JP2014186451A - Emotion estimation device, emotion estimation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy of estimated emotion in estimating emotion of a user from an utterance of the user.SOLUTION: An emotion estimation device creates instructor data which is information showing a pair of [a word, a negative/positive value], in order to determine whether content of a user's utterance is negative or positive, for each user. In creating the instructor data, the emotion estimation device extracts an utterance including a negative word, and a related utterance word thereof (a related word), from the user's utterance. The extracted negative word is assigned with a prescribed negative/positive value, and the related word is assigned with a negative/positive value reduced in a negative degree. At this point, the emotion estimation device reduces the negative degree of a negative/positive value of each word if the user is a user of frequent negative utterances, whereas the negative degree of the negative/positive value of each word is increased if the user is a user of frequent positive utterances.

Description

  The present invention relates to an emotion estimation device, an emotion estimation method, and a program.

  2. Description of the Related Art Conventionally, there is a technique for estimating the emotion of a user who is a speaker by using, as teacher data, words that are generally used for negative speech and positive speech. For example, for each user, prepare as teacher data how to assemble words and sentences used when the user is angry, and use the teacher data to analyze the input text and voice There is a technique for determining whether or not a user is angry. According to this technique, when it is determined that the user (customer) is angry at the call center or the like, the fact can be notified to the call center staff or the like (see Patent Document 1).

JP 2012-113542 A

  However, the expression of intention display for each user is often biased based on the judgment criteria of the user (individual). For example, an utterance that includes a negative word of a user who always makes a negative utterance and an utterance that contains a negative word of a user who always makes a positive utterance, There is a high possibility that the negative meaning of the remarks is different among users. In other words, as in Patent Document 1, a technique for estimating a user's emotion using only general judgment criteria (generally negative utterances and words used for positive utterances) possessed by words is not necessarily the user's. Emotions cannot be estimated accurately. Therefore, an object of the present invention is to solve the above-described problems and improve the estimation accuracy of the user's emotion.

In order to solve the above-described problems, the present invention stores a past log recording unit that stores a past log that indicates the content of a message for each user, and an emotion notation word that is a phrase that indicates a specific emotion and an opposite emotion. A general emotion expression word recording unit, an emotion expression word search unit that refers to the general emotion expression word recording unit, and searches the user's past log for a statement including any of the emotion expression words, and the user From the past log, a front-rear speech extraction unit that extracts a peripheral speech that is a speech within a predetermined range around the speech including the searched emotion notation word, and for each user, each word included in the speech Using a feature quantity extraction unit that creates teacher data indicating a value of an emotion vector representing the degree of strength of the specific emotion, and using the teacher data of the user, A speech analysis unit that determines a tendency of the specific emotion that the user's speech means, and outputs the determination result, wherein the feature amount extraction unit determines the value of the emotion vector of each of the words , A value indicating the degree of strength of a predetermined emotion according to the type of emotion indicated by the emotion notation word included in the statement with respect to the emotion notation word included in the statement searched by the emotion notation word search unit The value of the emotion vector given to the emotion notation word is given as the value of the emotion vector of the emotion notation word, and the value of the emotion vector for each word other than the emotion notation word among the words of the surrounding speech of the statement In comparison, the emotion estimation apparatus is characterized in that a predetermined value that reduces the degree of strength of the specific emotion is given.

  According to such an emotion estimation apparatus, in addition to a commonly used emotion notation word for teacher data used for user speech analysis, a predetermined word is used for words used by the user in surrounding speech of the word (word). Gives the value of the emotion vector. That is, since the emotion estimation device creates teacher data reflecting the user's usual speech tendency, emotion accuracy when performing emotion estimation from the user's speech can be improved.

  According to the present invention, it is possible to improve the estimation accuracy of a user's emotion.

FIG. 1 is a diagram illustrating a configuration example of a system including an emotion estimation apparatus according to the present embodiment. FIG. 2 is a diagram illustrating a processing procedure example of the emotion estimation apparatus in FIG. 1. FIG. 3 is a diagram showing a processing procedure of the emotion estimation apparatus of FIG. FIG. 4 is a diagram showing details of the processing in S17 of FIG. FIG. 5 is a diagram illustrating a processing procedure of the speech analysis unit in FIG. FIG. 6 is a diagram showing a processing procedure of the teacher data creation unit of FIG. FIG. 7 is a diagram showing a processing procedure when the word extracted in S105 of FIG. 6 is a negative word. FIG. 8 is a diagram showing a processing procedure when the word extracted in S105 of FIG. 6 is a positive word. FIG. 9 is a diagram illustrating a computer that executes a program for realizing the function of the emotion estimation apparatus in FIG. 1.

(Embodiment)
Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, a configuration example of a system including the emotion estimation device 1 will be described with reference to FIG. For example, as shown in FIG. 1, the system includes an emotion estimation device 1 and a user terminal 2, and the emotion estimation device 1 and the user terminal 2 are connected by a network (such as the Internet or a LAN (Local Area Network)). The When the emotion estimation device 1 receives the user's message content (for example, the message content in Twitter (registered trademark), Facebook (registered trademark), blog, chat, etc.) from the user terminal 2 via the network, the emotion estimation device 1 A teacher model for analyzing the user's current utterance is created from the contents of the past log stored in the past log. The analysis of the utterance is to determine the tendency of the emotion that the user's utterance means. In the present embodiment, the emotion estimation device 1 is positive whether the utterance of the user means a negative emotion. An example will be described in which it is determined whether the emotions mean different emotions. In addition, this teacher model shows the value of an emotion vector indicating the degree of the strength of a specific emotion of each word for each user. In this embodiment, as the value of this emotion vector, A case where a negative / positive value is used will be described as an example. This negative / positive value indicates a positive degree (or negative degree), which is a degree of positive (or negative) emotional strength of each word, as a value of “0 to 1”. The emotion estimation apparatus 1 reflects the tendency of emotion (for example, whether there are a lot of negative statements or a lot of positive statements) in the negative / positive value of the teacher data. And the emotion estimation apparatus 1 estimates (analyzes) the emotion (for example, negative or positive) indicated by the user's current speech using the teacher model related to the user.

  Here, the outline of the emotion estimation apparatus 1 in FIG. 1 will be described with reference to FIG. The emotion estimation device 1 accumulates a past log (utterance log) of a speech for each user. Here, the past log is text data in which the utterance of the user is recorded in chronological order together with the date and time of utterance for each user. Here, the emotion estimation apparatus 1 extracts a utterance including the word “whatever” as a negative word and words in the vicinity thereof from the user's utterance, and sets a negative / positive value (negative degree (or positive) of each word to each word. A case will be described as an example in which a value indicating (degree) is assigned and recorded in the teacher data. In the following description, it is assumed that the negative / positive value is most negative “0” to neutral “0.5” to most negative “1”.

(S1)
The emotion estimation device 1 searches the user's past log, and extracts a comment including the negative word “whatever”. Here, let us consider a case where a comment “whatever I am” at time “10:04:30” is extracted.

(S2)
Next, the emotion estimation device 1 extracts a message (peripheral message) for several minutes before and after the message extracted in S1 from the past log (for example, “I am I anyway”). For example, as indicated by reference numeral 202, the emotion estimation apparatus 1 utters “Yabai”, “I couldn't do it at all” for several minutes around “Whatever I am” ”,“ Whatever I am, ”“. Extract the remarks of "mmmm".

(S3)
Thereafter, the emotion estimation device 1 performs morphological analysis on the user's utterance extracted in S1 and the surrounding utterance extracted in S2, and performs word extraction. And the emotion estimation apparatus 1 assign | provides the negative positive value provided to each extracted word. For example, the emotion estimation device 1 sets the negative / positive value of the general word (here, the negative word included in the utterance searched in S1) to “0” and the peripheral word (the word included in the peripheral utterance searched in S2). The negative / positive value is assigned with a relatively negative negative / positive value “0.2” (see reference numeral 203). That is, even if the word is not a negative word, if the word appears in the vicinity of the word, the word is likely to be used in a negative sense for the user. A negative / positive value with a slightly reduced negative degree than the negative / positive value is applied. Then, the emotion estimation device 1 records a pair of the extracted word and the assigned negative / positive value as teacher data of the user. Note that the negative / positive value assigned in S3 is adjusted depending on which of the user's daily statements is more negative (negative or positive).

  For example, when the emotion estimation device 1 extracts a utterance including the negative word “Anything” in S1, if the user who made the utterance including the negative word is a person who has the same negative / positive ratio of utterances in daily life, the general word (negative The negative / positive value of word) remains “0”, and the negative / positive values of the surrounding words remain “0.2”. On the other hand, if the user has a high percentage of negative (negative) utterances, the negative / positive value of a general word (negative word) is set to “0.4”, and the negative / positive values of surrounding words are set to “0.5”. To do. That is, the negative / positive value is corrected to be closer to the positive overall. Further, if the user usually has a lot of positive (positive) utterances, the negative / positive value of the general word (negative word) remains “0”, but the negative / positive value of the surrounding words becomes “0.1”. To. That is, the negative / positive value is corrected so as to be negative overall. Here, the reason why the negative / positive value of the general word (negative word) remains “0” is because the negative / positive value does not take a value less than zero.

  The emotion estimation apparatus 1 executes such processes of S1 to S3 every time a negative word is found from within a user's speech, and creates teacher data for the user.

(S4)
Then, the emotion estimation apparatus 1 analyzes whether the user's utterance is negative or positive using the teacher data of the user, and outputs the result. For example, when the emotion estimation apparatus 1 stores the analysis result in a predetermined area of the storage unit or determines that the user has made a negative statement as a result of the analysis, the emotion estimation apparatus 1 transmits the fact to the user terminal 2 of the user. Or Although explanation is omitted here, the emotion estimation apparatus 1 assigns a negative / positive value to each word by using the same procedure for the utterance including the positive word registered in advance and the word of the surrounding utterance. To record. The procedure at this time will be described later.

  Returning to FIG. 1, the emotion estimation apparatus 1 will be described in detail. Emotion estimation device 1 is realized by a computer including a storage unit, a control unit, and an input / output unit (all not shown). The storage unit stores various types of information (eg, negative / positive words and past logs) used by the control unit for creating teacher data and analyzing speech. The control unit mainly creates teacher data and performs speech analysis using the teacher data. The input / output unit accepts input of a user's speech to the emotion estimation device 1 or outputs an analysis result.

  The general negative / positive word recording unit 131, the past log recording unit 132, the teacher data recording unit 133, and the analysis result recording unit 134 of FIG. 1 are included in the storage unit described above, and the teacher data creation unit 121 and the speech analysis unit 125 is included in the control unit described above.

  The general negative / positive word recording unit (general emotion notation word recording unit) 131 stores a phrase (emotion notation word) indicating a specific emotion and an emotion opposite to that emotion. For example, the general negative / positive word recording unit 131 has, in a general sense, a negative word that is a word indicating a negative emotion, and a positive word that is a word that indicates a positive emotion in a general sense (collectively, “ Negative positive word "). For example, the negative word is “whatever” or the like, and the positive word is “happy” or “fun” or the like. This negative / positive word is registered via the input / output unit. In the following description, the case where the emotion notation word is the negative positive word will be described as an example, but it may of course be other than the negative positive word, for example, “happy, sad”, “like, dislike”, “ Phrases such as “safe and anxious”, “fun and scary”, and “cute and uncomfortable” may be used.

  The past log recording unit 132 stores a past log indicating the content of a message for each user. This past log is information indicating the content of the user's statement together with the date and time of the statement. This message content is input from the user terminal 2 via the input / output unit and stored, for example.

  The teacher data recording unit 133 stores the teacher data created by the teacher data creation unit 121. The teacher data is information indicating a negative / positive value indicating the negative degree (or positive degree) of each word of the user's utterance for each user. As described above, this negative / positive value is numerically expressed, for example, by setting the value of the most negative word as “0”, the value of the neutral word as “0.5”, and the value of the most positive word as “1”. It is. In other words, the negative positive value is closer to “0” as the negative degree is higher, and the negative positive value is closer to “1” as the positive degree is higher. Of course, the negative / positive value may be expressed by a method other than the above method.

  The analysis result recording unit 134 stores the analysis result of each user's utterance by the utterance analysis unit 125 (the analysis result of whether the utterance is negative or positive with respect to the user's utterance).

  The teacher data creation unit 121 includes a negative / positive word search unit 122, a forward / backward speech extraction unit 123, and a feature amount extraction unit 124.

  The negative / positive word search unit (emotion expression word search unit) 122 refers to the emotion expression word (for example, negative / positive word) stored in the general negative / positive word recording unit 131 and stores the user's past log (past log recording unit 132). ) To search for remarks including any of emotion expression words (for example, negative / positive words).

  The front / rear speech extraction unit 123 extracts peripheral speech that is a speech within a predetermined range before and after the speech including the searched negative / positive word. Peripheral utterances are, for example, utterances for several minutes before and after utterances including negative / positive words, and utterances for several words before and after. The range of the surrounding utterances can be changed according to the daily utterance frequency of the user. For example, for a user with a relatively high speech frequency, the range extracted as a peripheral speech may be widened.

  The feature amount extraction unit 124 creates teacher data in which a negative / positive value is assigned to each word of the user's utterance and surrounding words.

Details of the method for assigning a negative / positive value to each word here will be described later. For example, the following processing is performed. First, when the utterance searched by the negative / positive word search unit 122 is a utterance including a negative word, the feature amount extraction unit 124 (1) sets the negative / positive value assigned to the negative word included in the utterance as the most negative. Negative-positive value (for example, “0”). Next, (2) a negative positive value assigned to each word (neighboring word) of the peripheral utterance of this utterance is a value (eg, “0”) that is less negative than the negative positive value (eg, “0”) assigned to the negative word. , “0.2”). After that, the feature amount extraction unit 124 (3) If the user is a user who has a relatively large number of negative word utterances, the negative positive value determined in (1) and (2) is further reduced in the negative degree (for example, , Negative word “0.4”, peripheral word “0.5”). On the other hand, if the user is a user who has a relatively large number of positive word utterances, the feature amount extraction unit 124 increases the negative value (eg, negative word) for the negative / positive values determined in (1) and (2). “0” and surrounding word “0.1”). That is, even if it is the same word, the negative degree (positive degree) that the word means differs depending on the usual user's utterance tendency. Therefore, the feature amount extraction unit 124 determines the negative / positive of each word based on the utterance tendency of the user. Adjust the value. It should be noted that whether the user is a user who has a relatively large number of negative word utterances is determined by the feature amount extraction unit 124 and the user's past log (stored in the past log recording unit 132) and negative / positive word (general negative / positive word recording unit 131) To store).
That is, when the feature amount extraction unit 124 determines the negative / positive value of each word, the negative / positive word included in the utterance is indicated by a specific positive word included in the utterance searched by the general negative / positive word search unit 122. A value indicating the degree of strength of emotion (for example, positive emotion or negative emotion) is set as the negative / positive value of this word. In addition, the feature amount extraction unit 124 reduces the degree of emotional strength compared to the negative / positive value assigned to the negative / positive word as the negative / positive value for each word other than the negative / positive word among the words of the peripheral utterances of the utterance. To a predetermined value. Furthermore, the feature quantity extraction unit 124 refers to the user's past log and negative / positive word, and in the user's utterance, the utterance frequency of a word indicating a specific emotion (for example, negative) is higher than a predetermined threshold. If it is determined, the negative / positive value given to the negative / positive word is corrected to a value that reduces the degree of strength of a specific emotion (for example, negative emotion), and an emotion (for example, If it is determined that the utterance frequency of a word indicating “positive emotion” is higher than a predetermined threshold, the degree of strength of a specific emotion (for example, negative emotion) is determined for the negative-positive value assigned to the negative-positive word. Correct to an increased value.

  As described above, the feature amount extraction unit 124 reduces the negative degree indicated by the negative / positive value of the word if the user who made the utterance including the word even in the same negative meaning is the user who usually makes a negative utterance. If the user who made a statement including the word usually makes a positive statement, the negative degree indicated by the negative / positive value of the word is increased. That is, the feature quantity extraction unit 124 adjusts the negative / positive value of the teacher data in consideration of whether each user has many negative utterances or many positive utterances. Can be created.

  The utterance analysis unit 125 uses the teacher data of the user, and for the user's utterance input from the user terminal 2 via the input / output unit, an emotion (for example, negative emotion or positive emotion) that the utterance means ). For example, when the negative degree indicated by the negative / positive value of the user's utterance (the average value of the negative / positive values of the words included in the utterance) exceeds a predetermined threshold, the utterance analysis unit 125 analyzes that the user's utterance is negative. (Judgment result) is output. This analysis result is output to the analysis result recording unit 134 and the like. In addition, the speech analysis unit 125 may transmit the analysis result to the user terminal 2 via the warning transmission unit 126 (described later).

  The warning transmission unit 126 and the result display unit 127 may or may not be equipped, and the case of being equipped will be described later.

  Next, the outline of the processing procedure of the emotion estimation apparatus 1 will be described with reference to FIG. First, as a preliminary preparation, the emotion estimation device 1 registers a negative / positive word in the general negative / positive word recording unit 131 (S11). Moreover, the emotion estimation apparatus 1 accumulate | stores the past remark log for every user in the past log recording part 132 (S12).

  Then, the teacher data creation unit 121 refers to the negative positive word registered in the general negative / positive word recording unit 131 and the past log of the user's utterance accumulated in the past log recording unit 132, and the teacher data of the user Is created (S13). Details of the teacher data creation procedure will be described later. Then, the teacher data creation unit 121 stores the created teacher data in the teacher data recording unit 133 (S14).

  When a utterance is input from the utterance input unit 21 of the user terminal 2 (S15) and the utterance analysis unit 125 of the emotion estimation apparatus 1 receives this utterance (S16), the teacher data of the user who has made the utterance of this utterance is recorded as teacher data. The message is read from the unit 133, and the utterance is analyzed with reference to the teacher data (S17). Then, the statement analysis unit 125 outputs the analysis result to the analysis result recording unit 134 (S18).

  Next, S13 of FIG. 3 will be described in detail with reference to FIG. First, the feature amount extraction unit 124 of the teacher data creation unit 121 indicates how often the user is making a negative statement or a positive statement from the past log of each user's statement. Analyze. That is, the teacher data creation unit 121 first acquires each user's past log from the past log recording unit 132 (S21). Then, the feature quantity extraction unit 124 refers to the negative / positive words registered in the general negative / positive word recording unit 131 and analyzes the negative / positive of the past log of the user (S22). That is, the feature amount extraction unit 124 refers to the negative / positive words registered in the general negative / positive word recording unit 131 and determines the negative / positive (negative or positive) of each word appearing in the user's past log. Then, the feature amount extraction unit 124 counts the appearance frequencies of the negative words and the positive words in the entire past log of the user (S23), and uses the counted appearance frequencies as negative / positive frequency information in the predetermined storage unit of the emotion estimation device 1. Record in the area (S24).

  Further, the negative / positive word search unit 122 of the teacher data creation unit 121 refers to the negative / positive word shown in the general negative / positive word recording unit 131 and the past log of the user's utterance accumulated in the past log recording unit 132, An utterance including a negative / positive word is extracted from the user's past log (S25), and an utterance hit by this search and its utterance date and time are output to the preceding and following utterance extraction unit 123 (S26).

  The forward / backward utterance extraction unit 123 extracts utterances (peripheral utterances) for several minutes before and after the utterance hit in the search from the user's past log (S27). Then, the utterance searched in S25 and its peripheral utterances are output to the feature amount extraction unit 124 (S28).

  The feature quantity extraction unit 124 assigns a negative / positive value of each word included in the utterance extracted in S25 and its peripheral utterances (S29). That is, a pair of [word, negative / positive value] is recorded in the teacher data of the user. It should be noted that the feature amount extraction unit 124 assigns a negative / positive value of each word in S24 to a value obtained by reducing the negative degree (or positive degree) of the negative / positive value of the words of surrounding utterances than the negative word (or positive word). To do. For example, the negative / positive value of the utterances around the negative word is set to “0.2”, which is a value obtained by reducing the negative degree than “0 (value having the highest negative degree)”. Further, the negative / positive value of the peripheral utterances of the positive word is set to “0.8” which is a value obtained by reducing the positive degree than “1 (value having the highest positive degree)”.

  Further, the feature amount extraction unit 124 refers to the negative / positive frequency information recorded in S24, and determines whether the user is a user with many negative utterances or a user with many positive utterances. For example, the feature amount extraction unit 124 refers to the negative / positive frequency information described above, and determines that the user has many negative utterances when the utterance frequency of the negative word of the user is higher than a predetermined threshold. On the other hand, when the utterance frequency of the positive word of the user is higher than a predetermined threshold, it is determined that there are many positive utterances. The feature amount extraction unit 124 adjusts the negative / positive value depending on whether the user has a lot of negative utterances or a user who has a lot of positive utterances. Details of the adjustment of the negative / positive value will be described later. When the feature value extraction unit 124 determines the negative / positive value of each word, the feature value extraction unit 124 records the determined negative / positive value and the word pair in the teacher data of the user.

  Next, S17 in FIG. 3 will be described in detail with reference to FIG. As shown in FIG. 5, when the speech analysis unit 125 of the emotion estimation device 1 receives a speech input via the speech input unit 21 of the user terminal 2, based on the teacher data of the user of the user terminal 2, The utterances are classified into positive utterances, negative utterances, and neutral utterances, and stored in the analysis result recording unit 134. The user's speech classification is, for example, a value a obtained by the speech analysis unit 125 dividing the total negative / positive value of each word included in the speech by the number of words, that is, an average value of negative positive values per word. Is performed by determining which range shown in Table 1 below.

  Note that the categorization of comments is not limited to the above three stages. For example, it may be classified into five levels of positive, slightly positive, neutral, slightly negative, and negative, or may be classified at a higher level. In this case, it is assumed that a range of a values corresponding to each classification is set in the emotion estimation device 1.

  Next, details of the processing from S25 to S29 in FIG. 4 will be described with reference to FIG. 1 and FIG. First, when the teacher data creation unit 121 in FIG. 1 detects that a new message log is stored in the past log recording unit 132 (S100), the negative / positive word search unit 122 of the teacher data creation unit 121 reads the message. It is determined whether or not the teacher data of the user who made the statement is the first time (S101). When it is determined to be the first time (Yes in S101), the negative / positive word search unit 122 searches the oldest log (utterance log) of the user from the past log recording unit 132 in chronological order and registers it in the general negative / positive word recording unit 131. The utterance including the negative word or the positive word being extracted is extracted (S102). On the other hand, if it is determined that the teacher data is not created for the first time (No in S101), the negative / positive word search unit 122 searches the past log recording unit 132 from the previously searched location in chronological order, and the general negative / positive word recording unit A negative word or a positive word registered in 131 is extracted (S103).

  After S101 and S102, the forward / backward utterance extraction unit 123 acquires utterances (peripheral utterances) for several minutes before and after the utterance extracted in S102 or S103 from the past log of the user in the past log recording unit 132 (S104).

  After S104, the feature quantity extraction unit 124 refers to the negative / positive word indicated in the general negative / positive word recording unit 131 and determines that the word extracted in S102 or S103 is a negative word (Yes in S105). It progresses to S110 of 7. On the other hand, when it is determined that the word extracted in S102 or S103 is not a negative word (No in S105), the process proceeds to S120 in FIG.

  In S110 of FIG. 7, the feature amount extraction unit 124 sets the negative / positive value of a general word (referred to as “general negative word” because the general word is a negative word here) to “0”, which is the most negative value, and its surroundings. The negative / positive value of the word is set to “0.2”, which is a relatively negative value. Thereafter, the feature amount extraction unit 124 refers to the user's past log and the negative / positive word registered in the general negative / positive word recording unit 131 to determine whether the user of the utterance usually has many negative utterances. (S111). Various determination methods are conceivable here. For example, the feature amount extraction unit 124 refers to the user's past log and determines each user's remark based on the determination criteria such as Table 1 described above. Categorized as positive, negative, or neutral. Then, the feature amount extraction unit 124 determines that the user who has the most negative utterances among the three types of utterances is the user who usually has many negative utterances (Yes in S111), and the user's general negative word The negative / positive value is corrected from “0” to “0.4”, and the negative / positive value of the surrounding word is corrected from “0.2” to “0.5” (S112). That is, the feature amount extraction unit 124 reduces the negative degree of the negative / positive values of the general negative word and its surrounding words assigned in S110. In other words, the negative / positive values of the general negative word and the surrounding words are corrected closer to positive.

  On the other hand, the feature amount extraction unit 124 determines that a user whose number of utterances is not a negative utterance among the three types of utterances is usually a user who does not have many negative utterances (No in S111). The value is left as “0”, and the negative / positive value of surrounding words is corrected from “0.2” to “0.1” (S113).

  The feature amount extraction unit 124 determines whether or not the word that gave the negative / positive value in S112 and S113 has already been registered in the teacher data of the user (S114), and if not yet registered (Yes in S114), A pair of [word, negative / positive value] is registered in the teacher data of the user (S115). Then, the process returns to S100. On the other hand, if the word that gave the negative / positive value in S112 and S113 has already been registered in the teacher data of the user (No in S114), the negative / positive value given in S112 or S113 was calculated as -0.1. A pair of [word, negative / positive value] to which a value is assigned is registered in the teacher data of the user (S116). Then, the process returns to S100. However, the negative / positive value is 0 or more. If the value is smaller than 0 as a result of the calculation in S116, the value is 0. In this way, for example, each time a word around a negative word is extracted as a word around the negative word, the negative / positive value of the word decreases (increases the negative degree). Therefore, the accuracy of the negative / positive value in the teacher model can be improved.

  Next, the processing after S120 in FIG. 8 will be described. The feature amount extraction unit 124 sets the negative / positive value of a general word (here, the general word is a positive word and is referred to as “general positive word”) to be the most positive value “1”, and the negative / positive values of the surrounding words are A relatively positive value “0.8” is set (S120). Thereafter, the feature amount extraction unit 124 refers to the user's past log and the negative / positive word registered in the general negative / positive word recording unit 131 to determine whether the user of the utterance usually has many positive utterances. (S121). Similar to S111, the feature amount extraction unit 124 makes the determination here based on which of the three types of statements is the most. Here, when the feature amount extraction unit 124 determines that the user is a user who usually has many positive utterances (Yes in S121), the negative / positive value of the user's general positive word is changed from “1” to “0.6”. The negative / positive value of the surrounding word is corrected from “0.8” to “0.5” (S122). That is, the feature amount extraction unit 124 reduces the positive degree of the negative / positive value of the general positive word and its surrounding words assigned in S110. That is, the negative / positive values of the general positive word and the surrounding words are corrected toward negative.

  On the other hand, when the feature amount extraction unit 124 determines that the user is a user who usually does not have many positive utterances (No in S121), the negative / positive value of the general positive word remains “1” and the negative / positive value of the surrounding words. Is corrected to "0.9" (S123). Then, the process returns to S100.

  The feature amount extraction unit 124 determines whether or not the word to which the negative / positive value is given in S112 and S113 is not already registered in the teacher data of the user (S124), and if not registered (Yes in S124). , [Word, negative / positive value] pair is recorded in the teacher data of the user (S125). On the other hand, if the word that gave the negative / positive value in S112 and S113 has already been registered in the teacher data of the user (No in S124), the negative / positive value given in S122 or S123 +0.1 is calculated, and the calculated value The pair of [word, negative / positive value] to which is added is registered in the teacher data of the user (S126). Then, the process returns to S100. However, the negative / positive value is set to a value of 1 or less, and when the value is larger than 1 as a result of the calculation in S126, the value is set to “1”. By doing so, for example, each time a word around a positive word is extracted as a word around the positive word, the negative / positive value of the word increases (increases the degree of positiveness). Therefore, the accuracy of the negative / positive value in the teacher model can be improved.

  As described above, the teacher data creation unit 121 of the emotion estimation apparatus 1 creates teacher data for each user in consideration of whether the user has many negative utterances or many positive utterances for each user. That is, the teacher data creation unit 121 creates teacher data with high emotion estimation accuracy.

  Note that the feature amount extraction unit 124 of the emotion estimation device 1 may adjust the negative / positive value after assigning the negative / positive value of each word of the teacher data based on the processing shown in FIGS. For example, the feature amount extraction unit 124 may include words that appear in the peripheral utterances of the utterances including the positive words in the past log of the user among the words included in the negative words and the utterances around the negative words in the user's teacher data. For example, the negative / positive value of the word is corrected to a value that reduces the negative degree. That is, since the feature amount extraction unit 124 may be a relatively negative utterance, the negative degree may not be so high for words that appear as peripheral words of the positive word in the user's past utterances. Correct the negative / positive value to a value closer to positive.

  The feature amount extraction unit 124 also includes words that appear in peripheral utterances of utterances including negative words in the user's past log among words included in positive words and utterances around the positive words in the user's teacher data. Then, the negative / positive value of the word is corrected to a value that reduces the positive degree. That is, even if the feature amount extraction unit 124 is a relatively positive utterance, there is a possibility that the positive degree is not so high for a word that appears as a peripheral word of the negative word in the past utterance of the user. The negative / positive value is corrected to a more negative value. Here, however, the negative / positive value is between 0 and 1. That is, the negative / positive value reaches a peak at “1” and bottoms out at “0”.

  By adjusting the negative / positive value in this way, the feature amount extraction unit 124 can further improve the accuracy of the user's teacher data.

  The emotion estimation device 1 may further include a warning transmission unit 126 as shown in FIG. If the message analysis unit is included, the warning transmission unit 126 transmits a warning to that effect to the user terminal 2. For example, the emotion estimation apparatus 1 receives a user's utterance input via the utterance input unit 21 of the user terminal 2 in real time, and the utterance analysis unit 125 analyzes the received utterance. Then, as a result of the utterance analysis, if the utterance includes a negative utterance, the warning transmission unit 126 transmits a warning to the user terminal 2. This warning is displayed on a display device (not shown) or the like by the warning display unit 22 of the user terminal 2. Thereby, the user of the user terminal 2 can know in real time whether he / she has made a negative statement. The emotion estimation device 1 receives a setting as to whether or not to transmit such a warning from the user terminal 2 and issues a warning to the user terminal 2 that is set to transmit such a warning. You may make it transmit.

  Moreover, the emotion estimation apparatus 1 may be provided with the result display part 127 shown in FIG. In response to a request from the user terminal 2, the result display unit 127 transmits a speech analysis result in the past predetermined period of the user. When the result display unit 127 accepts a browsing request for the analysis result of the speech in the past predetermined period from the browsing unit 23 of the user terminal 2, the result display unit 127 displays the analysis result of the user in the past predetermined period stored in the analysis result recording unit 134. It transmits to the user terminal 2. The user terminal 2 that has received the analysis result causes the browsing unit 23 to display the analysis result on a display device or the like. Thereby, the user of the user terminal 2 can know the transition of emotion (negative, positive) in his / her past utterance.

  Further, the negative / positive words registered in the general negative / positive word recording unit 131 of the emotion estimation apparatus 1 include not only negative words and positive words in a general sense but also negative words and positive words set by the respective users. Also good. For example, when the word “subtle” is registered as the negative word of the user A, when the emotion estimation apparatus 1 accepts an input of a utterance including the word “subtle” of the user A, the negative / positive value of the word Is “0”. As described above, when the general negative / positive word recording unit 131 further registers a negative / positive word specific to the user, the feature amount extracting unit 124 can create teacher data with high emotion estimation accuracy.

  Further, the general negative / positive word recording unit 131 may further include setting information of a negative / positive value to be assigned to each word of the negative / positive word. For example, the general negative / positive word recording unit 131 further includes information that the negative / positive value is set to “0” for the word “anything”, but the negative / positive value is set to “0.1” for the word “troublesome”. You may go out. That is, in the above-described embodiment, the feature amount extraction unit 124 assigns a negative positive value “0” to the negative words registered in the general negative / positive word recording unit 131 among the words appearing in the utterance (see FIG. 7 (see S110 in FIG. 7), a negative positive value “1” is assigned to the positive word (see S120 in FIG. 8), but the negative / positive value of the word set in the general negative / positive word recording unit 131 is assigned. Also good. In this way, the feature quantity extraction unit 124 can set a finer value as the negative / positive value of each word in the teacher data, so that teacher data with high emotion estimation accuracy can be created.

  In the above-described embodiment, the negative / positive frequency information created by the feature amount extraction unit 124 indicates the appearance frequency of a word having a negative / positive value in that range for each range of negative / positive values of the word included in the user's utterance. Information (see Table 2). The creation of the negative / positive frequency information here is performed with reference to the teacher data of the user.

  Then, the feature amount extraction unit 124 calculates a score based on the appearance frequency value for each range of the negative / positive value, and based on the score value, each user has a lot of negative statements or is positive. You may make it judge whether there are many remarks (or it is neither of them).

  Further, the processing executed by the emotion estimation apparatus 1 described in the above embodiment may be realized by a program described in a language that can be executed by a computer. In this case, the same effect as the embodiment can be obtained by executing the program by the computer. Further, such a program may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read by a computer and executed to execute the same processing as in the above-described embodiment. Below, an example of the computer which performs the program which implement | achieves the function similar to the emotion estimation apparatus 1 shown in FIG. 1 is demonstrated.

  As illustrated in FIG. 9, the computer 1000 includes, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These units are connected by a bus 1080.

  The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012 as illustrated in FIG. The ROM 1011 stores a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1031 as illustrated in FIG. The disk drive interface 1040 is connected to the disk drive 1041 as illustrated in FIG. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive. The serial port interface 1050 is connected to a mouse 1051 and a keyboard 1052, for example, as illustrated in FIG. The video adapter 1060 is connected to a display 1061, for example, as illustrated in FIG.

  Here, as illustrated in FIG. 9, the hard disk drive 1031 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program is stored in, for example, the hard disk drive 1031 as a program module in which a command to be executed by the computer 1000 is described.

  The various data described in the above embodiment is stored as program data, for example, in the memory 1010 or the hard disk drive 1031. Then, the CPU 1020 reads the program module 1093 and program data 1094 stored in the memory 1010 and the hard disk drive 1031 to the RAM 1012 as necessary, and executes an access monitoring procedure, an access control procedure, a process monitoring procedure, and a process control procedure.

  Note that the program module 1093 and the program data 1094 related to the program are not limited to being stored in the hard disk drive 1031, but may be stored in, for example, a removable storage medium and read out by the CPU 1020 via the disk drive or the like. . Alternatively, the program module 1093 and the program data 1094 related to the monitoring program are stored in another computer connected via a network (LAN, WAN (Wide Area Network), etc.) and read by the CPU 1020 via the network interface 1070. May be.

DESCRIPTION OF SYMBOLS 1 Emotion estimation apparatus 2 User terminal 21 Statement input part 22 Warning display part 23 Browse part 121 Teacher data creation part 122 Negative positive word search part (Emotion expression word search part)
123 Pre- and post-utterance extraction unit 124 Feature amount extraction unit 125 Speech analysis unit 126 Warning transmission unit 127 Result display unit 131 General negative / positive word recording unit (general emotion expression word recording unit)
132 Past log recording unit 133 Teacher data recording unit 134 Analysis result recording unit

Claims (8)

A past log recording unit for storing a past log indicating the content of a message for each user;
A general emotion notation word recording unit for storing an emotion notation word which is a phrase indicating a specific emotion and an opposite emotion;
Referring to the general emotion expression word recording unit, an emotion expression word search unit for searching a remark including any of the emotion expression words from the past log of the user;
Before-and-after utterance extraction unit for extracting a peripheral utterance that is a utterance within a predetermined range around the utterance including the searched emotion expression word from the user's past log;
A feature amount extraction unit that creates teacher data indicating a value of an emotion vector representing the degree of strength of the specific emotion of each word included in the utterance for each user;
Using the teacher data of the user, and determining a tendency of the certain emotion that means the user's remarks, and a remark analysis unit that outputs the determination result,
The feature amount extraction unit includes:
When determining the value of the emotion vector for each of the words, a predetermined value corresponding to the type of emotion indicated by the emotion expression word included in the statement with respect to the emotion expression word included in the statement searched by the emotion expression word search unit A value indicating the degree of emotion strength is given as the value of the emotion vector of the emotion notation word, and the value of the emotion vector for each of the words other than the emotion notation word among the words of the surrounding speech of the statement, An emotion estimation apparatus, characterized in that a predetermined value that reduces the degree of intensity of the specific emotion compared to the value of the emotion vector assigned to an emotion notation word is provided. The feature amount extraction unit further includes:
With reference to the user's past log and the emotion notation word, in the user's remarks, when it is determined that the remark frequency of the emotion notation word indicating the specific emotion is greater than a predetermined threshold, The value of the emotion vector assigned to the emotion notation word is corrected to a value that reduces the degree of strength of the specific emotion, and the frequency of the emotion notation word indicating an emotion that conflicts with the specific emotion Is determined to be greater than a predetermined threshold, the value of the emotion vector assigned to the emotion notation word is corrected to an increased value of the strength of the specific emotion. The emotion estimation apparatus according to claim 1. The feature amount extraction unit further includes:
When the emotion type indicated by the emotion notation word included in the statement is negative,
As the value of the emotion vector for the emotion notation word indicating negative emotion, a predetermined value as a negative degree indicating the degree of strength of negative emotion is given, and the emotion among the words of the peripheral statements of the statement As a value of the emotion vector for each word other than a notation word, a predetermined value with a reduced negative degree compared to the value of the emotion vector given to the emotion notation word indicating the negative emotion is given,
When the emotion type indicated by the emotion notation word included in the statement is positive,
As a value of the emotion vector for the emotion notation word indicating positive emotion, a predetermined value as a positive degree indicating the degree of strength of positive emotion is given, and the emotion among the words of the peripheral statements of the statement The emotion vector value for each word other than the notation word is given a predetermined value with a reduced positive degree compared to the emotion vector value given to the emotion notation word indicating the positive emotion, The emotion estimation apparatus according to claim 1 or 2. The feature amount extraction unit further includes:
In the user's utterance, when it is determined that the utterance frequency of a negative word, which is an emotion expression word indicating negative emotion as the specific emotion, is higher than a predetermined threshold, the emotion vector assigned to the negative word The value of the positive word is corrected to a value that reduces the negative degree, which is the degree of strength of negative emotion, and the utterance frequency of a positive word that is an emotion expression word indicating a positive emotion opposite to the negative emotion is a predetermined threshold value If it is determined that there are more than, the value of the emotion vector given to the emotion expression word that is the negative word is corrected to a value with a higher negative degree,
In the user's utterance, when it is determined that the utterance frequency of a positive word that is an emotion expression word indicating a positive emotion as the specific emotion is higher than a predetermined threshold, the emotion vector assigned to the positive word The value of the negative word is corrected to a value that reduces the positive degree, which is the degree of the strength of positive emotion, and the negative word utterance frequency indicating the negative emotion opposite to the positive emotion is a predetermined threshold value. If it is determined that there are more than, the value of the emotion vector assigned to the emotion notation word that is the positive word is corrected to a value with a higher positive degree.
The emotion estimation apparatus according to claim 3, wherein: The feature extraction unit further includes:
Among the words included in the negative word and the surrounding utterances in the user's teacher data, the negative emotion value of the word that also appears in the peripheral utterance of the utterance including the positive word in the user's past log. The emotion estimation device according to claim 4, wherein the emotion is corrected to a value with reduced degree. The feature extraction unit further includes:
Among the words included in the positive word and the surrounding utterances in the user's teacher data, positive about the emotion vector value of the word that also appears in the peripheral utterances of the utterance including the negative word in the user's past log The emotion estimation device according to claim 4 or 5, wherein the emotion is corrected to a value with a reduced degree. Emotion estimation device
Refers to the general emotion notation word recording unit that stores the emotion notation word, which is a phrase indicating a specific emotion and the opposite emotion, and indicates negative or positive emotion from the past log showing the content of each user's utterance An emotion expression word search step for searching for a statement including any of the emotion expression words that are phrases;
Before-and-after utterance extraction step of extracting peripheral utterances that are utterances within a predetermined range around the utterance including the searched emotion expression word from the user's past log;
A feature amount extraction step for creating teacher data indicating a value of an emotion vector representing the degree of the strength of personal emotion that each of the words of the speech means for each user,
Using the teacher data of the user, determining a tendency of the certain emotion that the user's statement means, and executing a statement analysis step of outputting the determination result;
The feature amount extraction step includes:
When determining the value of the emotion vector for each of the words, a predetermined value corresponding to the type of emotion indicated by the emotion notation word included in the statement with respect to the emotion notation word included in the statement searched by the emotion notation word search step A value indicating the degree of emotion strength is given as the value of the emotion vector of the emotion notation word, and the value of the emotion vector for each of the words other than the emotion notation word among the words of the surrounding speech of the statement, An emotion estimation method comprising: assigning a predetermined value in which the degree of strength of the specific emotion is reduced compared to the value of the emotion vector assigned to an emotion notation word. Emotion estimation device
Refers to the general emotion notation word recording unit that stores the emotion notation word, which is a phrase indicating a specific emotion and the opposite emotion, and indicates negative or positive emotion from the past log showing the content of each user's utterance An emotion expression word search step for searching for a statement including any of the emotion expression words that are phrases;
Before-and-after utterance extraction step of extracting peripheral utterances that are utterances within a predetermined range around the utterance including the searched emotion expression word from the user's past log;
A feature amount extraction step for creating teacher data indicating a value of an emotion vector representing the degree of the strength of personal emotion that each of the words of the speech means for each user,
Using the teacher data of the user to determine the tendency of the certain emotion that the user's statement means, and to execute a statement analysis step of outputting the determination result;
The feature amount extraction step includes:
When determining the value of the emotion vector for each of the words, a predetermined value corresponding to the type of emotion indicated by the emotion notation word included in the statement with respect to the emotion notation word included in the statement searched by the emotion notation word search step A value indicating the degree of emotion strength is given as the value of the emotion vector of the emotion notation word, and the value of the emotion vector for each of the words other than the emotion notation word among the words of the surrounding speech of the statement, A program for assigning a predetermined value in which the degree of strength of the specific emotion is reduced compared to the value of the emotion vector assigned to an emotion notation word.

Images