JP7186080B2 - Text information judgment device and its program - Google Patents

Description

The present invention relates to a text information determination device for determining text information with high topicality and a program therefor.

The development of social networking services (SNS) has made it possible for individuals to easily transmit information in real time. The social big data of these SNSs has become an influential information source, and is utilized to understand social trends and social problems. In recent years, a technique has been proposed that classifies a large amount of texts posted on SNS into several types of emotions and uses the classification results to predict stock price trends (Non-Patent Document 1). Also, a method for detecting a hot post on an SNS has been proposed (Non-Patent Document 2). In some broadcasting stations, posts to SNSs are always monitored by a plurality of people, and posts that are hot topics and information indicated by the posts (SNS information) may be utilized in program production.

"Stock Price Prediction Using Twitter's Multiaxial Emotional Information", Yusuke Masui, Iwao Fujino, The Association for Natural Language Processing, Proceedings of the 24th Annual Conference of the Association for Natural Language Processing, March 2018 "Detection and Analysis of Flares Using Emotion Analysis on Twitter", Naoki Takahashi, Yasuhiko Higaki, Institute of Electronics, Information and Communication Engineers, Institute of Electronics, Information and Communication Engineers Technical Research Report, March 2017

In order to find topical SNS information, the number of responses such as "retweet", "reply", and "like" may be used as a reference. However, if you check the SNS information that actually has a large number of responses, there are many posts related to advertisements such as company products and services. In the example of FIG. 9, words (keywords) frequently included in tweets (registered trademark) posted on a certain day are shown in order of the number of appearances. As shown in FIG. 9, many words such as "coupon code", "birthday", and "recruitment of participants" are included in the SNS information, and it is difficult to directly utilize such SNS information in program production. Therefore, it is required to accurately determine not only the number of responses and the number of appearances of words, but also highly topical SNS information.

SUMMARY OF THE INVENTION An object of the present invention is to provide a text information determination device and its program that can accurately determine topical text information.

In view of the above-described problems, a text information determination device according to the present invention is a text information determination device for determining highly topical text information based on changes in emotional polarity, comprising: text information input means; means, emotion polarity ratio calculation means, burst determination means, reversal determination means, and text information determination means.

According to such a text information determination device, the text information input means receives a plurality of pieces of text information including preset keywords and to which time information is added.
The emotional polarity label adding means calculates an emotional polarity score of each piece of input text information and adds an emotional polarity label corresponding to the calculated emotional polarity score to each piece of text information.
The emotional polarity ratio calculation means aggregates the number of cases of text information for each emotional polarity label in a preset aggregation period, and calculates the ratio of the total number of cases of text information to the number of cases for each emotional polarity label as the ratio of emotional polarity. do.

The burst determination means determines whether or not an emotional polarity burst has occurred based on the increase in the emotional polarity ratio.
The reversal determining means determines whether or not the reversal of the emotional polarity has occurred by comparing the ratios of the emotional polarities.
Then, the text information determination means determines text information corresponding to the emotional polarity label, keyword, and aggregation period when the burst and/or reversal occurs as highly topical text information.
In this way, the text information determination device recognizes the occurrence of a burst or a reversal of emotional polarity in text information as a change in emotional polarity, and determines highly topical text information.

The present invention can also be implemented by a program that causes hardware resources such as a CPU, a memory, and a hard disk provided in a computer to operate cooperatively as the above-described text information determination device.

According to the present invention, the occurrence of a burst or a reversal of emotional polarity in text information is treated as a change in emotional polarity, so highly topical text information can be accurately determined.

It is a block diagram which shows the structure of the SNS posted text extraction apparatus which concerns on embodiment. FIG. 10 is an explanatory diagram illustrating an example of an emotional polarity label added to an SNS posted message in the embodiment; In an embodiment, it is a table showing an example of the ratio of emotional polarities. In an embodiment, it is a graph showing an example of bursts of emotional polarity. In an embodiment, it is a graph showing an example of reversal of emotional polarity. 4 is a flow chart showing the operation of the SNS posted text extraction device according to the embodiment; 10 is a graph showing an example of changes in the ratio of emotional polarities in the example. In the example, it is a table showing an example of the rate of emotional polarity. It is a list showing words that appear frequently in SNS posts.

(embodiment)
[Overview of SNS Posted Sentence Extraction Device]
BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings as appropriate.
With reference to FIG. 1, an outline of an SNS posted text extraction device (text information determination means) 1 according to an embodiment will be described.

The SNS-posted message extraction device 1 extracts SNS-posted messages (text information) that are highly topical based on changes in emotional polarities. As shown in FIG. 1, the SNS posted text extraction device 1 includes storage means 10, keyword setting means 20, SNS posted text collection means (text information input means) 30, emotion classification means 40, and stationary value determination means. 50 , and SNS posted message extraction means (text information determination means) 60 .

This SNS posted text extraction device 1 efficiently and accurately extracts highly topical SNS posted texts by focusing on the emotional polarity indicated by the SNS posted texts. That is, the SNS-posted text extraction device 1 extracts SNS-posted texts that have triggered a significant change in the emotions of SNS users and SNS-posted texts that have triggered a change in public opinion.

First, the SNS-posted text extraction device 1 uses the emotion classification means 40 to determine the emotional polarity of the SNS-posted text posted on the SNS. Next, the SNS posted sentence extraction device 1 calculates the stationary ratio of each emotional polarity by the stationary value determination means 50, and determines whether there is a change in the emotional polarity such as an emotional polarity burst or an emotional polarity reversal. determine whether When there is a change in the emotional polarity, the SNS posted text extracting device 1 extracts the SNS posted text by considering the SNS posted text as a popular SNS posted text by the SNS posted text extracting means 60. - 特許庁

Note that the emotional polarity is the emotional expression of the SNS user for various objects such as services, products, or events, such as "○○ land was fun!!!". For example, the word “enjoyable” is an important vocabulary for determining the emotional polarity of SNS posts.
In this embodiment, three emotional polarities of positive, neutral, and negative are set in advance. Positive represents a positive emotional polarity, Neutral represents a neutral emotional polarity, and Negative represents a negative emotional polarity.
Each emotional polarity corresponds to an emotional polarity label that can uniquely identify the emotional polarity.

Note that the SNS posted text (social media information) is text information in units of posts posted on the SNS. For example, a tweet (registered trademark) can be cited as an SNS post. To this SNS posted message, as time information, a posted time indicating the time when the message was posted to the SNS is added.

In this embodiment, the emotion classification means 40 operates in two modes of operation. The first operation mode is a learning mode for learning an emotion classifier (emotion classification model), which will be described later. The second operation mode is an extraction mode in which a highly topical SNS posted message is determined and extracted using a learned emotion classifier. In this embodiment, the user of the SNS posted text extraction device 1 manually switches between the two operation modes.

[Configuration of SNS posted text extraction device]
The configuration of the SNS posted text extraction device 1 will be described in detail below.
The storage means 10 is a storage device such as a memory or a hard disk that stores emotion classifiers. This emotion classifier is written by the emotion classification means 40 (emotion polarity label addition means 41) described later in the learning mode, and is referenced by the emotion classification means 40 in the extraction mode.

The keyword setting means 20 presets keywords for collecting SNS posts. In this embodiment, the keyword setting means 20 sets one or more keywords.
For example, the keyword setting means 20 morphologically analyzes a large number of previously acquired SNS posts, counts the number of appearances of each word, and automatically sets words with a large number of appearances as keywords (see FIG. 9). At this time, the user of the SNS posted text extraction apparatus 1 can arbitrarily designate the part of speech of the keyword such as noun, verb, compound word, etc., using operating means such as a keyboard and a mouse (not shown).
Alternatively, the user of the SNS posted text extraction device 1 may manually set any keyword using the operation means.
The keyword setting means 20 outputs the set keywords to the SNS posted message collecting means 30 .

The SNS posted message collection unit 30 collects SNS posted messages including the keyword input from the keyword setting unit 20 for a predetermined collection period. In the present embodiment, the SNS posted text collection unit 30 extracts SNS posted texts containing one or more of the keywords set in the keyword setting unit 20 . Here, the user of the SNS posted text extraction device 1 manually sets an arbitrary collection period using the operation means. If the user of the SNS posted text extraction device 1 does not set the collection period, a predetermined initial value (for example, 10 hours) may be set as the collection period.
The SNS posted text collection means 30 outputs the collected SNS posted texts to the emotion classification means 40 .

The emotion classification means 40 classifies the emotion polarity of the SNS posted message input from the SNS posted message collection means 30 , and includes an emotion polarity label adding means 41 .

The emotional polarity label adding means 41 calculates the emotional polarity score of each SNS posted message, and adds an emotional polarity label corresponding to the calculated emotional polarity score to each SNS posted message. In the present embodiment, the emotion polarity label adding means 41 operates differently in the learning mode and the extraction mode, and therefore will be described in order.

In the learning mode, the emotion polarity label adding means 41 learns an emotion classifier using learning data in which emotion polarity labels are added to SNS posted messages in advance. For example, the emotion classifier consists of a neural network such as a recurrent neural network (RNN), a support vector machine (SVM), or a machine learning model such as CRF (Conditional Random Fields). . At this time, the emotion classifier may be learned in units of letters, morphemes, or words. By using this emotion classifier, it is possible to calculate, for each emotion polarity, an emotion polarity score representing the possibility of corresponding to that emotion polarity.

In the extraction mode, the emotion polarity label adding means 41 calculates the emotion polarity score of each SNS posted text using the emotion classifier, and adds the emotion polarity label with the highest emotion polarity score to each SNS posted text.

For example, the emotion polarity label adding means 41 divides each SNS posted sentence input from the SNS posted sentence collecting means 30 into character units, morpheme units, or word units in correspondence with the emotion classifier. Next, the emotional polarity label adding means 41 inputs each of the divided SNS posted sentences to the emotion classifier, and calculates an emotional polarity score for each emotional polarity. That is, the emotional polarity label adding means 41 calculates positive, neutral, and negative emotional polarity scores for each SNS posted message. Then, the emotional polarity label adding means 41 obtains the emotional polarity with the highest emotional polarity score for each SNS posted message, and adds an emotional polarity label indicating the emotional polarity to each SNS posted message.
The emotional polarity label adding means 41 outputs the SNS posted message to which the emotional polarity label and the posting time are added to the stationary value determination means 50 .

FIG. 2 shows an example of SNS posted sentences collected with the keyword “XX land”. As shown in FIG. 2, in the first line of the SNS post text, "○○ land was fun!!!" .626” was the highest score. In this case, the emotion polarity label of the SNS posted message on the first line is positive.
In addition, in the SNS post text "○○ land, open today as usual" on the second line, it is assumed that the neutral emotional polarity score "0.731" was the highest score among the emotional polarity scores on the three axes. . In this case, the emotion polarity label of the SNS posted text on the second line is neutral.
In addition, in the SNS post sentence “The meal at ○○ land was very bad” on the third line, the negative emotional polarity score “0.526” was the highest score among the three axis emotional polarity scores. do. In this case, the emotion polarity label of the SNS posted text on the third line is negative.

Returning to FIG. 1, the description of the configuration of the SNS posted text extraction device 1 is continued.
The stationary value determination means 50 treats the emotional polarity ratio as a stationary value and determines a change in the emotional polarity, and includes emotional polarity ratio calculation means 51, burst determination means 52, and reversal determination means 53. .

The emotional polarity ratio calculating means 51 counts the number of SNS posts input from the emotional polarity label adding means 41 for each emotional polarity label during a preset counting period, and calculates the total number of SNS posted messages and the emotional polarity label. The ratio of the number of cases for each is calculated as the ratio of emotional polarity. That is, the emotional polarity ratio calculation unit 51 calculates the ratio of the emotional polarity for each predetermined aggregation period (aggregation interval) based on the emotional polarity label added by the emotional polarity label addition unit 41 and the posting time of the SNS posted message. .
The emotional polarity ratio calculation means 51 outputs the calculated ratio of emotional polarity to the burst determination means 52 .

<Percentage of emotional polarity>
With reference to FIG. 3, the method of calculating the ratio of emotional polarities will be specifically described.
FIG. 3 shows an example of the emotional polarity ratio calculated by the emotional polarity ratio calculating means 51 as a table. In this example, the keyword is "XX land", the collection time is 10 hours, and the aggregation period is 1 hour.

First, since the number of SNS-posted messages differs in each aggregation period, the emotional polarity ratio calculation unit 51 calculates the total number of SNS-posted messages for each aggregation period. As described above, each SNS posted message containing the keyword "XX land" is attached with an emotional polarity label of positive, neutral, or negative. Therefore, the emotional polarity ratio calculation means 51 counts the number of SNS posts to which each emotional polarity label is added for each counting period, and obtains the number of times for each emotional polarity label. Furthermore, the emotional polarity ratio calculation means 51 calculates the ratio between the total number of SNS posted messages and the number of SNS posted messages aggregated for each emotional polarity label for each aggregation period.

In FIG. 3, when the aggregation period is the first hour, the total number of SNS posts is 1580, the number of positive SNS posts is 452, and the number of negative SNS posts is 351. , the number of neutral SNS posts is 777. In this case, the ratio of positive emotional polarity to the total number of SNS posts is 452/1580≒0.286, the ratio of negative emotional polarity is 351/1580≒0.222, and the ratio of neutral emotional polarity is is 777/1580≈0.492.
Although detailed explanation is omitted, the rate of emotional polarity is calculated in the same manner for the aggregation period after the second hour.

The burst determination means 52 determines whether or not an emotional polarity burst has occurred due to an increase in the emotional polarity ratio calculated by the emotional polarity ratio calculation means 51 . A burst is a phenomenon in which the ratio of a certain emotional polarity suddenly increases in a certain period of time.

In this embodiment, the burst determination means 52 determines whether or not a burst has occurred with an emotional polarity other than neutral, that is, with positive or negative emotional polarities. For example, the burst determination means 52 can determine the occurrence of a burst using Kleinberg's burst detection technique (Reference 1).
Reference 1: Bursty and Hierarchical Structure in Streams, Kleinberg, J., 2003.

<Burst Judgment>
The burst determination method will be specifically described with reference to FIG.
FIG. 4 illustrates a graph when a burst of emotional polarity occurs. In this example, the keyword is "XX land", the collection time is 10 hours, and the aggregation period is 1 hour. In the graph of FIG. 4, the solid line represents the percentage of positive emotional polarities, and the dashed line represents the percentage of negative emotional polarities. In FIG. 4, the vertical axis represents the ratio of emotional polarity, and the horizontal axis represents time (aggregation period).

As shown in FIG. 4, the percentage of positive emotional polarities increased sharply at 7-8 hours, suggesting that a burst of positive emotional polarities occurred. In this case, the burst determination means 52 determines that a burst has occurred at the emotional polarity label “positive”, the keyword “○○ land”, and the aggregation period “7th to 8th hours”, and the determination result is posted on the SNS. Output to extraction means 60 .

The reversal determining means 53 determines whether or not the reversal of the emotional polarity has occurred by comparing the emotional polarity ratios calculated by the emotional polarity ratio calculating means 51 . In this embodiment, the reversal determining means 53 determines whether or not the emotional polarities other than neutral, that is, the positive and negative emotional polarities are reversed.

<Determination of reversal of emotional polarity>
Three specific examples of the method for determining the reversal of emotional polarity will be described below.
In the first determination method, the ratios of positive and negative emotional polarities are compared in the order of aggregation periods, and when the ratios of both emotional polarities are equal, it is determined that the emotional polarities are reversed in that aggregation period. That is, in the first determination method, as shown in the following equation (1), during the counting period in which the difference between the positive emotional polarity ratio pos and the negative emotional polarity ratio neg becomes 0, the emotional polarity is Determine that it is reversed.
pos−neg=0 Expression (1)

In the second determination method, the ratios of positive and negative emotional polarities are compared in the order of aggregation periods, and when the ratio of both emotional polarities is switched, it is determined that the emotional polarities have been reversed in that aggregation period. At this time, the ratio of the emotional polarity with the smaller ratio is used as a reference, and the ratio of the emotional polarity with the larger ratio is compared with the ratio of the emotional polarity.

Consider a case where the percentage of negative emotional polarities is smaller than the percentage of positive emotional polarities in a certain aggregation period. In this case, in the next aggregation period, as shown in the following formula (2-1), if the negative emotional polarity ratio neg becomes greater than the positive emotional polarity ratio pos, it is determined that the emotional polarity has been reversed. do.
neg>pos...Equation (2-1)

Also, consider a case where the percentage of positive emotional polarities is smaller than the percentage of negative emotional polarities in a certain aggregation period. In this case, when the positive emotional polarity ratio pos becomes greater than the negative emotional polarity ratio neg in the next aggregation period, it is determined that the emotional polarity has been reversed, as shown in the following equation (2-2). do.
pos>neg Expression (2-2)

In the third determination method, the percentages of positive and negative emotional polarities are compared in order of aggregation period, and when the difference in the percentages of bi-emotional polarities becomes equal to or greater than a predetermined threshold after the magnitude of the percentage of bi-emotional polarities is switched. , it is determined that the emotional polarity has been reversed during the aggregation period. That is, in the third determination method, as shown in the following equation (3), during the counting period when the difference between the positive emotional polarity ratio pos and the negative emotional polarity ratio neg exceeds the threshold th, the emotional polarity is It is determined that the rotation has reversed (0<th<1).
|pos−neg|>th Expression (3)

In this third determination method, the threshold th may be manually set by the user of the SNS posted text extraction device 1 using the operating means.
Furthermore, the threshold th may be automatically set based on the ratio of the emotional polarity having the smaller ratio. For example, consider a case in which the ratio pos of positive emotional polarities is smaller than the ratio neg of negative emotional polarities in a certain aggregation period. In this case, in the next aggregation period, the value obtained by multiplying the positive emotional polarity ratio pos by a predetermined coefficient x is set as the threshold th (where 0<x<1).
The user of the SNS posted message extraction device 1 can arbitrarily set which of the first to third determination methods described above is used.

FIG. 5 shows a graph when the reversal of emotional polarity occurs. As shown in FIG. 5, before the 7th hour, the percentage of positive emotional polarities continues to be greater than the percentage of negative emotional polarities. At 7-8 hours, the proportion of negative emotional polarities was greater than the proportion of positive emotional polarities, suggesting that the negative emotional polarities reversed the positive emotional polarities. In this case, the reversal determination means 53 determines that the reversal of the emotional polarity has occurred at the emotional polarity label "negative", the keyword "○○ land", and the counting period "7th to 8th hour", and the determination result is Output to the SNS posted text extraction means 60 .

Returning to FIG. 1, the description of the configuration of the SNS posted text extraction device 1 is continued.
The SNS posted sentence extraction means 60 judges and extracts SNS posted sentences corresponding to the emotional polarity label, keyword, and aggregation period when burst and/or reversal occurs, as highly topical SNS posted sentences. .

In this embodiment, the SNS posted message extracting means 60 extracts SNS posted messages corresponding to the determination result of the burst determination means 52 or the reverse determination means 53 as highly topical SNS posted messages. Specifically, the SNS posted message extracting means 60 determines that the posted time of the SNS posted message among the SNS posted messages output by the emotion polarity label adding means 41 is included in the aggregation period of the determination result, and that the SNS posted message is an SNS posted message. An SNS posted text that matches the emotional polarity label with the result and includes the keyword of the determination result is extracted. In other words, the highly topical SNS posted text is the SNS posted text in which at least one of the emotional polarity burst or reversal occurs. Of course, highly topical SNS-posted texts include SNS-posted texts in which both a burst and a reversal of emotional polarity occur.

Here, the SNS posted message extracting means 60 can output the extracted SNS posted message by any method. For example, the SNS posted message extracting means 60 can transmit the extracted SNS posted message by data file or e-mail. In addition, the SNS posted message extraction means 60 may display the extracted SNS posted message on a display device (not shown).

Further, the SNS posted message extracting means 60 may output the extracted SNS posted message itself or a summary of the extracted SNS posted message.
When there are a large number of extracted SNS posted texts, it is difficult for the user of the SNS posted text extraction device 1 to check all the extracted SNS posted texts. In this case, the SNS posted message extracting means 60 may output the determination result (emotional polarity label, keyword, and aggregation period) when the SNS posted message with high topicality is extracted. Further, the SNS posted message extracting means 60 may output the extracted SNS posted message with a large number of responses. Furthermore, the SNS posted message extraction means 60 may display each word included in the extracted SNS posted message in a word cloud format. Note that the word cloud is a method of displaying words in sizes corresponding to the number of appearances.

[Operation of the SNS posted text extraction device]
The operation (extraction mode) of the SNS posted text extraction device 1 will be described with reference to FIG.
In addition, in FIG. 6, description will be made on the assumption that the emotion classifier has already been learned in the learning mode.

In step S1, the keyword setting unit 20 sets keywords for collecting SNS posted messages. For example, the keyword setting means 20 automatically sets words that appear frequently in SNS posts as keywords.
In step S2, the SNS posted message collection unit 30 collects SNS posted messages including the keyword set in step S1 for a predetermined collection period.

In step S3, the emotional polarity label adding means 41 calculates the emotional polarity score of each SNS posted message, and adds an emotional polarity label corresponding to the calculated emotional polarity score to each SNS posted message. That is, the emotion polarity label adding means 41 calculates the emotion polarity score of each SNS posted text using the learned emotion classifier in the learning mode, and adds the emotion polarity label with the highest emotion polarity score to each SNS posted text.

In step S4, the emotional polarity ratio calculation unit 51 totalizes the number of SNS posted texts for each aggregation period and emotional polarity label, and calculates the ratio of the aggregated number as the emotional polarity ratio. That is, the emotional polarity ratio calculation unit 51 calculates the emotional polarity ratio based on the emotional polarity label added in step S3 and the posting time of the SNS posted message.

In step S5, the burst determination means 52 determines whether or not an emotional polarity burst has occurred due to the increase in the emotional polarity ratio calculated in step S4. For example, the burst determination means 52 determines the occurrence of a burst using the Kleinberg burst detection technique described above.

In step S6, the reversal determining means 53 determines whether or not the reversal of the emotional polarity has occurred by comparing the ratios of the emotional polarities calculated in step S4. For example, the reversal determining means 53 determines whether or not the reversal of the positive and negative emotional polarities has occurred by the determination methods of the first to third examples described above.

In step S7, the SNS posted text extraction means 60 determines whether or not an emotional polarity burst or reversal has occurred.
Here, if a burst or reversal of emotional polarity occurs (Yes in step S7), the SNS posted text extraction device 1 proceeds to the process of step S8.
On the other hand, if the burst and the reversal of the emotional polarity do not occur (No in step S7), the SNS posted text extraction device 1 proceeds to the process of step S9.

In step S8, the SNS posted text extraction means 60 extracts SNS posted texts corresponding to the emotional polarity label, keyword, and counting period when the burst or reversal of the emotional polarity occurs, as highly topical SNS posted texts. .

In step S9, the SNS posted message extraction unit 60 determines whether or not to end extraction of SNS posted messages.
Here, if the SNS posted text extraction means 60 determines that the emotional polarity bursts or reverses in all the aggregation periods (Yes in step S9), it determines that the extraction of the SNS posted text is finished, and the process of step S10 is performed. proceed to
On the other hand, if the determination of emotional polarity burst or reversal has not been completed in all the aggregation periods (No in step S9), the SNS posted text extraction means 60 determines that the extraction of the SNS posted text is not finished, and Return to processing.

In step S10, the SNS posted text extraction unit 60 outputs the SNS posted text extracted in step S8.

[Action/effect]
As described above, the SNS-posted text extraction device 1 captures occurrences of bursts and emotional polarity reversals in SNS-posted texts as changes in the emotional polarity, so it is possible to accurately extract topical SNS-posted texts. That is, the SNS post extracting device 1 considers not only the number of reactions to the SNS post but also the emotions contained in the SNS post, and extracts information that has moved the SNS viewer's emotions (highly topical SNS post). ) can be extracted. Furthermore, since the user can set any keyword in the SNS posted text extraction device 1, it is possible to submit information specialized for the needs of the user.

Although the embodiments of the present invention have been described in detail above, the present invention is not limited to the above-described embodiments, and includes design changes and the like within the scope of the present invention.

(Modification 1: Emotion Polarity Dictionary)
In the above-described embodiment, the emotion classifier is learned, but the present invention is not limited to this. If it is difficult to prepare learning data in advance, a known emotional polarity dictionary (emotional polarity score information) can be used (references 2 and 3). This emotion polarity dictionary stores an emotion polarity label and an emotion polarity score for each word.

In this case, an emotion polarity dictionary is stored in advance in the storage means. Also, the emotion polarity label adding means calculates the average value of the emotion polarity scores of all the words included in the emotion polarity dictionary (emotion polarity score average value). Then, the emotion polarity label adding means refers to the emotion polarity dictionary, calculates the emotion polarity score of each word included in the SNS posted message, and totals the emotion polarity score for each SNS posted message. Furthermore, the emotional polarity label adding means calculates the emotional polarity score of the SNS posted message by dividing the total emotional polarity score by the number of words included in the SNS posted message. Furthermore, the emotional polarity labeling means classifies the emotional polarity score of the SNS posted text into binary values representing positive or negative, using the average emotional polarity score as a reference (boundary). Furthermore, the emotional polarity label adding means adds an emotional polarity label corresponding to the classified binary to each SNS post message.

Reference 2: Japanese Evaluation Polarity Dictionary, Tohoku University Inui Okazaki Laboratory, [online], [searched November 21, 2018], Internet <URL: http://www.cl.ecei.tohoku.ac .jp/index.php?Open%20Resources%2FJapanese%20Sentiment%20Polarity%20Dictionary>
Reference 3: Emotion Polarity Correspondence Table, Dai Takamura et al., [online], [searched November 21, 2018], Internet <URL: http://www.lr.pi.titech.ac.jp/~takamura /pndic_en.html>

Note that since the emotion polarity dictionary of Reference 2 does not include an emotion polarity score, the emotion polarity score may be manually set according to the emotion polarity. For example, if the emotional polarity is positive, the emotional polarity score is set to 1, and if the emotional polarity is negative, the emotional polarity score is set to -1.

(Modification 2: Emotional polarity of 10 axes)
In the above-described embodiment, the emotional polarities are described as having three axes of positive, neutral, and negative, but the present invention is not limited to this. For example, the SNS post extracting device can use 10 axes of emotional polarities (joy, anger, sorrow, fear, shame, like, disgust, anger, anxiety, surprise) (Reference 4). In this case, the reversal determination means may determine the reversal of the emotional polarity based on any combination of two or more of the ten emotional polarities.
Reference 4: Emotional Expression Dictionary, Akira Nakamura, Tokyodo Publishing Co., Ltd., May 1993

(Modification 3: Keyword ranking by number of appearances)
In the present invention, keywords may be ranked when setting keywords. For example, the keyword setting means ranks the words included in the SNS posted message in descending order of frequency of appearance as keywords. Then, if both the burst and the reversal of the emotional polarity do not occur, the SNS posted text extraction means switches to a low-rank keyword, recalculates the emotional polarity ratio, and re-determines the SNS posted text with high topicality. . As a result, the SNS posted text extraction device can more accurately extract SNS posted texts with high topicality.

(Modification 4: Ranking of keywords according to the maximum value of ratios of emotional polarities)
As shown in FIG. 7, in a certain keyword, the reversal of emotion polarity may occur multiple times. At this time, if an SNS posted message is extracted each time a reversal of the emotional polarity occurs, the number of SNS posted messages to be extracted becomes too large, which is troublesome for the user of the SNS posted message extraction device. Therefore, the keyword setting means may rank the keywords according to the maximum value of the ratio of the emotional polarities after the reversal of the emotional polarities.

FIG. 8 shows an example of the ratio of emotional polarities in the form of a table when a certain keyword is set. The status in FIG. 8 indicates the state of reversal of emotional polarity. Specifically, when the status is "0", it indicates that the reversal of the emotional polarity has not occurred during the counting period. Also, when the status is "1", it indicates that a reversal has occurred in which the percentage of positive emotional polarities is higher than the percentage of negative emotional polarities. At this time, the ratio in FIG. 8 is a value pos/neg obtained by dividing the ratio pos of the reversed positive emotional polarity by the ratio neg of the reversed negative emotional polarity. Also, when the status is "2", it indicates that a reversal has occurred in which the ratio of negative emotional polarities is higher than the ratio of positive emotional polarities. At this time, the ratio is neg/pos, which is the ratio neg of the reversed negative emotional polarity divided by the ratio pos of the reversed positive emotional polarity. The ranks in FIG. 8 are the results of ranking the ratio values from first to fifth. FIG. 8 shows that the maximum value of the ratio for a certain keyword is 2.778 when the aggregation period is the 14th hour.

The keyword setting means calculates the maximum value of the ratio of the emotional polarities for each of the set keywords by the procedure described above. Then, the keyword setting means may rank the keywords according to the maximum value of the ratio of the calculated emotional polarities.

(Modification 5: Deletion of keywords)
In the present invention, keywords that do not contribute to change in emotional polarity may be deleted. In other words, words with a large proportion of neutral emotional polarities and flattened words are deleted. Specifically, the keyword setting means deletes from the keywords words whose ratio of neutral emotional polarity is equal to or greater than a preset threshold value (for example, 90%). In addition, when the ratio of positive and negative emotional polarities does not vary within a predetermined range, the keyword setting means may delete the corresponding word from the keyword. As a result, the SNS posted text extraction device can set meaningful keywords that reflect the user's needs, and more accurately extract highly topical SNS posted texts.

(Other modifications)
In the above-described embodiment, the text information is social media information such as SNS posts, but the present invention is not limited to this.
In the above-described embodiment, the SNS posted text extraction device was described as independent hardware, but the present invention is not limited to this. For example, the present invention can also be implemented by a program that causes hardware resources such as a CPU, memory, and hard disk provided in a computer to operate cooperatively as the SNS posted message extraction device described above. These programs may be distributed via a communication line, or may be distributed by being written in a recording medium such as a CD-ROM or flash memory.

1 SNS posted text extraction device (text information determination means)
10 storage means 20 keyword setting means 30 SNS post collection means (text information input means)
40 Emotion classification means 41 Emotion polarity label addition means 50 Stationary value determination means 51 Emotion polarity ratio calculation means 52 Burst determination means 53 Reversal determination means 60 SNS post text extraction means (text information determination means)

Claims (8)

A text information determination device for determining highly topical text information based on changes in emotional polarity,
text information input means for inputting a plurality of the text information including preset keywords and to which time information is added;
emotion polarity label adding means for calculating an emotion polarity score of each piece of input text information and adding an emotion polarity label corresponding to the calculated emotion polarity score to each of the text information;
The number of cases of the text information is aggregated for each emotional polarity label in a preset aggregation period, and the ratio of the total number of cases of the text information to the number of cases of each emotional polarity label is calculated as the ratio of the emotional polarity. a polarity ratio calculating means;
burst determination means for determining whether or not a burst of the emotional polarity has occurred due to an increase in the ratio of the emotional polarity;
reversal determination means for determining whether or not the reversal of the emotional polarity has occurred by comparing the ratios of the emotional polarities;
text information determination means for determining the text information corresponding to the emotional polarity label, the keyword, and the aggregation period when the burst and/or the reversal occurs, as the text information having high topicality;
A text information determination device comprising: The emotional polarity label adding means includes:
learning an emotion classification model using learning data in which the emotion polarity label is added in advance to the text information during learning;
2. The text according to claim 1, wherein, at the time of determination, the emotion polarity score of each of the text information is calculated by the emotion classification model, and the emotion polarity label with the highest emotion polarity score is added to each of the text information. Information judgment device. The emotional polarity label adding means refers to emotional polarity score information in which the emotional polarity score for each word is stored, and determines the emotional polarity score of each text information and the emotional polarity score of each word included in the emotional polarity score information. Calculating an average emotional polarity score by averaging the emotional polarity scores, and classifying the emotional polarity score of each of the text information based on the average emotional polarity score, thereby obtaining an emotional polarity label corresponding to the classified emotional polarity score 2. The text information judging apparatus according to claim 1, wherein is added to each of the text information. The reverse rotation determination means is
Two emotional polarities to be determined for reversal are set in advance,
When the ratios of the two emotional polarities become equal, when the ratios of the two emotional polarities are exchanged, or after the ratios of the two emotional polarities are exchanged, the ratios of the two emotional polarities 4. The text information determination apparatus according to claim 1, wherein when the difference between the two is greater than or equal to a predetermined threshold value, it is determined that the reversal of the emotion polarity has occurred. keyword setting means for setting the keywords in order of the frequency of appearance of the words included in the text information, or in accordance with the maximum ratio of the emotional polarity ratios after the reversal of the emotional polarity occurs. further prepared,
1 to 4, wherein the text information determination means determines the text information with high topicality by switching to a keyword with a low number of appearances when neither the burst nor the reversal occurs. 5. The text information determination device according to any one of 4. 6. The text information determination apparatus according to claim 5, wherein said keyword setting means presets said emotional polarity representing neutral, and deletes words belonging to said neutral emotional polarity from said keyword. 7. The text information determination device according to any one of claims 1 to 6, wherein the text information is social media information to which posted time is added. A program for causing a computer to function as the text information determination device according to any one of claims 1 to 7.

Images