JP2016090776A - Response generation apparatus, response generation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a response generation apparatus, a response generation method, and a program, which estimate an emotion in different method in accordance with a degree of intimacy with a user, thus generating a response.SOLUTION: A response generation apparatus 100 includes: a voice recognition part 101 for recognizing utterance content of a user; an emotion estimation part 104 for estimating a category of emotion of the user on the basis of the utterance content; a response generation part 106 for generating a response sentence in accordance with the estimated category of emotion; and an emotion category database 105 for storing a degree of intimacy with a user correlated with a combination of plural categories of emotions. In the emotion category database 105, the categories of emotion included in combinations are different depending on a difference in the degree of intimacy with a user. Referring to the emotion category database 105, the emotion estimation part 104 selects a category of emotion in accordance with utterance content from combinations in accordance with the degrees of intimacy, thus estimating the emotion of the user.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a response generation apparatus, a response generation method, and a program, and more particularly to a technique for generating a response by performing emotion estimation according to intimacy in a voice interactive robot or the like.

  Patent Document 1 discloses a robot apparatus that performs dialogue with emotion. The robot apparatus holds a closeness database, and updates the closeness between the user and the robot apparatus in accordance with the user's interaction. The interaction refers to how the user touches the robot apparatus, and refers to actions such as stroking, hitting, providing a meal, talking, playing a ball, and the like. The robot apparatus updates the intimacy based on the number and time of these interactions.

  Also, the robot apparatus recognizes the user's utterance and selects a response sentence corresponding to the familiarity from the response sentence list defined in the dialogue database. And the response which changed the ending and the inflection of the response sentence according to the emotion estimated from the user's utterance content is performed.

  Here, in the response sentence list used by the robot apparatus, the types and number of emotions are constant regardless of the familiarity.

JP 2004-090109 A

  The robot apparatus described in Patent Literature 1 selects a response sentence defined according to the familiarity and the user's emotion from a list. Here, the type and number of emotions associated with intimacy are constant regardless of intimacy. For this reason, there is a problem that variations in response from the robot apparatus side are limited.

  That is, since the types and number of emotions that can be estimated by the robot apparatus are constant, even if the robot device is more intimate with the user, it cannot respond to the emotion of the user more than a certain amount. In addition, even though it is not so intimate, it sometimes gives a sense of incongruity due to too much emotion, making conversation difficult to maintain a sense of distance according to intimacy. This is problematic in that the degree of empathy for the user's emotion cannot be changed according to the familiarity, and an appropriate response has not been selected.

  The response generation apparatus according to the present invention includes a voice recognition unit that recognizes a user's utterance content, an emotion estimation unit that estimates a user's emotion type based on the utterance content, and the estimated emotion type. A response generation unit that generates a response sentence, and further includes an emotion classification database in which a closeness with the user and a combination of a plurality of emotion types are associated with each other, The types of emotions included in the combination are different according to a difference in intimacy with the user, and the emotion estimation unit refers to an emotion classification database from the combination according to the intimacy, The user's emotion is estimated by selecting the type of emotion according to the utterance content.

  That is, the response generation apparatus of the present invention estimates the familiarity with the user, changes the emotion classification for emotion estimation according to the familiarity, and changes the response accordingly. Specifically, when the closeness with the user is high, a deeper emotion estimation is performed and a response is made by using a more detailed emotion type. On the other hand, when the intimacy is low, superficial shallow emotion estimation and impression estimation are performed and a response is made.

  According to the present invention, it is possible to provide a response generation device, a response generation method, and a program that perform emotion estimation using different methods according to intimacy with a user and generate a response.

It is a figure which shows the structure of the response production | generation apparatus 100 concerning Embodiment. It is a figure which shows operation | movement of the response generation apparatus 100 concerning embodiment.

Embodiments of the present invention will be described below with reference to the drawings.
First, the configuration of the response generation apparatus 100 according to the embodiment of the present invention will be described using the block diagram of FIG.

  The response generation device 100 is a device that recognizes a user's utterance and returns a response, and is typically a voice interactive robot device. The response generation device 100 includes, for example, a central processing unit (CPU), a volatile or nonvolatile memory, a voice input / output device (such as a microphone, a speaker, an A / D and a D / A converter), and the like. By executing information processing according to the program stored in the program, various target functions can be realized.

  The response generation device 100 includes at least a speech recognition unit 101, a familiarity calculation unit 102, a familiarity database 103, an emotion estimation unit 104, an emotion classification database 105, and a response generation unit 106. Furthermore, a speech synthesizer 107 may be included.

  The voice recognition unit 101 performs processing for recognizing the user's utterance content. That is, the voice recognition unit 101 performs a process of inputting a user's utterance as voice data and generating text data from the voice data. Typically, the microphone acquires the user's utterance as an analog voice signal, and the A / D converter converts the analog voice signal into voice data and inputs the voice data to the voice recognition unit 101. Generation of text data from voice data can be realized by various known voice recognition techniques.

  The emotion estimation unit 104 analyzes the content of the user's utterance using the text data generated by the voice recognition unit 101 and performs a process of estimating the user's emotion. The intimacy calculating unit 102 performs processing for calculating intimacy with the user using information related to the interaction with the user, including the emotion estimation result by the emotion estimating unit 104. The intimacy calculation unit 102 records information and intimacy regarding the dialogue with the user in the intimacy database 103 described later.

  Alternatively, the familiarity calculation unit 102 may calculate the familiarity with the user according to the interaction with the user, as in the related art described in Patent Document 1. In this case, the familiarity calculation unit 102 can update the familiarity without requiring emotion estimation processing by the emotion estimation unit 104.

  The intimacy database 103 is a storage unit that records information related to user interaction and intimacy.

  The emotion estimation unit 104 estimates the user's emotion again with reference to the emotion classification database 105 described later after the familiarity calculation unit 102 calculates the familiarity. That is, the type of emotion is specified. The higher the intimacy, the deeper the emotion estimation, and the lower the familiarity, the more superficial emotion estimation and impression estimation.

  The emotion classification database 105 is a storage unit that defines different combinations of emotion types for each intimacy. As for the types of emotions, deeper and more detailed emotions are defined as the familiarity is higher, and superficial emotions or impressions are defined as the familiarity is lower.

  The response generation unit 106 performs processing for generating a response sentence for the user according to the closeness with the user calculated by the closeness calculation unit 102 and the user's emotion estimated by the emotion estimation unit 104. The response sentence is typically text data.

  The voice synthesizer 107 converts the response sentence generated by the response generator 106 into voice data. Generation of voice data from text data can be realized by various known voice synthesis techniques. After that, typically, the D / A converter converts the audio data into an analog audio signal, and the speaker outputs the analog audio signal as audio.

  Next, the operation of the response generation apparatus 100 according to the embodiment of the present invention will be described using the flowchart of FIG.

S101: Voice recognition A user speaks to the response generation apparatus 100. The voice recognition unit 101 inputs and recognizes the user's utterance content and generates text data.

  Specifically, the microphone acquires the user's utterance as an analog voice signal, and the A / D converter converts the analog voice signal into voice data. The voice recognition unit 101 receives the voice data and converts the voice data into text data using various known voice recognition techniques. For example, when the user speaks “Yes, yesterday I got a present”, the voice recognition unit 101 generates text data with the content “Yes, yesterday, I got a present”.

S102: Emotion estimation (1)
The emotion estimation unit 104 analyzes the text data generated in S101 and tries to estimate the user's emotion. Various methods are known as a method for analyzing the text data of the user's utterance content and estimating the user's emotion, and the present invention is not limited to any emotion estimation method. As one of the emotion estimation methods, there is “Emotion estimation based on the emotion-causing factor corpus acquired from the Web” (Yoshiko Tokuhisa et al., Proc. 14th Annual Conference of the Language Processing Society, March 2008).

  The emotion estimation unit 104 estimates the type of emotion of the user through the emotion estimation process. For example, from the text data “I got a gift from a friend yesterday”, a kind of emotion “happy” is estimated. In addition, various types of emotions such as “fun” and “dislike” can be estimated according to the contents of the text data.

  Note that the emotion estimation in S102 does not necessarily require precise emotion estimation using the emotion classification database 105 described later. It is only necessary to be able to estimate some emotion by the known emotion estimation method as described above. Of course, precise emotion estimation using the emotion classification database 105 described later may be used.

S103: Intimacy Calculation The intimacy calculation unit 102 calculates intimacy based on information already recorded in the intimacy database 103 and the user's emotion estimated in S102.

  Here, the familiarity database 103 cumulatively records the number of utterances of the user and the estimated number of times for each emotion type (the number of times of emotion appearance) up to now. Further, the closeness database 103 may record closeness with the user calculated in the past. When there are a plurality of users who speak, the familiarity database 103 can record the above information for each user.

The familiarity calculation unit 102 acquires the number of user utterances and the number of appearances of each emotion type from the familiarity database 103 to the present. If any emotion is estimated in S102, the number of appearances of the emotion type is updated (+1). Further, the number of utterances is updated (+1). Then, the familiarity calculating unit 102 calculates the familiarity using, for example, the following calculation formula using the updated number of utterances and number of emotion appearances.
Intimacy = number of emotion appearances / number of utterances For example, the number of utterances so far has been estimated to be 20 times, and “joyful” is estimated as 12 emotions, “fun” is 3 times, and “dislike” is estimated twice. The intimacy is (12 + 3 + 2) /20=0.85
It becomes.

  According to this calculation formula, the familiarity is represented by a numerical value of 0 to 1.0. The closer the familiarity is to 1.0, the higher the familiarity with the user.

  The intimacy calculation unit 102 records the number of utterances and the number of emotion appearances updated in step S <b> 103 in the intimacy database 103. The intimacy calculated in S103 may be recorded together.

  Instead of the processing of S102 and S103 described above, the familiarity calculation unit 102 may calculate the familiarity with the user based on the interaction with the user as in the related art described in Patent Document 1. Alternatively, the familiarity with the user may be calculated using any other known technique.

S104: Emotion estimation (2)
The emotion estimation unit 104 estimates the user's emotion again with reference to the emotion classification database 105 using the familiarity calculated by the familiarity calculation unit 102. The emotion estimation here is performed using a combination of different emotion types according to the familiarity with the user. The correspondence between the familiarity and the combination of emotion types is defined in the emotion classification database 105.

  The emotion classification database 105 holds combinations of different emotion types according to the level of intimacy. For example, the emotions used when the intimacy is high (eg, 0.8 or higher) include “happy”, “like”, “relief”, “sad”, “dislike”, “anxiety”, “spicy”, “lonely”, etc. Is set. Also, “fun”, “interesting”, “dull”, “boring”, “beautiful”, “dirty”, etc. are set as the types of emotions used when the intimacy is moderate (for example, 0.5 or more). Furthermore, “good”, “nice”, “great”, etc. are set as the types of emotions used when the intimacy is low (less than 0.5).

  Here, in the emotion classification database 105, it is preferable to set deeper and more detailed emotions as the intimacy is higher, and more superficial emotions or impressions as the intimacy is lower. That is, the higher the intimacy, the greater the number of emotion types. Moreover, the kind of emotion which shows deeper inner emotion is used, so that intimacy becomes high.

  The emotion estimation unit 104 analyzes the text data generated in S101 again, and estimates the user's emotion. Various methods are known as a method for analyzing the text data of the user's utterance content and estimating the user's emotion, and the present invention is not limited to any emotion estimation method. As one of the emotion estimation methods, there is “Emotion estimation based on the emotion-causing factor corpus acquired from the Web” (Yoshiko Tokuhisa et al., Proc. 14th Annual Conference of the Language Processing Society, March 2008).

  However, the emotion estimation unit 104 generates an emotion estimation result by selecting an appropriate one from the types of emotions defined in the emotion classification database 105. That is, the emotion estimation unit 104 first refers to the emotion classification database 105 and acquires a combination of emotion types associated with the familiarity calculated by the familiarity calculation unit 102. Next, the emotion estimation unit 104 estimates an emotion using a known emotion estimation method with the above-mentioned combination of emotion types as a candidate. That is, an appropriate emotion is selected from a combination of emotion types.

  For example, if the familiarity calculated by the familiarity calculation unit 102 is “0.85”, the emotion estimation unit 104 is first associated when the familiarity is high (0.8 or higher) in the emotion classification database 105. The combination of the types of emotions that are present, “happy”, “like”, “relief”, “sad”, “dislike”, “anxiety”, “spicy”, “lonely” are acquired. Next, emotion estimation by a known method is performed using combinations of these emotion types as estimation result candidates. As a result, the emotion estimation unit 104 can estimate, for example, “happy” as the user's emotion from the above-described combination of emotion types.

S105: Response Generation The response generation unit 106 generates a response sentence corresponding to the user's emotion estimated in S104. For example, if the user's emotion estimation result is “happy”, a response sentence “I was happy” can be created. Alternatively, if the user's emotion estimation result is “good”, a response sentence “like” can be created.

  Furthermore, for example, when the intimacy is a predetermined level or higher, a response sentence including a question answer for deepening the user's story may be generated. For example, it is possible to generate a response sentence such as “I was happy about it” or “What are you talking about?”.

  Typically, the response generation unit 106 can generate a response text by referring to a response text table held by a storage unit (not shown) and selecting an appropriate response text from the response text table. For example, the response sentence table includes response sentences associated with emotion types, and the response generation unit 106 selects response sentences previously associated with user emotion types from the response sentence table. select. For example, the response sentence “I was happy” associated with the emotion “I am happy” is selected.

S106: Speech Synthesis The speech synthesis unit 107 converts the response sentence generated in S105 into speech data using a known speech synthesis technique or the like. Thereafter, the speaker typically outputs the voice data of the response sentence that has been D / A converted as voice.

  According to the present embodiment, the response generation device 100 responds according to the familiarity and feeling with the user by changing the type of emotion and impression estimated by the familiarity. Specifically, for a partner with lower intimacy, a shallower emotion or impression is estimated and a response with a sense of distance is maintained. On the other hand, a more detailed emotion estimation is made to a partner with a high degree of intimacy, and a response is taken into the partner's emotion.

  As a result, when the intimacy is low, a response based on shallow emotion estimation is performed, and therefore, a response that goes into emotion is not performed. On the other hand, when the intimacy is high, a deeper emotion estimation makes it possible to respond more deeply to emotions. That is, since an appropriate response according to the intimacy is performed, smoother communication between the user and the response generation device 100 can be realized.

DESCRIPTION OF SYMBOLS 100 Response production | generation apparatus 101 Speech recognition part 102 Intimacy calculation part 103 Intimacy database 104 Emotion estimation part 105 Emotion classification database 106 Response generation part 107 Speech synthesis part

Claims (5)

A voice recognition unit for recognizing the user's utterance content;
An emotion estimation unit that estimates the type of emotion of the user based on the utterance content;
A response generation unit that generates a response sentence according to the estimated type of emotion,
An emotion classification database in which a closeness with the user and a combination of a plurality of emotion types are associated with each other;
The types of emotions included in the combination are different depending on the intimacy with the user,
The response estimation device, wherein the emotion estimation unit estimates an emotion of the user by selecting an emotion type corresponding to the utterance content from the combination corresponding to the familiarity with reference to an emotion classification database. The response generation device according to claim 1, wherein, in the emotion classification database, the higher the intimacy, the more the emotion type combinations include more emotion types. 3. The response generation device according to claim 1, wherein, in the emotion classification database, the higher the intimacy, the more a combination of the emotion types includes an emotion type indicating a deeper emotion. A speech recognition step for recognizing the user's utterance content;
An emotion estimation step of estimating the type of emotion of the user based on the utterance content;
A response generation step of generating a response sentence corresponding to the estimated type of emotion, comprising:
In the emotion estimation step,
An affinity classification database in which the intimacy with the user is associated with a combination of a plurality of emotion types, and the types of emotion included in the combination are different according to the difference in intimacy with the user. Browse
A response generation method of estimating the user's emotion by selecting a type of emotion according to the utterance content from the combination according to the familiarity.   The program for making a computer perform the method of Claim 4.

Images