ng, Michael I. Jordan. “Latent Dirichlet Allocation”. Journal of Machine Learning Research 3 (2003) 993-1022.

For each user, the importance calculation unit 32 uses the user for each classified issue set based on which of the classified issue sets each of the issues extracted from the utterance text by the user belongs to. The importance of the set of issues for the user is calculated and stored in the utility value parameter storage unit 34.

Specifically, the importance of the user to each set of issues is calculated by the following method.

(A) As shown in Fig. 2, a co-occurrence network of issues is created. A co-occurrence network is an undirected graph showing whether two issues appear at the same time in the same utterance text. When two nodes are connected, it indicates that the issue words of that node appear in the same sentence.

(B) Find the importance of each node in the co-occurrence network. The importance of each node (issue word) is determined by, for example, a PageRank algorithm (Non-Patent Document 4) (see FIG. 3).

[Non-Patent Document 4] Lawrence Page, Sergey Brin, Rajeev Motwani, Terry Winograd. “The PageRank Citation Ranking: Bringing Order to the Web”. Stanford Digital Libraries Working Paper, 1998.

(C) Find the average value of the importance of the issues included in the issue set. _{For example, using the importance V i, n} for each node n of the user i calculated by the PageRank algorithm, the _{importance I i, N} for the issue set N of the user i is obtained by the following equation.

The issue point extraction unit 36 extracts each of the nouns satisfying the predetermined conditions as the issue points from the utterance text expressing the opinion for which the utility value is to be calculated, as in the issue point extraction unit 24.

The issue classification unit 38 determines which of the issue sets stored in the classification result storage unit 30 belongs to each of the issues extracted by the issue extraction unit 36, and sets each of the extracted issues into an issue set. Classify into one of.

The utility value calculation unit 40 determines which of the issue sets the issues extracted by the issue extraction unit 36 belongs to, and the issues for the target user determined for each issue set stored in the utility value parameter storage unit 34. Estimate the utility value of the target user for the utterance text to be evaluated based on the importance of the set.

_{Specifically, the utility values U i and o} of the target user i for the utterance text o to be evaluated are calculated. _{Using the importance I i,} N for the issue set N of user i, the utility values U _{i, o} are calculated by the following equation.

However, issues represents a set of issues extracted from the utterance text to be evaluated, and f (n) represents a set of issues to which the issue n belongs.

Then, the estimated utility value is normalized between 0.0 and 1.0 using the following formula.

However, U _o is the utility value to be normalized, and U _min and U _max are the minimum and maximum values of the predetermined utility value. Here, it is assumed that the more important the opinion refers to the set of issues, the higher the utility value (degree of satisfaction) from the user's point of view.

The finally calculated value is used as the utility value estimated by the user i for the opinion o, and is output by the output unit 50.

<Operation of utility value estimation device according to the first embodiment>
Next, the operation of the utility value estimation device 100 according to the first embodiment will be described. First, when a plurality of utterance texts (for example, posted texts) expressing opinions about each of the plurality of users are received, the utility value estimation device 100 stores them in the opinion database 22. Then, the utility value estimation device 100 executes the issue set importance calculation processing routine shown in FIG.

First, in step S100, the issue point extraction unit 24 extracts a plurality of nouns satisfying predetermined conditions as issues from a plurality of utterance texts stored in the opinion database 22 by a plurality of users.

Next, in step S102, the issue set classification unit 26 sets a plurality of issue sets based on the plurality of issues extracted in step S100 and the topic model stored in the topic model storage unit 28. It is generated as a classification and stored in the classification result storage unit 30.

Next, in step S104, the importance calculation unit 32 classifies each of the users based on which of the classified issue sets each of the issues extracted from the utterance text by the user belongs. For each issue set, the importance of the issue set for the user is calculated, stored in the utility value parameter storage unit 34, and the issue set importance calculation processing routine is terminated.

Further, when the utterance text expressing the opinion to be calculated of the utility value is received, the utility value estimation device 100 executes the utility value estimation processing routine shown in FIG.

First, in step S110, the issue point extraction unit 36 extracts each of the nouns satisfying the predetermined conditions as the issue points from the utterance text representing the opinion for which the utility value is to be calculated.

Next, in step S112, the issue classification unit 38 determines which of the issue sets stored in the classification result storage unit 30 each of the issues extracted in step S110 belongs to, and the extracted issues Classify each of the above into one of the set of issues.

In step S114, the utility value calculation unit 40 determines which of the issue sets the issues extracted in step S110 belongs to, and the target user determined for each issue set stored in the utility value parameter storage unit 34. Based on the importance of the set of issues for the user, the utility value of the target user for the utterance text to be calculated is estimated, output by the output unit 50, and the utility value estimation processing routine is terminated.

As described above, according to the utility value estimation device according to the first embodiment, which of the issue sets classified according to the latent topic model belongs to the issues extracted from the utterance text to be evaluated. The utility value for the user is accurate by estimating the utility value of the target user for the utterance text to be evaluated based on the importance of the issue set for the target user, which is determined for each classified issue set. It can be estimated well.

In addition, based on a plurality of issues extracted from a plurality of utterance texts by a plurality of users, a set of a plurality of issues according to a latent topic model is generated as a classification of the issues, and for each user, from the utterance text by the user. The utility value for the user is calculated by calculating the importance of the issue set for the user for each classified issue set based on which of the classified issue sets each of the extracted issues belongs to. It is possible to calculate the importance of a set of issues for accurate estimation.

In addition, in social media such as discussion boards that express opinions on the Web, it is possible to automatically estimate the utility value that quantifies the importance and preference of each user with respect to the posted opinion.

In addition, the point that words that point to the same topic, which is a problem of the conventional technology, is classified into different problem groups, is extracted from the opinions of all users and converted into a problem set (set of related nouns) by topic estimation. It was solved by automatic classification.

[Second Embodiment]
Next, the utility value estimation device according to the second embodiment will be described. The same processing as in the first embodiment is designated by the same reference numerals and the description thereof will be omitted.

The second embodiment is different from the first embodiment in that the utility value is estimated in consideration of the polarity of the utterance text having the issue set.

<Structure of utility value estimation device according to the second embodiment>
Next, the configuration of the utility value estimation device according to the second embodiment will be described. As shown in FIG. 6, the utility value estimation device 200 according to the second embodiment is functionally configured to include an input unit 10, a calculation unit 220, and an output unit 50.

The calculation unit 220 includes an opinion database 22, an issue extraction unit 24, an issue set classification unit 26, a topic model storage unit 28, a classification result storage unit 30, an importance calculation unit 32, and a polarity determination unit 232. It is composed of a utility value parameter storage unit 234, an issue extraction unit 36, an issue classification unit 38, and a utility value calculation unit 240.

The polarity determination unit 232 sets the issue set for the user for each of the classified issue sets based on the dependency analysis result of the utterance text by the user and the word to which the polarity is given in advance for each user. The polarity of the spoken text to be held is determined and stored in the utility value parameter storage unit 234.

Specifically, when the polarity of the user's issue set is positive, the user is positive for the issue set, and when it is negative, it is negative.

To determine the polarity with respect to the issue set, first determine the polarity with respect to the issue by dependency analysis. Dependency analysis is a technique for analyzing the modified / modified relationship between two clauses in a text. For example, CaboCha (Non-Patent Document 5) may be used as the dependency analysis engine.

[Non-Patent Document 5] Taku Kudo, Yuji Matsumoto. “Japanese Dependency Analysis by Support Vector Machine”. Technical Report of the Society of Electronics, Information and Communication Engineers, 2000.

Then, if there is a polar word in the word related to the noun, it is assumed that the polarity is the polarity for the noun, and the procedure for finding the polarity for the issue is shown below.

(A) Create a tree structure by intercepting and analyzing sentences.
(B) Move the tree to each clause by breadth-first search.
(C) If the clause contains an issue, find the polarity of the issue.
(D) If the clause contains a negative element, the polarity is reversed. If double negation is detected, the polarity is determined by template matching.

The case of finding the polarity for the noun "cat" included in the example sentence "My cat is sometimes unfriendly but basically cute" shown in FIG. 7 will be described. First, a tree structure is created by dependency analysis of this sentence. Next, move from the parent phrase "kawaii" to each phrase by breadth-first search, and check whether the phrase contains an issue (noun). Since the phrase "cat is" contains the noun "cat", it goes back to the parent phrase of this phrase. We found a phrase "cute" that includes the polar word "cute" whose polarity is predetermined, and let the polarity of this polar word "cute" be the polarity with respect to "cat". A Japanese appreciation evaluation expression dictionary is used for polarity determination (Non-Patent Document 6).

[Non-Patent Document 6] Daiki Sano. "Japanese Appraisal Evaluation Expression Dictionary-Attitude Evaluation-JAppraisal Dictionary ver1.0" Published by Language Resources Association, 2011.

Since the polarity of "cute" is positive because it is set to 1 in the case of positive and -1 in the case of negative, the polarity values p _{i and n} of the issue "cat" of user i are set to +1. Also, when the child clause of the noun is a polar word such as "cute cat", it is also searched (see Fig. 8).

_{Then, the polarities P i and N} with respect to the issue set N of the user i are obtained by the following equation.

However, the polarity values of the issues n included in the issue set N are p _{i and n} . _{According to the above equation, when the sum of the polar values p i and n} of the issues n included in the issue set N _{is 0 or more, the polarities P i and N} with respect to the issue set N of the user i are set to +1 and when they are negative, they are -1. To do.

If the clause containing the issue n contains a negative element, the polarity value is inverted. Here, the expression expressing negation in a sentence is called a negation element. As a method for determining a negative expression, in the embodiment of the present invention, the following three types shown in (1) to (3) are defined as negative elements, registered in a list in advance, and matched with each clause. Judge with.

(1) Negative words are used as negative elements. For example, the auxiliary verbs "not" and "zu", the suffix "not", the prefix "non", "not", "nothing", "not", "anti", and "different" are negative words.

(2) Non-existent content words are used as negative elements. For example, the adjective "not" and the noun "none" are non-existent content words.

(3) A compound word representing negation is used as a negation element. For example, "not", "not", "cannot be", etc. are compound words that express negation.

In addition, regarding double negation, create a template as shown in the table below, and do not invert or invert the polarity value according to the basic meaning corresponding to the case where it is matched with them.

The utility value calculation unit 240 determines which of the issue sets the issues extracted by the issue extraction unit 36 belongs to, and the issues for the target user determined for each issue set stored in the utility value parameter storage unit 234. Based on the importance of the set and the polarity of the utterance text having the issue set for the target user, the utility value of the target user for the utterance text to be evaluated is estimated.

_{Specifically, the utility values U i and o} of the target user i for the utterance text o to be evaluated are calculated. Using the importance I _{i, f (n)} for the issue set f (n) of the user i and the polarities P _{i, f (n)} for the issue set f (n) of the user i, the utility values U _{i, o} are , Calculated by the following formula.

However, issues represents a set of issues extracted from the utterance text to be evaluated, and f (n) represents a set of issues to which the issue n belongs.

Further, in the utterance text o to be evaluated, when the clause including the issue n contains a negative element, the polarities P _{i and f (n)} are inverted and the above equation is calculated. In addition, in the utterance text o to be evaluated, when the clause containing the issue n includes a double negation, the polarities P _{i and f (n)} are reversed according to the basic meaning corresponding to the double negation. Alternatively, the above equation is calculated without inversion.

Then, the estimated utility value is normalized between 0.0 and 1.0 by the following formula.

Here, it is assumed that the opinion that refers to a set of issues that are more important and affirmative has a higher utility value (satisfaction) from the user's point of view.

The finally calculated value is used as the utility value estimated by the user i for the opinion o, and is output by the output unit 50.

<Operation of utility value estimation device according to the second embodiment>
Next, the operation of the utility value estimation device 200 according to the second embodiment will be described. The same processing as in the first embodiment is designated by the same reference numerals and detailed description thereof will be omitted.

The issue set importance calculation processing routine in the second embodiment will be described with reference to FIG.

First, in step S100, the issue point extraction unit 24 extracts a plurality of nouns satisfying predetermined conditions as issues from a plurality of utterance texts stored in the opinion database 22 by a plurality of users.

Next, in step S102, the issue set classification unit 26 responds to the latent topic model based on the plurality of issues extracted in step S100 and the topic model stored in the topic model storage unit 28. A plurality of issue points are generated as a classification of issues and stored in the classification result storage unit 30.

In step S200, the polarity determination unit 232 transfers the polarity determination unit 232 to the user for each set of issues classified based on the result of the dependency analysis of the utterance text by the user and the word to which the polarity is given in advance. The polarity of the utterance text having the issue set of is determined and stored in the utility value parameter storage unit 234.

Next, in step S104, the importance calculation unit 32 classifies each of the users based on which of the classified issue sets each of the issues extracted from the utterance text by the user belongs. For each issue set, the importance of the issue set for the user is calculated, stored in the utility value parameter storage unit 234, and the issue set importance calculation processing routine is terminated.

Further, when the utterance text expressing the opinion to be calculated of the utility value is received, the utility value estimation device 200 executes the same processing routine as the utility value estimation processing routine shown in FIG.

At this time, in step S114, the utility value calculation unit 240 is determined for each issue set stored in the utility value parameter storage unit 234 as to which of the issue sets the issues extracted in step S110 belongs to. , The utility value of the target user for the utterance text to be calculated is estimated based on the importance of the issue set for the target user and the polarity of the utterance text having the issue set for the target user.

As described above, according to the utility value estimation device according to the second embodiment, which of the issue sets classified according to the latent topic model belongs to the issues extracted from the utterance text to be evaluated. To estimate the utility value of the target user for the utterance text to be evaluated based on the importance of the issue set to the target user and the polarity of the utterance text having the issue set determined for each classified issue set. Therefore, the utility value for the user can be estimated with high accuracy.

In addition, based on a plurality of issues extracted from a plurality of utterance texts by a plurality of users, the issues are classified into a plurality of issue sets according to a latent topic model, and for each user, the issues extracted from the utterance texts by the users are classified. Based on which of the classified issue sets each belongs to, the result of the dependency analysis of the utterance text by the user, and the word to which the polarity is given in advance, each of the classified issue sets is the issue set for the user. By determining the polarity of the utterance text having the above, it is possible to determine the polarity of the utterance text having a set of issues for accurately estimating the utility value for the user.

In addition, a polar word with a predetermined polarity that indicates whether each word is positive or negative is used to indicate that an opinion having a content that is negative to the user, which is a problem of the prior art, has a high utility value. It was solved by determining the polarity of the utterance text having a set of issues and introducing the polarity to the user's posted opinion when estimating the utility value.

[Third Embodiment]
Next, the utility value estimation device according to the third embodiment will be described. The same processes as those in the first embodiment and the second embodiment are designated by the same reference numerals, and the description thereof will be omitted.

The third embodiment is different from the second embodiment in that the consensus proposal candidates in the plurality of users are extracted from the plurality of utterance texts expressing the opinions of the plurality of users and visualized.

<Structure of utility value estimation device according to the third embodiment>
Next, the configuration of the utility value estimation device according to the third embodiment will be described. As shown in FIG. 10, the utility value estimation device 300 according to the third embodiment functionally includes an input unit 10, a calculation unit 320, and an output unit 50.

The calculation unit 320 includes an opinion database 22, an issue extraction unit 24, an issue set classification unit 26, a topic model storage unit 28, a classification result storage unit 30, an importance calculation unit 32, and a polarity determination unit 232. It includes a utility value parameter storage unit 234, an issue extraction unit 36, an issue classification unit 38, an utility value calculation unit 240, an agreement proposal extraction unit 342, and a visualization unit 344.

The input unit 10 receives a plurality of utterance texts (for example, posted texts) expressing opinions about each of the plurality of users, and stores them in the opinion database 22 described later.

Further, the input unit 10 receives a plurality of utterance texts expressing opinions about each of the plurality of users, which are the targets for calculating the utility value.

The issue point extraction unit 36 extracts each of the nouns satisfying a predetermined condition as an issue point from each of the plurality of utterance texts expressing the opinion for which the utility value is to be calculated.

In the issue classification unit 38, for each of the plurality of utterance texts expressing the opinions for which the utility value is calculated, each of the issues extracted by the issue extraction unit 36 is stored in the classification result storage unit 30. It is determined which one it belongs to, and each of the extracted issues is classified into one of the issue sets.

The utility value calculation unit 240 determines, for each user, which of the issue sets the issues extracted by the issue extraction unit 36 belongs to each of the plurality of utterance texts expressing the opinions for which the utility value is calculated, and the utility value. The utterance of the calculation target is based on the importance of the issue set for the user and the polarity of the utterance text having the issue set for the user, which are stored in the parameter storage unit 234 and are determined for each issue set. Estimate the utility value of the user for the text.

The consensus draft extraction unit 342 has a plurality of utterances expressing opinions for which the utility value is calculated based on the estimated utility value for each of the plurality of users and each of the plurality of utterance texts for which the utility value is calculated. From the text, utterance text candidates as an agreement between multiple users are extracted.

Specifically, a main user is selected from a plurality of users, a pair of two people is created, and a consensus proposal candidate between the pair of users is extracted for each pair of users. In negotiation issues, it is said that Pareto optimality is the state in which the utility value of one negotiator cannot be increased unless the utility value of the other negotiator is reduced. Become a candidate.

Therefore, as an utterance text (a candidate for an agreement) expressing an opinion that can be agreed upon by a pair of users, a Pareto optimal utterance text obtained from the utility value obtained for the pair of users is obtained (see FIG. 11).

The visualization unit 344 generates a graph structure connecting the utterance text extracted as the draft agreement and each of the two or more users who agree with the opinion expressed by the utterance text, and presents it by the output unit 50.

Specifically, the utterance text that expresses the Pareto optimal opinion extracted for the pair of users is visualized as a consensus candidate for discussion among a plurality of users (see FIG. 12).

First, the candidate agreements for each pair are displayed on the same screen. Then, for each of the draft agreement candidates, a graph structure is generated and presented in which the candidate agreement proposal and each of the two or more users who agree with the candidate agreement proposal are connected by an edge. At this time, the user is set as an external node, the utterance text expressing an opinion is set as an internal node, and the graph is drawn with whether or not each user agrees with the consensus proposal candidate as an edge.

In addition, the utterance text that expresses the opinion that many pairs of users have selected as consensus candidates is displayed in large size. In addition, the utterance text expressing the opinion should be written only in the first few letters, and the full text and the user name (poster name) should be displayed by clicking.

In addition, as many users as possible may present a consensus proposal that is agreed upon. In this case, for example, the top 10 people with the most posts are selected as the main users, and a pair of two people is created. That is, ₁₀ C ₂ = 45 pairs are created. Then, the utterance text expressing the Pareto optimal opinion of each pair is sought as a candidate for the agreement. Then, for each consensus candidate, the number of agreed pairs is normalized so as to have a value of 0 to 1, and many pairs are consensus candidates for which the normalized number of pairs is 0.1 or more. Ask for it as an utterance text that expresses the opinion selected in, and draw a graph as a candidate for a consensus proposal for the entire discussion.

<Operation of utility value estimation device according to the third embodiment>
Next, the operation of the utility value estimation device 300 according to the third embodiment will be described. The same reference numerals are given to the same processes as those of the first embodiment and the second embodiment, and detailed description thereof will be omitted.

First, the utility value estimation device 300 executes a processing routine similar to the issue set importance calculation processing routine shown in FIG.

Further, when a plurality of utterance texts expressing opinions for which the utility value is to be calculated are received, the utility value estimation device 300 processes each of the plurality of utterance texts in the same manner as the utility value estimation processing routine shown in FIG. Repeat the routine.

At this time, in step S114, the utility value calculation unit 240 determines which of the issues in the issue set the issues extracted in step S110 belongs to for each of the plurality of users, and the issues stored in the utility value parameter storage unit 234. The utility value of the user for the utterance text for which the utility value is to be calculated, based on the importance of the issue set for the user and the polarity of the utterance text having the issue set for the user, which is determined for each set. To estimate.

Then, the utility value estimation device 300 executes the agreement draft visualization processing routine shown in FIG.

First, in step S300, the Pareto-optimal utterance text is extracted as a consensus proposal candidate based on the utility value obtained for each of the utterance texts for the pair of users.

In step S302, it is determined whether or not the processing of step S300 has been completed for all the pairs of users, and if there is a pair for which the processing of step S300 has not been completed, the process returns to step S300. Extract potential agreements for the pair. On the other hand, when the process of step S300 is completed for all the pairs of users, the process proceeds to step S304.

In step S304, based on the agreement proposal candidates extracted for all pairs of users, the utterance text extracted as the agreement proposal candidate and each of the two or more users who agree with the opinion expressed by the utterance text are separated from each other. The graph structure connected by the edges is generated and presented by the output unit 50, and the agreement draft visualization processing routine is terminated.

As described above, according to the utility value estimation device according to the third embodiment, the Pareto optimal utterance text is agreed for each pair of users based on the utility value obtained for each of the utterance texts. By extracting as candidates and visualizing the consensus proposal candidates, it is possible to lead to an appropriate consensus proposal by a plurality of users.

[Fourth Embodiment]
Next, the utility value estimation device according to the fourth embodiment will be described. The same processes as those in the first embodiment and the second embodiment are designated by the same reference numerals, and the description thereof will be omitted.

The fourth embodiment is different from the second embodiment in that the utility value for a plurality of virtual users is estimated and the comprehensive evaluation value is calculated.

<Structure of utility value estimation device according to the fourth embodiment>
Next, the configuration of the utility value estimation device according to the fourth embodiment will be described. As shown in FIG. 14, the utility value estimation device 400 according to the fourth embodiment is functionally configured to include an input unit 10, a calculation unit 420, and an output unit 50.

The calculation unit 420 includes an opinion database 22, an issue extraction unit 24, an issue set classification unit 26, a topic model storage unit 28, a classification result storage unit 30, an importance calculation unit 32, and a polarity determination unit 232. The utility value parameter storage unit 234, the issue extraction unit 36, the issue classification unit 38, the utility value calculation unit 240, and the comprehensive evaluation unit 442 are included.

The input unit 10 receives a plurality of utterance texts (for example, posted texts) expressing opinions about each of the plurality of users, and stores them in the opinion database 22 described later.

Further, the input unit 10 receives an utterance text representing an opinion for which the utility value is to be calculated.

The utility value parameter storage unit 234 stores the importance to the user and the polarity of the utterance text having the issue set for the user for each classified issue set for each user. By using each of the users here as a virtual user, the utterance text representing the opinion for which the utility value is to be calculated is evaluated thereafter.

The issue point extraction unit 36 extracts each of the nouns satisfying a predetermined condition as an issue point from the utterance text expressing the opinion for which the utility value is to be calculated.

The issue classification unit 38 indicates which of the issue sets stored in the classification result storage unit 30 belongs to each of the issues extracted by the issue extraction unit 36 with respect to the utterance text expressing the opinion for which the utility value is calculated. And classify each of the extracted issues into one of the issue sets.

The utility value calculation unit 240 determines which of the issue sets the issues extracted by the issue extraction unit 36 belongs to for each virtual user in the utterance text expressing the opinion for which the utility value is calculated, and the utility value parameter storage unit. The utterance text to be calculated based on the importance of the issue set for the virtual user and the polarity of the utterance text having the issue set for the virtual user, which is stored in 234 and determined for each issue set. Estimate the utility value of the virtual user for.

The comprehensive evaluation unit 442 calculates the comprehensive evaluation value for the utterance text based on the utility value estimated for each virtual user for the utterance text to be calculated. For example, the total value or the average value of the utility values estimated for each virtual user is calculated as a comprehensive evaluation value and output by the output unit 50.

<Operation of utility value estimation device according to the fourth embodiment>
Next, the operation of the utility value estimation device 400 according to the fourth embodiment will be described. The same reference numerals are given to the same processes as those of the first embodiment and the second embodiment, and detailed description thereof will be omitted.

First, the utility value estimation device 400 executes a processing routine similar to the issue set importance calculation processing routine shown in FIG.

Further, when the utterance text expressing the opinion to be calculated of the utility value is received, the utility value estimation device 400 executes the utility value estimation processing routine shown in FIG.

First, in step S110, the issue point extraction unit 36 extracts each of the nouns satisfying the predetermined conditions as the issue points from the utterance text representing the opinion for which the utility value is to be calculated.

Next, in step S112, the issue classification unit 38 determines which of the issue sets stored in the classification result storage unit 30 each of the issues extracted in step S110 belongs to, and the extracted issues Classify each of the above into one of the set of issues.

In step S400, the utility value calculation unit 240 determines for each of the plurality of virtual users which of the issue sets the issues extracted in step S110 belongs to, and for each issue set stored in the utility value parameter storage unit 234. The utility of the virtual user with respect to the spoken text for which the utility value is to be calculated, based on the importance of the issue set for the virtual user and the polarity of the spoken text having the issue set for the virtual user. Estimate the value.

Then, in step S402, the comprehensive evaluation unit 442 calculates the comprehensive evaluation value for the utterance text based on the utility value estimated for each virtual user for the utterance text to be calculated, outputs it by the output unit 50, and outputs the utility value. End the estimation processing routine.

As described above, according to the utility value estimation device according to the fourth embodiment, it is possible to automatically evaluate the posted opinion or the like from the utility values estimated for a plurality of virtual users. It is also possible to automatically extract important opinions.

The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

For example, in the above embodiment, the case where the calculation of the utility value is realized by the utility value estimation device together with the calculation of the importance of the issue set and the determination of the polarity has been described as an example, but the present invention is not limited to this. Instead, the calculation of the importance of the issue set, the determination of the polarity, and the calculation of the utility value may be realized by separate devices such as the issue set importance calculation device and the utility value estimation device. In this case, the issue set importance calculation device includes the opinion database 22, the issue extraction unit 24, the issue set classification unit 26, the topic model storage unit 28, the classification result storage unit 30, and the importance calculation unit 32. And the utility value parameter storage unit 34 or 234. Further, the utility value estimation device may include a utility value parameter storage unit 34 or 234, an issue extraction unit 36, an issue classification unit 38, and a utility value calculation unit 40 or 240.

Further, the processing by the issue set classification unit 26 may be performed by an external device, and a plurality of issue sets according to the latent topic model may be input from the external device to the utility value estimation device.

Further, in the third embodiment and the fourth embodiment, the utility value may be estimated without considering the polarity as in the first embodiment.

Further, in the specification of the present application, the program has been described as an embodiment in which the program is pre-installed, but the program can be stored in a computer-readable recording medium and provided, or provided via a network. It is also possible to do.

10 Input unit 20, 220, 320, 420 Calculation unit 22 Opinion database 24, 36 Issue extraction unit 26 Issue set classification unit 28 Topic model storage unit 30 Classification result storage unit 32 Importance calculation unit 34, 234 Utility value parameter storage unit 38 Issue classification unit 40, 240 Utility value calculation unit 50 Output unit 100, 200, 300, 400 Utility value estimation device 232 Polarity determination unit 342 Agreement draft extraction unit 344 Visualization unit 442 Comprehensive evaluation unit

Claims (4)

It is a utility value estimation device that estimates the utility value indicating the degree of preference for the target user for the utterance text to be evaluated.
An issue extraction unit that extracts nouns as issues from the utterance text to be evaluated,
Which of the issue sets classified according to the latent topic model belongs to the issues extracted by the issue extraction unit, and the importance of the issue set for the target user, which is determined for each classified issue set. A utility value estimation unit that estimates the utility value of the target user for the utterance text to be evaluated based on the degree and
Visualization part and
Including
The target user is a plurality of target users,
The utterance text to be evaluated is a plurality of utterance texts to be evaluated.
The utility value estimation unit is defined for each of the plurality of target users and each of the utterance texts of the plurality of evaluation targets for each of the classified issue sets and for each of the target users. Using the importance of the issue set, the utility value of the target user for the utterance text to be evaluated is estimated.
The visualization unit selects utterance text candidates agreed between two or more target users based on the utility values estimated for each of the plurality of target users and each of the plurality of evaluation target utterance texts. A utility value estimation device that generates a graph structure that connects the extracted utterance text candidates and each of two or more target users who agree with the utterance text candidates and presents them by the output unit. It is a utility value estimation device that estimates the utility value indicating the degree of preference for the target user for the utterance text to be evaluated.
An issue extraction unit that extracts nouns as issues from the utterance text to be evaluated,
Which of the issue sets classified according to the latent topic model belongs to the issues extracted by the issue extraction unit, and the importance of the issue set for the target user, which is determined for each classified issue set. A utility value estimation unit that estimates the utility value of the target user for the utterance text to be evaluated based on the degree and
Comprehensive evaluation department and
Including
The target user is a plurality of virtual users.
The utility value estimation unit uses the importance of the issue set for the virtual user, which is determined for each of the classified issue sets and for each virtual user, for each of the plurality of virtual users. Estimate the utility value of the virtual user for the utterance text to be evaluated,
The comprehensive evaluation unit is a utility value estimation device that calculates a comprehensive evaluation value for the utterance text to be evaluated based on the utility value estimated for each of the plurality of virtual users. The utility value estimation unit determines which of the issue sets classified according to the latent topic model belongs to the issues extracted by the issue extraction unit, and determines for each of the classified issue sets, the target user. Claim 1 that estimates the utility value of the target user for the utterance text to be evaluated based on the importance of the issue set to the user and the polarity of the utterance text having the issue set to the target user. Or the utility value estimation device according to 2. A program for causing a computer to function as each part constituting the utility value estimation device according to any one of claims 1 to 3.

Images