JP2017134787A - Device, program, and method for analyzing topic evaluation in multiple areas - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device with which it is possible to more precisely grasp feeling about a topic, and analyze the evaluation state of the topic in mutually different areas in such a way that the cause of this feeling is easily understandable.SOLUTION: The present evaluation analysis device comprises: keyword group generation means for extracting, for each area, a representative keyword from a document retrieved using a query that includes a topic keyword, and collecting representative keywords together that are in different languages and will be translated and generating a keyword group; feeling information determination means for determining, for each area, feeling information related to a submitted document that is in a language used in the area and was retrieved using a query that includes a topic keyword related to the topic, the determination being made using a database of vocabularies related to feelings; and score addition means for adding a score to the generated keyword group on the basis of one or both of the number of occurrences of the representative keywords and the feeling information related to the submitted document.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a technique for acquiring or generating target information by analyzing a document distributed on the Internet.

  Conventionally, a great number of users have expressed their opinions and sentiments about various topics (social media) through social media such as Twitter (registered trademark). For example, critique and commentary on a topical product or application, state an opinion about a topical person, company or region, and present a discussion of a topical event or news case. It is often done. In addition, there are often comments and feelings about topics related to daily life such as food, housing, health and crime.

  Information related to such impressions and feelings is an important judgment material for companies, governments, etc. to evaluate the current situation and determine future business / government policies. Therefore, a technique called “opinion mining” that automatically acquires and analyzes such impressions and feelings is being actively developed.

  For example, Non-Patent Document 1 discloses a number of sentiment analysis techniques that have been developed over the past 10 years. It includes emotion analysis for Twitter (registered trademark) messages as disclosed in Non-Patent Document 2.

  Research and development is also underway on techniques for searching the latest hot topics in opinion mining. For example, Non-Patent Document 3 discloses a technique for identifying a popular topic by recognizing a sudden increase in the frequency of occurrence of a keyword that is usually low in frequency.

  Further, Non-Patent Document 4 discloses a technique of feature-based sentiment analysis that performs emotion analysis using features of an object. In this method, for example, features such as “display size” and “battery life” are identified and used for the target device product “iPhone (registered trademark)”.

  In addition, although there are very few research examples, so-called cross-language sentiment analysis is also performed. For example, Non-Patent Document 5 discloses a technique that combines cross-language emotion analysis and feature emotion analysis, and uses a feature in one language and a feature similar in meaning to another language. It is disclosed. Furthermore, Non-Patent Document 6 that published the research by the inventors of the present application discloses a technique for performing characteristic emotion analysis in different languages and comparing the analysis results between languages.

  The application “talkwalker” of Non-Patent Document 7 analyzes tweets of positive and negative emotions and visualizes the percentage of tweets expressing such emotions on a world map. Furthermore, the application “Trendsmap” of Non-Patent Document 8 displays keywords such as frequent words, hash tags, and user names on each region of the world map.

B. Liu, L. Zhang, “A Survey of Opinion Mining and Sentiment Analysis”, Mining Text Data, 2012, 415-463. SM Mohammad, S. Kiritchenko, X. Zhu, "NRC-Canada: Building the State-of-the-Art in Sentiment Analysis of Tweets", Proceedings of International Workshop on Semantic Evaluation Exercises (SemEval), 2013, 321-327 page M. Mathioudakis, N. Koudas, “TwitterMonitor: Trend Detection over the Twitter Stream”, Proceedings of the ACM SIGMOD International Conference on Management of Data, 2010, 1155-1158 N. Naveed, T. Gottron, S. Staab, "Feature Sentiment Diversification of User Generated Reviews: The FREuD Approach", Proceedings of International AAAI Conference on Weblogs and Social Media (ICWSM), 2013 H. Guo, H. Zhu, Z. Guo, X. Zhang, Z. Su, "OpinionIt: A Text Mining System for Cross-Lingual Opinion Analysis", Proceedings of the ACM international conference on Information and knowledge management (CIKM), 2010, 1199-1208 M. Erdmann, K. Ikeda, H. Ishizaki, G. Hattori, Y. Takishima, "Feature Based Sentiment Analysis of Tweets in Multiple Languages", Proceedings of Web Information Systems Engineering (WISE), 2014, 109-124. Talkwalker, "social data intelligence platform", [online], [December 4, 2015 search], Internet <http://www.talkwalker.com> "trendsmap", [online], [December 4, 2015 search], Internet <http://trendsmap.com>

  However, conventional techniques such as the emotion analysis method described above have the following problems.

  First of all, there is a problem that emotions expressed throughout one document on social media do not necessarily match emotions for a topic to be evaluated. However, in the conventional emotion analysis as described above, it has been rarely performed to distinguish between these two emotions and accurately grasp the emotion based on the topic.

  As one example where these emotions do not match, a posted document that says “It is very sad not to have an iPhone (registered trademark)” can be regarded as expressing negative emotions as a whole. However, it should be judged that the topic of “iPhone (registered trademark)” represents a positive feeling as to what it wants to have. That is, emotions based on topics are often expressed in an indirect manner. Furthermore, although the posted document includes keywords related to the topic, the emotion expressed there is often unrelated to the topic. For example, the sentiment that says “She turned up and threw me an iPhone (registered trademark)” can be said to be unrelated to the topic “iPhone (registered trademark)”.

  As described above, emotion analysis that distinguishes emotions of the entire document from emotions about a target topic, and extraction of keywords that are characteristic with emotion analysis results have not been performed in the past.

  Also, as a second problem, using the result of emotion analysis, for example, simply calculating the ratio of positive and negative posted documents, the cause of this emotion behind the emotions for the topic being evaluated It is still difficult to understand.

  For example, for the purpose of understanding the cause of emotions as a result of emotion analysis, representative keywords are extracted from tweets using the Twitter (registered trademark) message emotion analysis method described in Non-Patent Document 2. It is also possible. However, since a posted document as an extraction source usually has a short sentence length and there is no formal description format or configuration, it is very difficult to extract representative keywords. Under such circumstances, the extraction of representative keywords has hardly been performed except for the characteristic emotion analysis disclosed in Non-Patent Document 4.

  Furthermore, as a third problem, it is mentioned that the conventional emotion analysis is performed in one specific country or region. In contrast, for example, sentiment analysis on the topic to be evaluated is conducted in different countries and regions, and by comparing the analysis results, an appropriate global business strategy can be established across cultural barriers. It becomes possible. However, in the past, such a global comparison method has, of course, hardly been studied since emotion analysis has been performed in one country or region.

  For example, in cross language sentiment analysis as described in Non-Patent Document 5, it is still difficult to perform accurate translation and match search queries and representative keywords between languages. There has also been no visualization of sentiment analysis results that can easily compare the results of each country / region on the world map.

  Further, in the feature emotion analysis disclosed in Non-Patent Document 6, analysis is performed specifically for a product name, and for example, the analysis result is not visualized on a map. In addition, automatic topic detection and topic-based sentiment analysis have not been implemented.

  Further, the application of Non-Patent Document 7 does not automatically detect a topic, nor does it analyze emotion based on the topic. The application of Non-Patent Document 8 displays keywords such as frequent words, hash tags, and user names on the world map, but has a function of performing emotion analysis on tweets including these keywords. Absent.

  Therefore, the present invention provides an apparatus, a program, and a method capable of more accurately grasping emotions about a topic and analyzing the evaluation status of topics in different regions in a form that makes it easy to understand the cause of the emotions. For the purpose.

According to the present invention, an evaluation analyzer for analyzing evaluations on topics in a plurality of regions,
For each relevant region, a representative document is extracted from a document in a language used in the relevant region and searched using a query including a topic keyword related to the topic, and a language used in a different region. A keyword group generation means for generating a keyword group by collecting representative keywords that are representative keywords and are translated;
For each of the regions, a posted document in a language used in the region and searched for using a query including a topic keyword related to the topic, using a database of emotional vocabulary, Emotion information determination means for determining emotion information related to the posted document;
A score is assigned to a generated keyword group based on one or both of the number of appearances of a representative keyword belonging to the keyword group and emotion information related to a posted document related to the topic to the representative keyword. And an evaluation analyzer having means.

  This evaluation analysis apparatus according to the present invention is generated as an embodiment related to the topic at a position corresponding to the area in an image displayed on a display device provided inside or outside the apparatus. Display information related to the representative keyword of the region included in the keyword group given a score satisfying a predetermined condition among the keyword groups, and information related to emotion information of the region determined to be related to the topic It is also preferable to further include display control means for causing the display control means to be operated.

Furthermore, as another embodiment, the evaluation analysis apparatus searches for the value of the position of the topic keyword in the posted document searched using the query including the topic keyword and the posted document for each region. And using a feature amount including at least one of a value related to the distance between the query used to perform the predetermined emotion-inducing word and a value related to the number of nouns appearing in the posted document, A topic-specific document determining means for determining a topic-specific posted document that is a topic-specific posted document;
The emotion information determination means uses at least information related to a word appearing in the topic-specific posted document for each topic as a feature quantity element and uses a database of emotional vocabulary It is also preferable to determine emotion information related to the topic-specific posted document.

In addition, the score giving means, as one embodiment,
(1) The difference in the degree of emotion information between representative keywords in different regions belonging to the keyword group for the keyword group to be assigned, and the degree of positive or negative in the emotion information of the topic-specific posted document related to the topic The greater the difference between, the higher the score,
(2) With respect to the keyword group to be assigned, the difference in the number of appearances between representative keywords in different regions belonging to the keyword group, and the larger the difference in the number of appearances in the topic-specific posted document related to the topic, the more Granting a high score,
(3) For a given keyword group, the higher the sum of the positive or negative degree in the emotion information of the topic-specific posted document related to the topic in all the representative keywords included in the keyword group, the higher the score. (4) For the keyword group to be assigned, the total number of appearances of the representative keywords included in the keyword group, and the more the total number of appearances appearing in the topic-specific posted document related to the topic, the more It is also preferable to employ at least one of giving a high score.

  Further, as still another embodiment, the evaluation analysis apparatus is a document in a language used in the region for each region, and has a document type that has been previously recognized as having a larger amount of information than the posted document. It is also preferable to further include topic keyword generation means for extracting the topic keyword from the document based on the appearance frequency of the word and / or the number of documents in which the word appears.

Furthermore, the keyword group generation means, as one embodiment,
For each region, collect posted documents that include both the topic keyword and the extracted representative keyword, and relate to the average distance between the selected representative keyword included in the collected posted document and a predetermined emotion-inducing word. A collected post document including a value, a value related to whether or not the representative keyword of the selection target is a noun, a value related to the distance between the representative keyword of the selection target and the topic keyword, and the representative keyword of the selection target In order to generate the keyword group using a feature quantity having at least one of the values relating to the degree of bias indicating whether the emotion information determined for is biased to one of positive and negative It is also preferable to have representative keyword determining means for selecting the representative keywords.

According to the present invention, there is also provided a program for causing a computer mounted on an apparatus for analyzing evaluations on topics in a plurality of regions to function.
For each relevant region, a representative document is extracted from a document in a language used in the relevant region and searched using a query including a topic keyword related to the topic, and a language used in a different region. A keyword group generation means for generating a keyword group by collecting representative keywords that are representative keywords and are translated;
For each of the regions, a posted document in a language used in the region and searched for using a query including a topic keyword related to the topic, using a database of emotional vocabulary, Emotion information determination means for determining emotion information related to the posted document;
A score is assigned to a generated keyword group based on one or both of the number of appearances of a representative keyword belonging to the keyword group and emotion information related to a posted document related to the topic to the representative keyword. An evaluation analysis program for causing a computer to function as means is provided.

According to the present invention, for each region, a representative keyword is extracted from a document in a language used in the region and searched using a query including a topic keyword related to the topic. Collecting representative keywords of languages used in different regions and related to each other to generate a keyword group;
For each of the regions, a posted document in a language used in the region and searched for using a query including a topic keyword related to the topic, using a database of emotional vocabulary, Determining emotion information related to the posted document;
A step of assigning a score to the generated keyword group based on one or both of the number of appearances of the representative keyword belonging to the keyword group and emotion information related to the posted document related to the topic to the representative keyword; An evaluation analysis method is provided.

  According to the evaluation analysis apparatus, program, and method of the present invention, it is possible to more accurately grasp emotions for a topic and analyze the evaluation status for topics in different regions in a manner that makes it easy to understand the cause of the emotion. .

It is the flowchart which showed the outline | summary of one Embodiment of the evaluation analysis method implemented in the evaluation analysis apparatus by this invention. It is a functional block diagram which shows the function structure in one Embodiment of the evaluation analysis apparatus by this invention. It is a flowchart which shows one Embodiment of the topic keyword production | generation process in a topic keyword (KW) production | generation part. It is a flowchart which shows one Embodiment of the topic specific document and emotion information determination process in a topic specific document determination part and an emotion information determination part. It is a flowchart which shows one Embodiment of the representative keyword extraction and selection process in a representative KW determination part. It is a program code which shows one Embodiment of the score provision process in a score provision part. It is a schematic diagram which shows one Embodiment of the emotion world map displayed on a display by a display control part.

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

  FIG. 1 is a flowchart showing an outline of an embodiment of an evaluation analysis method implemented in an evaluation analysis apparatus according to the present invention.

  The evaluation analysis apparatus 1 according to the present embodiment shown in FIG. 1 analyzes evaluation on a certain topic (topic) in a plurality of regions, for example, a plurality of countries, and the analysis result is displayed on a world map displayed on a display, for example. It is a device that can be presented at the corresponding position. Specifically, it can be an information device such as a personal computer (PC), a notebook computer, a smartphone, or a tablet computer on which the evaluation analysis program according to the present invention is installed.

In this embodiment, the evaluation / analysis apparatus 1 is connected to various document management / distribution servers 2 installed in a plurality of areas via an access network, the Internet, or the like.
(A) Documents submitted by users on social media including Twitter (registered trademark) miniblogs and various blogs (blogs, weblogs), and (b) news articles, newspaper editorials, It is possible to acquire a document (large information document) having a larger amount of information (number of characters) or the like, such as various analysis articles, as compared with the above (a). Here, the large information amount document can be a document of a type that is recognized in advance as having a larger amount of information than a posted document such as Twitter (registered trademark).

Furthermore, the evaluation analysis apparatus 1 determines, as a specific functional configuration, a “topic” to be evaluated in each region included in the analysis target,
(A) For each region included in the analysis target, a document (a large amount of information document) in a language used in the region and searched from a document searched using a query including a topic keyword related to “topic” Extract "representative keywords"
(B) A “representative keyword” of languages used in different regions and collecting “representative keywords” in a relation to be translated to generate a keyword group,
(C) For each region included in the analysis target, it is a posted document in a language used in the region, and is related to emotions for a posted document searched using a query including a topic keyword related to “topic” By performing machine learning using a vocabulary database (emotion vocabulary DB), “emotion information” related to the posted document is determined,
(D) For the generated keyword group, the number of appearances of the “representative keyword” belonging to this keyword group and “emotion information” related to the posted document related to the topic related to the “representative keyword” belonging to this keyword group A “score” can be assigned based on one or both of the above.

As an example of such processing by (A) to (D), after “iPhone (registered trademark)” is first determined as a topic,
(A) From the representative keywords extracted for each country in Japan, the United States, Brazil, France, and China, which are the target countries for analysis,
(B) A keyword group containing representative keywords such as “price” and “display size”
(C) “Impressions / feelings” (emotion information) on Twitter (registered trademark) submission documents (in Japanese, English, Portuguese, French and Chinese) in Japan, the United States, Brazil, France and China And
(D) Based on the number of appearances of “price” and “display size” belonging to the generated keyword group and “impressions / comments” with respect to the posted document related to “iPhone (registered trademark)” as the representative keyword, As the “score” for the generated keyword group, for example, dozens of hundreds of points are given.

  As described above, in the evaluation analysis apparatus 1 of the present invention, “emotion information” is determined by performing machine learning using a sentiment vocabulary database for a posted document searched using a query including a topic keyword. ing. This solves the problem that emotions expressed throughout the submitted document do not necessarily match the emotions of the “topic” to be evaluated, and as a result, more accurately grasps the emotion to the “topic” to be evaluated. It becomes possible. For example, emotional information for a document that says “She turned up and threw me an iPhone (registered trademark)” is excluded from the analysis because it is unrelated to the topic “iPhone (registered trademark)” It can be done.

  In the evaluation analysis apparatus 1 of the present invention, “representative keywords” included in the keyword group are determined. This “representative keyword” includes words that are in different languages and are translated into each other, and serves as a guide for understanding the cause of the emotion behind the emotion for the “topic” to be evaluated. For example, if positive “emotion information” (impression / feelings) is determined for one product (name) as an evaluation target topic and “price” is determined as a representative keyword, positive for this product The low price is presumed to be the cause of this feeling. In this way, it is possible to analyze the status of topic evaluation in a form that makes it easy to understand the cause of emotions for the topic.

  Furthermore, in the evaluation analysis apparatus 1 of the present invention, “emotion information” is determined for each region included in the analysis target, and the number of occurrences of “representative keyword” and “representative keyword” are determined for the generated keyword group. "Score" is given based on one or both of "emotion information" related to the posted document related to the topic. Here, the “representative keywords” to be handled exist for each language used in different regions, and are in a relation to be translated from each other. This makes it possible to select keyword groups suitable for analysis and comparison, and to effectively understand the status of topic evaluation in different regions.

  Therefore, it is possible not only to analyze emotions in one specific region (one country) but also to perform emotion analysis across countries using different languages, for example. As a result, for example, by comparing analysis results between countries, it is possible to establish an appropriate global strategy that transcends cultural barriers in fields such as business and politics.

Incidentally, the evaluation analysis apparatus 1 is one embodiment,
(E) In a position corresponding to a target area in an image (for example, a world map) displayed on a display device provided in the apparatus 1 or outside the apparatus 1,
(E1) Information related to the “representative keyword” of this region included in the keyword group assigned with the “score” that satisfies the predetermined condition among the keyword groups generated as related to the evaluation target topic;
(E2) It is also preferable to display information related to “emotional information” in this area determined as related to the evaluation target topic.

  When such an embodiment is adopted, for example, by displaying the “representative keyword” and “emotion information” for each country on the world map at the position of the country and visualizing the emotion analysis result, the country ( It is also possible to easily compare emotion analysis results for each region.

  Note that the evaluation analysis apparatus according to the present invention may be an apparatus separate from a PC as a user interface used by a user, for example. In this case, the evaluation analysis apparatus is connected to the PC by wireless or wired communication, inputs instruction information about the analysis processing operation from the PC, and outputs display image data to the display control unit of the PC. There may be. Further, the evaluation analysis apparatus according to the present invention performs the machine learning processing part for determining the “emotion information” in the configuration (C) by using another external server, and stores only the processing result database therein. It may be a thing.

  Next, the outline of the topic evaluation analysis process in the evaluation analysis apparatus 1 according to the present embodiment will be described in steps S1 to S5 using the flowchart shown in FIG.

(S1: Topic Detection) First, a topic that is suitable (worth worth performing), that is, a so-called hot topic, that is suitable for performing global sentiment analysis and comparison of sentiment analysis results in each region (country) is detected.
(S2: Topic-specific emotion analysis) Next, instead of evaluating the emotion of the entire posted document, the emotion of the posted document for the search query (topic) is analyzed.
(S3: Extraction of representative keywords) Next, for the purpose of making the user understand the cause of emotional bias as an analysis result, representative keywords are extracted using both a posted document positive for emotion and a negative posted document.

(S4: Comparison of emotional analysis results between regions) (including different language use countries) Compare emotional analysis results between multiple countries (regions) and emphasize differences in results judged to be the most interesting To summarize.
(S5: Generation and display of emotional information world map) Understand the causes of emotions as well as the “emotional information” such as the degree of emotional bias, including the comparison results, for the analyzed countries. "Representative keywords" are also displayed on the world map.

  Thus, according to the present evaluation analysis apparatus 1, topic-specific emotion analysis in a plurality of countries (regions) can be performed. In addition, after taking up a topic of value as an analysis target, it is possible to analyze the emotion of the topic just taken, not the emotion of the entire posted document. In addition to “emotional information” such as the degree of emotional bias, “representative keywords” that can explain this “emotional information” are extracted and displayed, and the differences in interesting results between countries are highlighted and displayed. It can also be done.

  Hereinafter, details of each of the above-described processing steps S1 to S5 will be described in detail, while clearly showing the functional components of the evaluation analyzer 1 in charge of the processing steps.

  FIG. 2 is a functional block diagram showing a functional configuration in an embodiment of the evaluation analyzer according to the present invention.

  According to FIG. 2, the evaluation analysis apparatus 1 includes a communication interface unit 101, a document database (DB) 102, an emotion vocabulary DB 103, a score assignment keyword (KW) group recording unit 104, a display 105, and an operation unit 106. And processor memory. Here, the processor memory realizes the evaluation analysis function by executing a program that causes the computer that is the main function unit of the evaluation analysis apparatus 1 to function.

  Further, the processor memory includes a topic KW generation unit 111, a topic-specific document determination unit 112, an emotion information determination unit 113, a KW group generation unit 114, a score assignment unit 116, and input / output control as function components. Part 117. In addition, according to FIG. 2, the flow of the process which connected each function structure part with the arrow is understood also as one Embodiment of the evaluation analysis method by this invention.

  The communication interface unit 101 receives, from the document management / distribution server 2, via a communication network, a posted document by a user on social media including Twitter (registered trademark), a large information document such as a news article, and the like. In addition, information related to emotion information world map including emotion information and representative keywords as an analysis result may be input from an input / output control unit 117 described later, and transmitted to an external device. The document DB 102 is a database in which acquired posted documents, large information content documents, and the like are stored in a form that can be used for various document searches.

[Topic Keyword Generation Processing: S1]
The topic KW generation unit 111 is a document of a language used in each country (region) included in the analysis target and is previously recognized as having a larger amount of information than the posted document. Topic keywords are extracted from the document (large information document) based on the appearance frequency of words and / or the number of documents in which words appear.

  FIG. 3 is a flowchart showing an embodiment of topic keyword generation processing in the topic KW generation unit 111.

(S11) For example, 100 keywords are extracted from a news article (large information document) in descending order of feature using an algorithm such as tf-idf (term frequency-inverse document frequency).
Here, for example, a topic (topic) taken up in an article in a newspaper is not necessarily the latest hot topic on social media. Then, next, step S12 is performed.

(S12) The extracted keywords are filtered to exclude keywords that are not characteristic (small weight) in a plurality of posted documents.
Here, as the filtering process, it is also preferable to calculate tf-idf scores in a plurality of posted documents for each keyword, and to exclude keywords having a score less than a predetermined threshold. A keyword selected by such filtering processing is determined as a topic keyword, and a topic related to the determined topic keyword becomes a topic to be evaluated. Thereafter, a topic-specific document that is a document specific to the topic is determined using the topic keyword.

(S13) Using a query including the selected topic keyword, a posted document including the topic keyword is searched.
Hereinafter, a topic-specific document is determined from the posted documents related to the topic thus obtained, and emotion analysis is performed on the determined topic-specific document.

[Topic-specific emotion analysis processing: S2]
Returning to FIG. 2, the topic-specific document determining unit 112 is searched for a country (region) included in the analysis target using a query that is a posted document in a language used in the region and includes a topic keyword. About posted documents
(A) a value related to the position of the topic keyword in the posted document;
(B) a value related to the distance between the query used to search the posted document and a predetermined emotion-inducing word;
(C) Machine learning is performed using a feature quantity having at least one of the values related to the number of nouns appearing in the posted document, and a topic-specific posted document that is a topic-specific posted document is determined. To do. In the topic specific document determination unit 112 shown in FIG. 2, the feature quantity elements (a), (b), and (c) are respectively calculated by the KW position analysis unit 112a, the query distance analysis unit 112b, and the part of speech analysis unit 112c. The

  For each country (region) included in the analysis target, the emotion information determination unit 113 uses at least information related to a word appearing in the topic-specific posted document as a feature element for the topic-specific posted document determined later. By performing machine learning using the emotion vocabulary DB 103, emotion information related to the topic-specific posted document is determined.

  The emotion vocabulary DB 103 is a database of vocabulary related to emotion. Vocabulary related to emotion can be a lexicon (lexicon, lexicon) in which emotion types are organized and stored. In particular, it is also preferable that the polarity of the emotion, which is a biased state (position between both emotions) of positive and negative emotions, is a polarity lexicon associated with the vocabulary.

  As a modification mode, the machine learning processing part in the topic specific document determination unit 112 and / or the emotion information determination unit 113 is performed in a processing server or the like installed outside the device 1, and a database of this processing result is The apparatus 1 can be accommodated and used.

  FIG. 4 is a flowchart showing an embodiment of topic-specific document and emotion information determination processing in the topic-specific document determination unit 112 and the emotion information determination unit 113.

(S21) First, the position of the topic keyword is analyzed and examined in the post document to be analyzed that is searched using the query including the topic keyword. For example, the entire posted document may be a number line from 0 to 1, and the value of the topic keyword appearing on the number line may be examined. Alternatively, the number of the topic keyword corresponding to the number of words from the document start position can be set as the position.
Accordingly, for example, if the topic keyword is close to the start position of the posted document, it is easy to determine that the keyword of the posted document is highly likely to match the topic keyword.

(S22) Next, all emotion-inducing words in the post document to be analyzed are specified using the emotion vocabulary DB 103, and the query (topic keyword) used for the search and the specified query closest to the query The distance from the emotion-inducing word, for example, the number of words or the number of characters is calculated.
Thus, for example, if the calculated distance is a small value less than a predetermined value, it is easy to determine that the emotion related to the posted document is likely to be related to the topic keyword. Here, the emotion-inducing word is a word meaning the expression of a certain sentiment (sentiment), and is a word stored in the sentiment vocabulary DB 103 in association with the sentiment (its polarity).

(S23, S24) Next, using an application such as a POS (part of speech) tagger, the query (topic keyword), emotion-inducing word and all other words of speech in the post document to be analyzed are identified, Count the number of nouns in the submitted document.
Thereby, the possibility that the topic (topic) of the posted document is different from the topic keyword can be measured by the number of nouns in the posted document.

  Specifically, for example, if many nouns appear in the posted document, it is easy to determine that the topic (topic) of the posted document is likely to be different from the topic keyword. For example, a posted document “I love iPhone (registered trademark)!” Contains only one noun, but the emotion represented by this posted document can be said to be an emotion for the iPhone (registered trademark). On the other hand, the posted document “It is also good to order coffee for this store's iPhone (registered trademark) app!” Contains four nouns, giving feelings about the iPhone (registered trademark). It is judged that it is not expressed.

  Here, the information relating to the position of the topic keyword, the distance from the emotion-inducing word, and the number of nouns acquired in steps S21 to S24 described above is stored in the feature amount memory of the topic-specific document determination unit 112 in each step. And a feature vector for the post document to be analyzed is generated using these pieces of information as components.

(S25, S26) Using the feature vector generated for the posted document to be analyzed, learning was performed by a machine learning algorithm such as SVM (Support Vector Machine) or Naive Bayes, and this posted document expressed an emotion related to the topic. A classification process for determining whether the document is a topic-specific document is performed. If it is determined that the document is not a topic-specific document, the posted document is not an emotion analysis target, and the analysis process ends.

(S27) On the other hand, when it is determined that the posted document is a topic-specific document, emotion analysis is performed on the posted document. This emotion analysis can be performed using a known standard emotion analysis algorithm based on machine learning such as SVM or Naive Bayes.
Here, as the emotion analysis result, for example, an emotion ratio (rentiment ratio) r that is a degree of emotional bias between the negative and positive poles may be output, or coordinate values in various emotional spaces may be output. Good. The emotion ratio r may take three values of r = −1 (negative), 0 (neutral), and 1 (positive), may take a negative or positive value, or a value in a range of −1 ≦ r ≦ 1. You may set to take.

  By selecting the topic-specific documents whose sentiments are related to the topic (keywords) by the processing described above, and performing emotion analysis only on the topic-specific documents suitable for the selected emotion analysis target It becomes possible to acquire appropriate emotion information that hit the target. For example, a posted document such as “She turned up and threw me an iPhone (registered trademark)” did not express an emotion related to the iPhone (registered trademark). Can be excluded from the target. In addition, since emotions are comprehensively determined using machine learning, for example, it is possible to handle indirect expressions of emotions such as “I am sorry that there is no iPhone (registered trademark)”.

[Representative keyword extraction and keyword group generation processing: S3]
Returning to FIG. 2, the KW group generation unit 114 includes a representative KW determination unit 114 a and a grouping unit 114 f. Among these, the representative KW determination unit 114a performs a search for each country (region) included in the analysis target using a query including a topic keyword related to the topic to be evaluated, which is a language document used in the region. The representative keyword is extracted from the document.

The representative KW determination unit 114a further
(A) Collect post documents containing both topic keywords and extracted representative keywords for each country (region) included in the analysis target,
(B) For each collected posted document, a feature vector having elements calculated by the emotion word distance analyzing unit 114b, the part of speech analyzing unit 114c, the topic KW distance analyzing unit 114d, and the KW appearance frequency analyzing unit 114e is generated. And
(C) It is also preferable to perform machine learning using the generated feature vector to select a representative keyword for generating a keyword group.

Here, in the above (b),
(B1) The emotion word distance analysis unit 114b calculates a value related to the average distance between the representative keyword to be selected included in the collected posted document and the predetermined emotion induction word,
(B2) The part-of-speech analysis unit 114c calculates a value related to whether the representative keyword to be selected is a noun,
(B3) The topic KW distance analysis unit 114d calculates a value related to the distance between the representative keyword to be selected and the topic keyword,
(B4) The KW appearance frequency analysis unit 114e calculates a value related to the degree of bias indicating whether the emotion information determined for the collected posted document including the representative keyword to be selected is biased to one of positive and negative. calculate.

  The representative KW determination unit 114a can perform machine learning using a feature amount including at least one, preferably all, of the calculated values (b1) to (b4). Incidentally, the machine learning process for extracting the representative keyword can be performed by applying a known standard algorithm such as SVM or Naive Bayes, and the process result is whether or not the representative keyword to be selected is selected. Such information may be output.

  Further, as a change mode, the machine learning processing part in the representative KW determination unit 114a is performed in a processing server or the like installed outside the apparatus 1, and the apparatus 1 accommodates and uses the database of the processing results. It is also possible.

  Similarly, in FIG. 2, the grouping unit 114 f is a representative keyword of a language used in a different country (region) among the representative keywords selected in each country (region) included in the evaluation target, and is translated. A keyword group is generated by collecting representative keywords that are related to each other. At this time, it is also preferable that a plurality of keyword groups are generated, each including one or more representative keyword sets having a relationship to be translated.

  Here, for the translation of the representative keyword, a known machine translation application or a link for translation between languages can be used. It is also preferable that representative keywords (in different languages) that are translated with each other are determined using a preset thesaurus. As such a thesaurus, it is also preferable to prepare in advance various types of dictionaries such as all target languages and normal dictionaries, slang dictionaries, new word dictionaries between the languages. In this embodiment, it is very important to generate an appropriate keyword group using a rich thesaurus in order to effectively compare analysis results between countries (regions).

  FIG. 5 is a flowchart showing an embodiment of representative keyword extraction / selection processing in the representative KW determination unit 114a.

(S31) First, for each topic keyword, representative keyword candidates are extracted from a document (large information amount document) existing on the Internet, and a representative keyword candidate list is generated. This process is specifically
(A) Search and collect web pages using a search query that combines a topic keyword and an appropriate keyword (for example, “news”) using a known web page search engine,
(B) Extract the body text of the collected web pages, and from the extracted text, tf-idf, df-idf, LDA (Latent Dirichlet Allocation), LSA (Latent Semantic Association), AIC (Akaike Information Criterion), etc. This can be implemented by extracting representative keyword candidates using a known algorithm.

(S32) Then, for each of the extracted representative keyword candidates, all the posted documents including both the keyword and the topic keyword are collected.
Here, it is also preferable to exclude a posted document corresponding to noise that does not include any emotion-inducing word and is not suitable for emotion analysis. Furthermore, it is also preferable to exclude posted documents such as advertisements that are said to be spam. In this case, for example, a URL (Uniform Resource Locator) of web content other than an image is included, or “win”, “deal”, “review”, “$”, and “$” All posted documents containing suspicious words / characters such as “offer” may be excluded.

  Next, in order to exclude inappropriate representative keyword candidates using machine learning, the following steps S33 to S36 are performed.

(S33) Using the emotion vocabulary DB 103, which is a polarity lexicon, the emotion-inducing word in the posted document is identified, and the average distance between the representative keyword candidate to be selected and the emotion-inducing word is calculated. .
This makes it easier to determine that the more representative keyword candidates are accompanied by emotion-inducing words, the more they are selected as representative keywords.

(S34) Using an application such as a POS tagger, it is determined whether the representative keyword candidate in the posted document is a noun, and the determination result is output.
As a result, if the representative keyword candidate is a noun, it becomes easier to be selected as a representative keyword than when it is a verb or an adjective.

(S35) The semantic distance between the topic keyword and the representative keyword candidate is calculated using a technique such as PMI (pointwise mutual information) or cosine similarity of co-occurrence vectors.
Thus, if the semantic distance is large, it is easy to determine that the representative keyword candidate is likely to be only a general word. For example, when the topic keyword is “Prius (registered trademark)”, “engine” having a small semantic distance is selected as the representative keyword among the representative keyword candidates “engine” and “people”. It becomes easy to be judged.

(S36) The co-occurrence frequency of representative keyword candidates in each of the posted document of positive emotion, the posted document of negative emotion, and the posted document of neutral emotion is determined.
As a result, it is easy to determine that a representative keyword candidate that appears frequently in one of the positive and negative posted documents but that appears in both has a low possibility of being selected as a representative keyword. For example, if the topic keyword is “iPhone (registered trademark)” and the representative keyword candidate “app” appears frequently in both positive and negative posted documents, “app” Is determined to be unsuitable for understanding the polarities of emotions, and there is a high possibility that it will not be selected as a representative keyword.

  Here, the information acquired in steps S33 to S36 described above is stored in the feature amount memory of the representative KW determination unit 114a in each step, and the representative keyword candidates to be selected using these pieces of information as components. Feature quantity vectors are generated.

(S37) Using the feature vector generated for the representative keyword candidate to be selected, learning is performed by a machine learning algorithm such as SVM or Naive Bayes, and it is determined whether or not the representative keyword candidate is selected as the representative keyword. Perform classification process. Here, only representative keyword candidates determined to be selected as representative keywords are determined as representative keywords, a list of determined representative keywords is generated, and the main extraction / selection process ends.

  The representative keyword determined by the above processing is of course related to the topic (topic keyword) that is deemed appropriate for the evaluation target, but it is characterized by feelings about this topic (impression / feelings). It is likely that the regional differences in emotions for the topic are conspicuous.

[Emotion analysis result comparison process: S4]

  Returning to FIG. 2, the score assigning unit 116 inputs the keyword group generated from the KW group generating unit 114 and the emotion information determined from the emotion information determining unit 113. A score is assigned based on one or preferably both of the number of appearances of the representative keyword belonging to the keyword group and the emotion information related to the posted document related to the topic to the representative keyword.

  Here, the score given to the keyword group is a ranking score at the time of presenting information, and information (representative keyword etc.) related to the keyword group whose score is greater than or equal to a predetermined value is presented as an evaluation analysis result. is there.

  Next, the score assigning unit 116 records information in which the score assigned for each keyword group is associated with the score giving KW group recording unit 104. Hereinafter, a specific embodiment of the score assignment will be described in detail with reference to FIG.

  FIG. 6 is a program code showing an embodiment of a score assignment process in the score assignment unit 116.

  The score assigning unit 116 adopts at least one of the four score assigning policies (a) to (d) described below as a specific score assigning process, preferably all the policies, and calculates the score. Can be granted.

(A) First, for the keyword group to be assigned, the total number of appearances of representative keywords (for all languages) included in the keyword group, and the total number of appearances appearing in the topic-specific post document related to the topic The higher the number, the higher the score. This policy is shown in FIG.
score = 0
score = score + absolute no. of occurrences score
According to this, for example, as the total number of appearances increases, it becomes more appropriate to present information on the target keyword group.

(B) Also, with respect to the keyword group to be given, the degree of positive or negative in the emotion information of the topic-specific posted document related to the topic in all of the representative keywords (for all languages) included in this keyword group ( The higher the sum of the emotion ratios, the higher the score. This policy is shown in FIG.
score = score + overall ratio of sentiment score
According to this, for example, a keyword group having a large sum of emotion ratios should be presented with more information than a keyword group that has a balanced emotion ratio.

(C) Furthermore, for the keyword group to be given, the difference in the number of appearances between different representative keywords in the country (region) belonging to this keyword group, and the difference in the number of appearances in the topic-specific posted document related to the topic The higher the value, the higher the score. This policy is shown in FIG.
score = score + difference in no. of occurrences score
According to this, for example, a representative keyword that has a high number of appearances for one language and shows a low number of appearances for one or more other languages may be presented. It will be more appropriate.

(D) Further, for a given keyword group, a topic-specific posted document related to a topic, which is a difference in the degree of emotion information (emotion ratio) between different representative keywords in countries (regions) belonging to the keyword group The higher the positive or negative degree (emotion ratio) in the emotion information, the higher the score. This policy is shown in FIG.
score = score + difference in ratio of sentiment score
According to this, for example, a representative keyword whose emotion ratio for one language is significantly different from the emotion ratio for one or more other languages is an emotion for which language. Compared to representative keywords that have the same ratio, it is more appropriate to be presented.

The four score assignment policies (a) to (d) described above are executed, and a value “score” obtained by adding the scores calculated in each of them is assigned to the keyword group to be given. Thereafter, for example, representative keywords included in the keyword group having the largest combined score (ranking score) are presented as evaluation analysis results. This is shown in FIG.
rank according to score
output top ranked keyword groups
In practice, however, it is a process in the input / output control unit 117 (FIG. 2) described below.

[Emotion information world map generation / display processing: S5]
Returning to FIG. 2, the display control unit 117 a of the input / output control unit 117 serves as an evaluation target in the world map image displayed on the display device (display 105 in FIG. 2) provided inside or outside the device 1. In the position corresponding to the country (region) included,
(A) Information related to the representative keyword of the country (region) included in the keyword group that is given a score satisfying a predetermined condition among the keyword groups generated as related to the topic to be evaluated (for example, representative keyword Itself) and
(B) Information related to emotion information of the country (region) determined to be related to the topic to be evaluated (for example, emotion ratio) is displayed. Here, the “score satisfying the predetermined condition” in (b) can be, for example, the score having the largest value among the scores assigned to the plurality of keyword groups. Moreover, it is good also as a score more than a predetermined threshold value. In this case, representative keywords and emotion ratios related to a plurality of keyword groups may be displayed.

  The input / output control unit 117 further generates other screen image signals to be displayed on the display 105 and outputs them to the display 105. In addition, an operation signal from the user is input from the input unit 101 and a command corresponding to the operation signal is output to each unit in the apparatus 1. For example, in response to a transmission instruction input by the user, the representative keyword and emotion ratio information associated with the position on the world map image may be transmitted to an external information processing apparatus via the communication interface 101.

  Here, the operation unit 106 is a device that accepts an input operation by a user or an interface for the device. For example, the operation unit 106 is a keyboard, a push button, a pointing device such as a mouse or a touch pad, a touch panel, or an interface for them. be able to. It is also possible to adopt a touch panel display as an integrated unit of the operation unit 106 and the display 105.

  FIG. 7 is a schematic diagram showing an embodiment of an emotion world map displayed on the display 105 by the display control unit 117a.

  According to FIG. 7, an emotion world map when “iPhone (registered trademark)” is adopted as a topic (keyword) is displayed on the display 105. Specifically, representative keyword groups (keyword groups) determined in each of the six regions (continents) in Asia, Oceania, Europe, Africa, North America, and South America, as well as these An emotion ratio (Sentiment ratio) is displayed as emotion information for the representative keyword group.

  As representative keywords, ten keywords such as “action” and “access” in the keyword group having the highest score are employed. These representative keywords are displayed in each region (continent) translated into a language representative of the region. The emotion ratio is displayed as either positive (feeling good) or negative (antisense (bad feeling)). In addition to the number of retrieved posted documents, the ratio of positive and negative in these posted documents is also displayed as a pie chart. Each representative keyword (related document) is also classified as positive or negative.

  As a result, it is possible to more accurately grasp the feelings about the topic (impression / sentiment) as the topic evaluation status in different regions. In addition, since it is possible to know emotions (such as positive or negative) with respect to the representative keyword (post document), it is easy to understand the cause of the emotions.

  The display of the representative keyword and emotion information (emotion ratio) is not limited to the form performed for each continent (region) on the displayed world map as shown in FIG. For example, it may be performed for each country in the displayed world map by selecting a menu by the user, or may be performed for each country in one displayed continent (region). In addition, representative keywords and emotion information (emotion ratio) in one country or one region can be displayed in detail. Furthermore, as an additional function, various trend and statistical information may be displayed together and directly compared among a plurality of countries (regions).

  As described above in detail, in the present invention, emotion information is determined by performing machine learning using a emotion vocabulary database for a posted document searched using a query including a topic keyword. Thereby, it becomes possible to grasp the feeling with respect to the topic to be evaluated more accurately.

  Further, representative keywords included in the keyword group are determined. These representative keywords include those that are in different languages and translated into each other, and serve as a guide to understand the cause of the emotion behind the emotions for the topic being evaluated in each region (country). . That is, by using the representative keyword, it is possible to analyze the topic evaluation status in each region (country) in a form that makes it easy to understand the cause of emotions for the topic.

  Further, emotion information is determined for each region (country) included in the analysis target, and a score is assigned to the generated keyword group. This makes it possible to select keyword groups suitable for analysis and comparison, and to effectively understand the status of topic evaluation in different regions (countries).

  In addition, this enables not only emotion analysis in a specific region (one country) but also emotion analysis across countries using different languages, for example. In other words, opinion mining on feelings for global topics can be implemented globally. As a result, for example, by comparing analysis results between countries, it is possible to establish an appropriate global strategy that transcends cultural barriers in fields such as business and politics.

  For the various embodiments of the present invention described above, various changes, modifications, and omissions in the technical idea and scope of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Evaluation analyzer 101 Communication interface part 102 Document database (DB)
103 Emotional Vocabulary DB
104 score assignment keyword (KW) group recording unit 105 display 106 operation unit 111 topic KW generation unit 112 topic specific document determination unit 112a KW position 112b query distance analysis unit 112c part of speech analysis unit 113 emotion information determination unit 114 KW group generation unit 114a representative KW determination unit 114b Emotion word distance analysis unit 114c Part of speech analysis unit 114d Topic KW distance analysis unit 114e KW appearance frequency analysis unit 114f Grouping unit 116 Score assigning unit 117 Input / output control unit 117a Display control unit 117a

Claims (8)

An evaluation analyzer that analyzes evaluations on topics in multiple regions,
For each relevant region, a representative document is extracted from a document in a language used in the relevant region and searched using a query including a topic keyword related to the topic, and a language used in a different region. A keyword group generation means for generating a keyword group by collecting representative keywords that are representative keywords and are translated;
For each of the regions, a posted document in a language used in the region and searched for using a query including a topic keyword related to the topic, using a database of emotional vocabulary, Emotion information determination means for determining emotion information related to the posted document;
A score is assigned to a generated keyword group based on one or both of the number of appearances of a representative keyword belonging to the keyword group and emotion information related to a posted document related to the topic to the representative keyword. And an evaluation analysis apparatus.   A score that satisfies a predetermined condition among the keyword groups generated as related to the topic is given to a position corresponding to the region in an image displayed on a display device provided inside or outside the device. The display control means for displaying the information related to the representative keyword of the area included in the keyword group and the information related to the emotion information of the area determined as related to the topic. The evaluation analyzer according to 1. For each region, a value related to the position of the topic keyword in the posted document searched using a query including the topic keyword, a query used to search the posted document, and a predetermined emotion-inducing word A topic-specific posted document that is a posted document specific to the topic is determined using a feature quantity having at least one of a value related to the distance and a value related to the number of nouns appearing in the posted document. A topic-specific document determination means for
The emotion information determination means uses at least information related to a word appearing in the topic-specific posted document for each topic as a feature quantity element and uses a database of emotional vocabulary The evaluation analysis apparatus according to claim 1, wherein emotion information related to the topic-specific posted document is determined. The score giving means is
(1) The difference in the degree of emotion information between representative keywords in different regions belonging to the keyword group for the keyword group to be assigned, and the degree of positive or negative in the emotion information of the topic-specific posted document related to the topic The greater the difference between, the higher the score,
(2) With respect to the keyword group to be assigned, the difference in the number of appearances between representative keywords in different regions belonging to the keyword group, and the larger the difference in the number of appearances in the topic-specific posted document related to the topic, the more Granting a high score,
(3) For a given keyword group, the higher the sum of the positive or negative degree in the emotion information of the topic-specific posted document related to the topic in all the representative keywords included in the keyword group, the higher the score. (4) For the keyword group to be assigned, the total number of appearances of the representative keywords included in the keyword group, and the more the total number of appearances appearing in the topic-specific posted document related to the topic, the more The evaluation analyzer according to claim 3, wherein at least one of giving a high score is adopted.   A document in a language used in the region for each region, and a word appearance frequency and / or a word appearance from a document of a document type that is recognized in advance as having a larger amount of information than the posted document. 5. The evaluation analysis apparatus according to claim 1, further comprising topic keyword generation means for extracting the topic keyword based on the number. The keyword group generation means includes
For each region, collect posted documents that include both the topic keyword and the extracted representative keyword, and relate to the average distance between the selected representative keyword included in the collected posted document and a predetermined emotion-inducing word. A collected post document including a value, a value related to whether or not the representative keyword of the selection target is a noun, a value related to the distance between the representative keyword of the selection target and the topic keyword, and the representative keyword of the selection target In order to generate the keyword group using a feature quantity having at least one of the values relating to the degree of bias indicating whether the emotion information determined for is biased to one of positive and negative The representative keyword determining means for selecting the representative keyword of any one of claims 1 to 5 is provided. Evaluation and analysis device. A program for operating a computer installed in a device for analyzing evaluations on topics in a plurality of regions,
For each relevant region, a representative document is extracted from a document in a language used in the relevant region and searched using a query including a topic keyword related to the topic, and a language used in a different region. A keyword group generation means for generating a keyword group by collecting representative keywords that are representative keywords and are translated;
For each of the regions, a posted document in a language used in the region and searched for using a query including a topic keyword related to the topic, using a database of emotional vocabulary, Emotion information determination means for determining emotion information related to the posted document;
A score is assigned to a generated keyword group based on one or both of the number of appearances of a representative keyword belonging to the keyword group and emotion information related to a posted document related to the topic to the representative keyword. An evaluation analysis program characterized by causing a computer to function as a means. An evaluation analysis method for analyzing evaluations on topics in a plurality of regions,
For each relevant region, a representative document is extracted from a document in a language used in the relevant region and searched using a query including a topic keyword related to the topic, and a language used in a different region. Collecting representative keywords that are representative keywords that are translated and generating a keyword group;
For each of the regions, a posted document in a language used in the region and searched for using a query including a topic keyword related to the topic, using a database of emotional vocabulary, Determining emotion information related to the posted document;
Assigning a score to the generated keyword group based on one or both of the number of appearances of a representative keyword belonging to the keyword group and emotion information related to a posted document related to the topic to the representative keyword; An evaluation analysis method characterized by comprising:

Images