JP2012093972A - Conversation processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a conversation processing device capable of creating a responsive sentence encouraging a user to change in emotion state.SOLUTION: A conversation processing device which recognizes emotion of a user and expresses emotion to the user using a knowledge database, comprises processing for obtaining a current point in a coordinate system having components of a type of emotion and strength of emotion from the recognized result of the emotion of the user, and determining a passing point nearer to a targeted convergence state of the emotion than the current point in the coordinate system. The knowledge database indicates a plurality of words and typical sentences in association with parameters in accordance with the type of the emotion and the strength of the emotion. The emotion is expressed to the user by searching the knowledge data base using the parameters in accordance with the type of the emotion and the strength of the emotion at the passing point to obtain at least one of the words and the typical sentences to be used for conversation, and creating a response sentence using the obtained result.

Description

  The present invention relates to a dialogue processing apparatus that creates a response sentence for a user's words.

Conventional dialog processing is generally one-way utterance from the device side like ATM, or answering a user's question by inserting a common fixed sentence like automatic telephone service. there were. Such a response gives a user a mechanical / office-like impression, and there is no room for enjoying a dialogue with the apparatus.
However, in recent years, the use of dialog processing devices is expected for applications such as pet toys for games and pets, and communication robots for the purpose of nursing and healing. In such an application, an interactive process that gives the user the impression that they are facing a human being is required.

  For example, there is Patent Document 1 as a communication robot. Patent Document 2 discloses a communication robot that performs communication behavior according to the personality of the user.

International Publication No. 2005/014242 JP 2008-278981 A

However, the device described in Patent Document 2 does not return a response that draws the user's emotion and prompts the emotional state to change.
An object of the present invention is to provide an interactive processing apparatus that can create a response sentence that prompts a change in a user emotional state.

  In order to achieve the above object, a dialogue processing apparatus according to the present invention is a dialogue processing apparatus that executes recognition of a user's emotion and expression of emotion for the user using a knowledge database, from the recognition result of the user's emotion A knowledge database that includes a process for obtaining a current point in a coordinate system having the emotion type and emotion strength as components, and determining a via point closer to the target emotion convergence state than the current point in this coordinate system. Indicates each of a plurality of vocabularies and fixed phrases in association with parameters according to the type of emotion and the strength of emotion, and the emotional expression for the user includes the type of emotion and the strength of emotion at the waypoint By searching the knowledge database with parameters according to the size, we obtain at least one of the vocabulary and fixed phrases that should be used in the dialogue, and the response sentence using the acquired results Characterized in that it is made by forming.

  The dialogue processing apparatus according to the present invention creates a response sentence by using an emotion type at a waypoint closer to the target emotion convergence state than the current user emotion and a vocabulary corresponding to the emotion strength. . As a result, it is possible to promote a change in user emotion that gently approaches the target emotion convergence state. Such a response that prompts a gentle emotional change can give the user the impression that they are facing humans.

The figure which shows the utilization form of the dialogue processing apparatus which concerns on this embodiment The block diagram which shows the internal structure of the dialogue processing apparatus 1 It is a figure which shows an example of an emotion vocabulary cluster file. The figure which shows the Mahalanobis table which made the vocabulary recorded in the emotion vocabulary cluster and the fixed sentence a population. The block diagram which shows the detailed structure of the emotion expression control part 13 Diagram showing the data structure of temporary learning data (A) is a relationship between a polar coordinate system in which the θ component is equally divided into 10 and assigned 10 types of emotions, and a unit vector indicating each emotion element of emotion information (joy, normality, sadness, anger) output by ST emotion (B) is a diagram showing a procedure for determining the current point of emotion in the polar coordinate system based on a synthesized vector v3 obtained by synthesizing emotion information output by ST emotion, and (c) is a diagram showing a current point, a transit point, and Diagram showing the relationship The figure which shows determination of the via point (theta) component using learning data f (theta) The figure which shows determination of the via point (theta) component using learning data fR (A) is a diagram showing conditions for continuing temporary learning even if there is a difference between the waypoint and the current point after response, and (b) is a case where there is a difference between the waypoint and the current point after answering to terminate temporary learning. That shows the conditions The flowchart which shows the operation | movement procedure of the dialogue processing apparatus 1 which concerns on this embodiment. Flow chart showing details of Mahalanobis acquisition processing Flowchart showing details of waypoint calculation / temporary learning process Flow chart showing details of emotion vocabulary narrowing processing Flow chart showing details of emotion sequence end processing The figure which shows the internal structure of the dialogue processing apparatus which concerns on a modification. The figure which shows the method of determining a present point in the dialogue processing apparatus which concerns on a modification.

DESCRIPTION OF EMBODIMENTS Hereinafter, an embodiment of a dialogue processing apparatus according to the present invention will be described with reference to the drawings.
FIG. 1 is a diagram showing a usage pattern of the dialogue processing apparatus according to the present embodiment. The dialogue processing apparatus 1 according to the present invention is a stuffed toy with a microphone, a speaker, and a microcomputer system. By repeating the operation of outputting a response sentence in voice to the user's utterance, the dialogue with the user is performed. Realize processing. In such a dialogue process, when an unpleasant emotion such as strong anger or strong sadness is recognized in the user's word, the dialogue processing device 1 uses a vocabulary that urges the user to soothe or change the user's emotion. Alternatively, a fixed sentence is used as a response sentence. The use of a vocabulary or a fixed sentence so that the user emotion converges to a target emotion state in the dialogue processing is referred to as emotion expression processing of the dialogue processing device 1. In emotion expression processing, recognition of user emotion based on the user's utterance and output of a response sentence to the user are regarded as one “trial”, and the dialogue processing device 1 targets the user's emotion by repeating the trial. When the emotion state is converged, the emotion expression process is terminated.

Next, the internal configuration of the dialogue processing apparatus 1 will be described. FIG. 2 is a block diagram showing an internal configuration of the dialogue processing apparatus 1. The dialogue processing apparatus 1 includes a microcomputer system 10, a microphone 20, and a speaker 30.
The microphone 20 generates an audio signal from the input audio, converts it into audio data in the PCM WAVE format, and outputs the audio data.

The speaker 30 converts audio data in PCM WAVE format into audio and outputs it.
The microcomputer system 10 includes a CPU, a RAM, and a ROM. When the CPU executes a program recorded in the ROM, the voice recognition unit 11, the emotion recognition unit 12, the emotion expression control unit 13, and the knowledge shown in the broken line in FIG. The functions of the database 14, the response sentence creation unit 15, and the speech synthesis unit 16 are realized.

The voice recognition unit 11 is a functional block realized by executing voice recognition software, analyzes voice data input from the microphone 20, and outputs words uttered by the user as text data.
The emotion recognition unit 12 analyzes the inflection of the voice data input from the microphone 20 and outputs emotion information as a recognition result of the emotion of the user who spoke. For the realization of the emotion recognition unit, for example, use of “ST Emotion” which is emotion recognition software manufactured by AGI Corporation can be considered. ST Emotion detects the intensity for each of the six emotional elements of “anger”, “joy”, “sadness”, “normal”, “laughter”, and “excitement”. "Anger", "joy", "sadness", "normal", and "excitement" are detected in 11 levels from 0 to 10, and "laughter" is detected in 2 levels. Hereinafter, in the present embodiment, it is assumed that ST emotion is used, and emotion information output by the emotion recognition unit 12 includes “anger”, “joy”, “sadness”, “sadness” among the detected values of “ST emotion”. The detected values of the four emotional elements “normal” are used.

  The emotion expression control unit 13 has a function of acquiring a vocabulary and a fixed sentence including emotion expressions to be used in the response sentence from the knowledge database and presenting them to the response sentence creating unit 15. Specifically, when strong anger or strong sorrow is shown in the emotion information output by the emotion recognition unit 12, vocabulary and fixed sentences that soothe or comfort them are acquired from the knowledge database 14, and the acquisition result Is output to the response sentence creation unit 15. Here, when a plurality of vocabularies and fixed phrases are acquired from the knowledge database 14, the emotion expression control unit 13 ranks the acquired vocabulary and fixed phrases according to the likelihood, and outputs them to the response sentence creation unit 15. .

  The knowledge database 14 is a vocabulary management database used for responses in dialogue processing. The knowledge database 14 collects vocabulary and fixed phrases for various conversation categories (hereinafter referred to as clusters), and creates and manages XML files for each cluster. In each cluster file, vocabulary and fixed phrases are classified by semantic attributes. Further, in the knowledge database 14, related information such as the co-occurrence frequency is added to the vocabulary and fixed phrases. FIG. 3 is a diagram illustrating an example of an emotion vocabulary cluster file. The emotional vocabulary cluster file contains vocabulary with emotional expressions and fixed phrases (hereinafter collectively referred to as “emotional vocabulary”), and conversations that it is desirable to use these emotional vocabularies in response sentences. Each emotional vocabulary is classified with the other party's emotion type and emotion strength as semantic attributes.

  The knowledge database 14 further has a function of searching for a vocabulary and a fixed sentence to be used for a response sentence from a vocabulary and a fixed sentence recorded in such a data structure. Specifically, as the search function of the knowledge database 14, a neutral vocabulary search function that returns a vocabulary and a fixed phrase that have a high probability from the Mahalanobis distance and co-occurrence frequency for the vocabulary used in the input text, and a response to user emotion And an emotion vocabulary search function that returns a suitable emotion vocabulary.

  In the emotion vocabulary search function, vocabulary selection is performed using the Mahalanobis distance with the emotion vocabulary recorded in the emotion vocabulary cluster file as a population. As the emotion vocabulary search function, when a search Mahalanobis distance is input, the knowledge database 14 returns a predetermined number (for example, four) of responses from the emotion vocabulary recorded in the emotion vocabulary cluster file from those whose Mahalanobis distance is close to the input. return.

The Mahalanobis distance used here is a covariance matrix A of a population in an orthogonal coordinate system in which the type of emotion and the strength of emotion are XY components, and a vector X representing the emotion type and strength of the search target, It is calculated by the following formula 1.
Mahalanobis distance = X * A inverse matrix * X Equation 1
Such Mahalanobis distance is uniquely determined according to the type and strength of the emotion to be searched if there is no change in the emotion vocabulary recorded in the emotion vocabulary cluster file. Therefore, a Mahalanobis table in which the Mahalanobis distance calculated from the emotion vocabulary cluster file is associated with the type and strength of the emotion is created and provided in advance to the emotion expression control unit 13 that is the inquiry source using the emotion vocabulary search function. Keep it. FIG. 4 is a diagram showing a Mahalanobis table in which the vocabulary and fixed sentences recorded in the emotion vocabulary cluster are used as a population. By using the Mahalanobis table, the emotion expression control unit 13 can easily obtain the Mahalanobis distance from the emotion type and the strength of the vocabulary to be used in the response sentence, and search the emotion vocabulary.

  The response sentence creation unit 15 creates a response sentence for the words uttered by the user, using the text data output by the speech recognition unit 11 as an input. In the response sentence creation by the response sentence creation unit 15, first, a neutral vocabulary search function provided by the knowledge database 14 is used to create a value-neutral sentence that does not involve emotional expression for the input text. Thereafter, when a vocabulary or fixed phrase including emotional expression is provided from the emotion expression control unit 13, the response sentence creation unit 15 includes a vocabulary provided according to the likelihood rank or co-occurrence frequency attached by the emotion expression control unit 13, Appropriate sentences are selected from the standard sentences, and the selected ones are inserted into the value-neutral sentences to complete the creation of the response sentence. When a vocabulary or fixed phrase including emotional expression is not provided from the emotional expression control unit 13, a value-neutral sentence is used as it is as a response sentence. The response sentence created in this way is output to the speech synthesizer 16 as response sentence text data.

The speech synthesizer 16 generates speech data in the PCM WAVE format based on the response text data output from the response text creation unit 15.
The above is the outline of the internal configuration of the dialogue processing apparatus 1.

<Details of Emotion Expression Control Unit 13>
A characteristic function of the dialogue processing apparatus 1 according to the present embodiment is emotion expression processing. Emotion expression processing is realized by inserting an emotion vocabulary into the response sentence. Therefore, the details of the emotion expression control unit 13 that determines the emotion vocabulary to be inserted into the response sentence will be described below.

FIG. 5 is a block diagram showing a detailed configuration of the emotion expression control unit 13. The emotion expression control unit 13 includes functional blocks of a coordinate conversion unit 101, a waypoint calculation unit 102, a learning data management unit 103, a vocabulary selection unit 104, and a Mahalanobis table holding unit 105.
The coordinate conversion unit 101 converts emotion information indicating the strength of four types of emotion elements into a single composite vector, thereby allowing the user to use a polar coordinate system in which the emotion type is the θ component and the emotion strength is the R component. It is a functional block that determines the current point indicating emotion.

The waypoint calculation unit 102 is a functional block that determines a waypoint. A via point is a user's emotional state that is intended to be changed by one trial, and is represented by one point in the polar coordinate system. A point closer to the convergence state of the target emotion than the current point is set as the via point in the polar coordinate system.
The learning data management unit 103 is a functional block that realizes reinforcement learning for waypoint determination by accumulating and updating learning data. The learning data includes a numerical sequence fθ {θ1, θ2,..., Θn} in which the θ component of the via point in the past episode is arranged in order from the first trial, and the R component of the via point in the past episode, It is stored in a data format of a number sequence fR {R1, R2,..., Rn} arranged in order from the first trial. Here, the “episode” is a series of trials repeated from the first trial until the user emotion converges to the target range.

  In order to update the learning data, the learning data management unit 103 accumulates the transit point data of all trials in each episode as temporary learning data in each work area as shown in FIG. Compare the number of trials with temporary learning data and the number of trials with existing learning data when it converges to the target range. If the number of trials is smaller in the temporary learning data, the learning data management unit 103 updates the temporary learning data as new learning data.

  The learning data management unit 103 includes “shame”, “scary”, “sorrow”, “厭”, “anger”, “昂”, “surprise”, “joy”, “good”, “low”. Ten θ component learning data fθ and 10 R component learning data fR starting from each region to which an emotion is assigned are accumulated. In this way, in order to manage and update the learning data for each region, the learning data management unit 103 holds the current point in the first trial as the conversation start position until the end of the episode, and the temporary learning data at the end of the episode For comparison and update, learning data in the region to which the conversation start position belongs are targeted.

  The vocabulary selection unit 104 is a functional block that provides a list of emotion vocabulary to be used in the response sentence to the response sentence creation unit 15. First, the vocabulary selection unit 104 refers to the Mahalanobis table of FIG. 4 held in the Mahalanobis table holding unit 105 to obtain the Mahalanobis distance according to the emotion type θ of the waypoint and the emotion strength R. By using the Mahalanobis distance of this waypoint to search the emotional vocabulary from the knowledge database 14, it is possible to obtain a plurality of emotional vocabularies that are close to the Mahalanobis distance. However, the emotional vocabulary obtained in this way includes words with different types of waypoints and emotions even when the Mahalanobis distance is close. Therefore, the vocabulary selection unit 104 ranks the emotional vocabulary obtained by the search and matches the emotional type with those of the emotional vocabulary, ranks them in order from the Mahalanobis distance of the transiting point, and registers them in the list. To provide.

The Mahalanobis table holding unit 105 is a functional block that holds the Mahalanobis table, and is realized by a recording area secured in a rewritable nonvolatile memory such as FeRAM.
The detailed configuration of the emotion expression control unit 13 has been described above. Next, details of the current point determination method in the coordinate conversion unit 101 and the waypoint determination method in the waypoint calculation unit 102 will be described.

<Determination Method of Current Point in Coordinate Conversion Unit 101>
The current point is determined by converting emotion information indicating the strength of four types of emotion elements into one composite vector, thereby determining the emotion type as one component in the polar coordinate system with the θ component and emotion strength as the R component. Is done.

  In the polar coordinate system used to indicate the current point of user emotion, as shown in FIG. 7 (a), there are 10 regions divided every 36 degrees of the θ component in the “Emotion Expression Dictionary” (1993 Akira Nakamura Publishing). The ten emotions that are classified are "shame", "fear", "sorrow", "厭", "anger", "昂", "surprise", "joy", "good", "low" Assigned counterclockwise. Hereinafter, the counterclockwise direction is the positive direction of θ, and the clockwise direction is the negative direction of θ.

  In order to convert emotion information into one synthetic vector on such a polar coordinate system, four unit vectors corresponding to emotion elements of emotion information are defined. Specifically, a unit vector corresponding to the emotional element of “normal” in the emotional information is defined in the center direction of the “cheap” area from the origin, and the emotional information in the central direction of the “angry” area from the origin. Define a unit vector corresponding to the emotion element of “anger”, define a unit vector corresponding to the emotion element of “joy” in the emotion information from the origin toward the center of the area of “joy”, and define “sorrow” from the origin. A unit vector corresponding to the emotion element of “sorrow” in the emotion information is defined in the central direction of the region. Based on the unit vector thus defined, the four emotion elements indicated by the emotion information can be expressed by four vectors according to their strengths as shown in FIG.

These four and vector V _p of "joy" in the vector obtained by combining the vector V _n of the "normal" and the vector V _1, also the vector V _a of "anger" and the vector V _s of "sorrow" Is a vector V ₂ and a vector V ₁ and a vector V ₂ are combined to obtain a combined vector V ₃ indicating the current point.
As a specific calculation method of the current point, if the declination angle θ1 in the “joy” direction, the declination angle θ2 in the “normal” direction, the declination angle θ3 in the “sorrow” direction, and the declination angle θ4 in the “anger” direction, The x and y components of the unit vector are
Joy (x) = cos θ1 Joy (y) = sin θ1
Normal (x) = cos θ2 Normal (y) = sin θ2
Sorrow (x) = cos θ3 Sorrow (y) = sin θ3
Anger (x) = cos θ4 Joy (y) = sin θ4
Therefore, by multiplying each component by the value of the intensity (0 to 10) of each emotion element indicated by the emotion information, “X” that is the x component of the composite vector, and y of the composite vector The component “Y” is determined. From the x and y components of the current point, the deviation angle θt of the combined vector, that is, the θ component of the current point, is obtained by the following equation 2.

θt = tan ⁻¹ (Y / X) Equation 2
The square root of the combined vector, (X ² + Y ² ), is the R component of the current point. The details of the current point determination method in the coordinate conversion unit 101 have been described above.

<Method for Determining Via Point in Via Point Calculation Unit 102>
Next, details of the waypoint determination method will be described. In the dialogue processing apparatus 1 according to the present embodiment, the user component is in a state where the θ component is an area to which any of the emotions of low, good, and joy is assigned, or the R component indicating the strength of the emotion is 1 or less. As shown in FIG. 7C, the via point is closer to the target emotion convergence state than the current point in the polar coordinate system, that is, it is cheaper and better than the current point. , It is set to a point that is close to the region of pleasure or has a smaller R component than the current point. In the following, the region where the θ component is assigned to one of the emotions of low, good and happy and the region where the R component indicating the strength of the emotion is 1 or less are indicated by hatching in FIG. The area is called “target range”.

  There are two ways to determine the waypoint. The first is a waypoint determination method in the initial state, and the second is a waypoint determination method using learning data. If the user emotion has previously converged to the target range by the interactive processing that repeats the acquisition of the user's speech and the output of the response sentence, the learning data management unit 103 causes the user emotion to converge from the start of the interactive processing. A series of processes up to this time is recorded as learning data. The initial state is a state in which such learning data is not accumulated.

<Via-point determination method in the initial state>
In the determination of the waypoint in the initial state, the θ component and the R component are determined independently. The range in which the θ component is changed from the current point by determining one elapsed point is 0 to 36 degrees, and the amount of change is determined for each trial using a probability density function such as a normal distribution. The direction in which the θ component is changed is when the current point is on the “low” side from the center of the area assigned with “sorrow”, “fear”, or “shame” and the area assigned with “厭”. , The direction approaching the area assigned “low”, that is, the negative direction. On the other hand, if the current point is on the “joy” side of the area to which any of “surprise”, “昂”, or “anger” is assigned and the center of the area to which “厭” is assigned, “joy” is assigned. It changes to the direction approaching the given area, that is, the positive direction.

The range in which the R component is decreased from the current point by determining the elapsed point once is 0 to 1, and the amount of change is determined for each trial using a probability density function such as a normal distribution.
<Via-point determination method using learning data>
Also in the waypoint determination using the learning data, the θ component and the R component are determined independently. This is the same as the waypoint determination in the initial state, and the learning data is used for determining the θ component or the R component.

  The learning data is a series of fθ {θ1, θ2,..., Θn}, fR {R1, R2,. Rn}. The learning data management unit 103 has emotions of “shame”, “scary”, “sorrow”, “厭”, “anger”, “昂”, “surprise”, “joy”, “good”, “low”. Ten θ component learning data fθ and 10 R component learning data fR starting from each assigned region are accumulated.

  FIG. 8 is a diagram illustrating determination of the via point θ component using the learning data fθ. In this figure, the arrow indicates the θ component of the current point, and the white circle is the t-th value of the sequence fθ. In the waypoint calculation unit 102, the trial in which the current point outside the target range is first input is selected as the first trial, and the sequence fθ starting from the region to which the current point belongs is selected, and the first of the sequence fθ is selected. Is the θ component of the waypoint. In the subsequent second, third, and fourth trials, the same number fθ is used, and the second, third, and fourth values are set as θ components of the waypoints. Similarly, when determining the via-point R component using the learning data fR, the numerical sequence fR starting from the region to which the current point in the first trial belongs is selected, and thereafter, as shown in FIG. Select a sequence fR starting from the region to which the current point θ component belongs, and try the first, second, third, and fourth values of the first, second, third, and fourth values of the sequence fR, respectively. This is used for the R component at the via point.

  Here, when the difference of the θ component exceeds a predetermined threshold (for example, 36 degrees) between the transit point in the t-th trial and the current point recognized in the t + 1-th trial, the difference in the R component is a predetermined value. When the threshold value (for example, 2.5) is exceeded, the t-th trial is referred to as an “invalid trial”. In the example of FIG. 10A, the fourth trial is an invalid trial. In the case of the trial next to the invalidation trial, in the example of FIG. 10A, if the fifth trial is not the invalid trial, the waypoint determination is continued using the previous learning data. However, when invalid trials are continued twice as shown in FIG. 10B, in the trial next to the second invalid trial, new learning data having a start position in the region to which the current point belongs in the trial is newly added. Re-select and make waypoint determination as a new first trial. Such determination of invalid trial and reselection of learning data are performed independently for the θ component and the R component.

The details of the waypoint determination method in the coordinate conversion unit 101 have been described above.
By the function of the emotion expression control unit 13 described so far, it is possible to select an emotion vocabulary that brings the user emotion closer to the target range, and by inserting such an emotion vocabulary into the response sentence, according to the present embodiment Emotion expression processing by the dialogue processing device 1 is realized.

<Operation of Dialogue Processing Device 1>
Subsequently, the operation of the dialogue processing apparatus 1 according to the present embodiment will be described.

FIG. 11 is a flowchart showing an operation procedure of the dialogue processing apparatus 1 according to the present embodiment.
In the operation of the dialog processing device 1, initialization processing of the dialog processing device 1 is executed in step S1, and the user's speech is awaited. The initialization process is a process in which the learning data management unit 103 secures a recording area for managing temporary learning data and a conversation start position in a work area on the RAM. A temporary learning data recording area and a conversation starting position recording area are separately provided for the θ component and the R component.

After the initialization, when the user speaks, voice input data is acquired through the microphone 20 (step S2). This voice input data is analyzed by the emotion recognition unit 12, and emotion information indicating the strength of the four emotion elements is acquired. (Step S3). Based on this emotion information, the emotion expression control unit 13 executes a Mahalanobis acquisition process (step S4).
Here, when the current point of user emotion obtained by converting emotion information is within the target range, the Mahalanobis distance is not acquired in the Mahalanobis acquisition process described in detail later (step S5: No). In such a case, no emotional vocabulary is provided to the response sentence creating unit 15, and the response sentence creating unit 15 creates a response sentence in step S8 without inserting the emotional vocabulary. For example, when the user's remark is the word “T team has lost” and a strong emotion whose current point of emotion is out of the target range is not recognized, the response sentence creation unit 15 reads “ Create a response sentence that does not include the emotional vocabulary.

  On the other hand, if the current point of user emotion obtained by converting emotion information is outside the target range, some Mahalanobis distance is acquired in the Mahalanobis acquisition process (step S5: Yes). In such a case, the knowledge database 14 is searched using the acquired Mahalanobis distance by the vocabulary selection unit 104 in the emotion expression control unit 13, and a plurality of emotional vocabularies having a short Mahalanobis distance are acquired (step S6). The vocabulary selection unit 104 that has acquired a plurality of emotional vocabularies excludes inappropriate emotional vocabulary by the emotional vocabulary narrowing process, which will be described in detail later, and creates an output list in which only appropriate emotional vocabularies are arranged in order of closest Mahalanobis distance, It outputs to the response sentence preparation part 15 (step S7). The response sentence creation unit 15 provided with the emotion vocabulary list creates a response sentence in which the emotion vocabulary is inserted in step S8.

  As an example, strong emotions that deviate from the target range due to the inflection of the user's word “T team lost” were recognized, and the Mahalanobis distance of the waypoint determined from this recognition result was 2.2 Assume a case. Referring to FIG. 4, the Mahalanobis distance is close to 2.2 because the emotional vocabulary of the emotion type “sorrow” strength “4” with the Mahalanobis distance 2.3, and the emotion type “sorrow” with the Mahalanobis distance of 2.0. "Emotion vocabulary with strength" 5 ", emotion vocabulary with emotion type" anger "intensity" 4 "with Mahalanobis distance 2.4, and emotion with" anger "intensity" 5 "with emotion type" Maharanobis distance 2.2 " Vocabulary. In this example, the emotional vocabulary of these attributes is extracted from the emotional vocabulary cluster file shown in FIG. 3 and acquired by the vocabulary selection unit 104. If the current point of the user emotion is included in the “sorrow” region, the emotion vocabulary of the emotion type “anger” is excluded from the extracted vocabulary, and the emotion type “Mahalanobis distance is 2.3”. Vocabulary of “sorrow” strength “4” “I'm sorry,” “I do n’t care”, “Maharanobis distance 2.0 emotion type“ sorrow ”strength“ 5 ”vocabulary“ sad, ”“ still quite Is registered in the output list in this order, and is output to the response sentence creation unit 15. The response sentence creation unit 15 selects a vocabulary to be used from the emotion vocabulary registered in the output list using the Mahalanobis distance and the co-occurrence frequency. When the emotional vocabulary “I'm sorry,” is selected here, a response sentence is created in which the emotional vocabulary “I'm sorry, but I'm sorry.” Is inserted. The response sentence created in this way is provided to the user as sound through the speaker 30.

In the above operation procedure, steps S2 to S8 are one trial and are repeated every time the user speaks.
<Mahalanobis acquisition processing>
Next, details of the Mahalanobis acquisition process will be described. FIG. 12 is a flowchart showing details of the Mahalanobis acquisition process.

In the Mahalanobis acquisition process, first, the coordinate conversion unit 101 converts emotion information into a composite vector, and determines the current point of user emotion as shown in FIG. 4B (step S11). Next, the waypoint calculation unit 102 executes a waypoint calculation / temporary learning process described later in detail, and determines a waypoint based on the current point (step S12).
Here, when the current point is outside the target range (step S13: No), the vocabulary selection unit 104 refers to the Mahalanobis table of FIG. 4 (step S14), and the θ component of the waypoint determined in step S11, R A Mahalanobis value is acquired based on the component (step S15).

When the current point is included in the target range (step S13: Yes), the vocabulary selection unit 104 notifies the response sentence creation unit 15 that the Mahalanobis value has not been acquired (step S16), and then learning data management The emotion sequence end process is executed by the unit 103.
The above is the details of the Mahalanobis acquisition process.
<Via-point calculation / temporary learning process>
Next, details of the waypoint calculation / temporary learning process will be described. FIG. 13 is a flowchart showing details of the waypoint calculation / temporary learning process.

  In the waypoint calculation / temporary learning process, first, in step S21 and step S22, the learning data management unit 103 confirms the number of accumulations of temporary learning data. If the accumulated number of temporary learning data exceeds a predetermined threshold Th1 (for example, 10 times) (step S21: Yes), the temporary learning data is reset (step S35). The temporary learning data reset is a process of erasing all data recorded in the temporary learning data recording area and the recording area of the conversation start position.

When no transit point is recorded as temporary learning data (step S22: Yes), that is, immediately after the temporary learning data is reset or at the first trial, the learning data management unit 103 records the current point as the conversation start position. The area is recorded (step S23), and the process proceeds to step S30.
If a transit point is recorded as temporary learning data (step S22: No), that is, if it is a second or later trial, the transit point calculation unit 102 invalidates the previous trial in the processing procedure of step S24 to step S29. It is determined whether it was a trial. When the difference between the R component of the previous trial via point and the current point is greater than 2.5 (step S24: Yes), or when the difference between the previous trial via component and the current point is greater than 36 degrees ( In step S27: Yes, assuming that the previous trial was an invalid trial, the variable for recording the out-of-range count for each component is incremented (step S26, step S29). When the previous trial is not an invalid trial (step S24: No, step S27: No), the variable for recording the out-of-range frequency for each component is initialized to 0 (step S25, step S28). Further, in step S30, the waypoint calculation unit 102 determines whether the variable for recording the out-of-range number is 2 or more. If the out-of-range count is 2 or more, the temporary learning data is reset (step S36). Note that the variable for recording the out-of-range count is initialized to 0 in the temporary learning data reset process.

  Thereafter, when the learning data management unit 103 additionally records the current point in the temporary learning data recording area and the current point is located outside the target range (step S32: No), the waypoint calculation unit 102 performs steps S33 and S33. In S34, the R component and the θ component are determined as the waypoints for the current trial. In the determination of the waypoint in step S33 and step S34, the waypoint determination method in the initial state already described as “the waypoint determination method in the waypoint calculation unit 102” or the waypoint determination method using learning data Is used.

The above is the details of the waypoint calculation / temporary learning process.
<Emotion vocabulary narrowing process>
Next, details of the emotion vocabulary narrowing process will be described. FIG. 14 is a flowchart showing details of emotion vocabulary narrowing processing.
The emotion vocabulary narrowing process is a process executed by the vocabulary selection unit 104. First, the emotion type assigned to the region to which the waypoint belongs is acquired based on the θ component of the waypoint (step S41). Is set to a semantic attribute character string (step S42). Thereafter, the output table recording area on the RAM is initialized (step S43).

  After the output table recording area is initialized, the processes in steps S45 to S46 are repeated for all emotional vocabularies detected from the knowledge database 14. The processing in step S45 selects the unprocessed emotion vocabulary detected from the knowledge database 14 and selects the one having the closest waypoint and Mahalanobis distance (step S45), and the emotion type that is the semantic attribute of this emotion vocabulary And the semantic attribute character string set in step S42 are compared (step S46). If the comparison results match (step S46: Yes), the selected emotion vocabulary is additionally recorded in the output table recording area (step S47).

The above is the details of the emotion vocabulary narrowing process.
<Emotion sequence end processing>
Next, the details of the emotion sequence end process will be described. FIG. 15 is a flowchart showing details of emotion sequence end processing.
The emotion sequence end process is a process executed when the current point indicating the user emotion converges to the target range by repeating trials.

  In the emotion sequence end process, the learning data management unit 103 acquires the value of the θ component recorded in the recording area of the conversation start position (step S51), and this value is “shame”, “fear”, “sorrow”, It is determined whether any kind of emotions “厭”, “anger”, “昂”, “surprise”, “joy”, “good”, “low” are included in the assigned region (step S52), When the learning data having this area as the start position has not been accumulated before (step S53: No), the waypoint recorded in the temporary learning data recording area is accumulated as new learning data (step S54).

  When the learning data having the area determined in step S52 as the start position has been accumulated before (step S53: Yes), the number of via points of the learning data and the temporary learning data recording area are recorded. The number of via points is compared (step S55). If the number of waypoints recorded in the temporary learning data recording area is smaller (step S55: Yes), the existing learning data is discarded, and the waypoints recorded in the temporary learning data recording area are replaced with new ones. Accumulate as learning data.

Finally, temporary learning data is reset in step S56. The above is the details of the emotion sequence end process.
As described above, according to the dialogue processing apparatus according to the present embodiment, using the vocabulary corresponding to the emotion type at the waypoint closer to the convergence state of the target emotion than the current user emotion, and the strength of the emotion, Response sentences can be created. As a result, it is possible to promote a change in user emotion that gently approaches the target emotion convergence state. Such a response that prompts a gentle emotional change can give the user the impression that they are facing humans.

(Modification)
As mentioned above, although the dialogue processing apparatus concerning the present invention was explained based on the embodiment, the present invention is not limited to these embodiments. For example, the following modifications can be considered.
(1) The present invention may be a recording method disclosed by the processing procedure of the flowchart described in each embodiment. Further, the present invention may be a program that realizes these methods by a computer, or may be a digital signal composed of the program.

The present invention also provides a computer-readable recording medium for the program or the digital signal, such as a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, BD (Blu-ray Disc). It may be recorded in a semiconductor memory or the like.
In the present invention, the program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication line, a network represented by the Internet, or the like.

In addition, the program or the digital signal is recorded on the recording medium and transferred, or the program or the digital signal is transferred via the network or the like, and is executed by another independent computer system. It is good.
(2) The via point may be a point closer to the target emotion convergence state than the current user emotion, and may be determined by a method different from the above embodiment.

For example, first, the θ component is changed from the current point, and the R component is determined as a via point set to the same value as the current point. As a result of repeating the trial using such a via point, invalid trials are performed a predetermined number of times (for example, In the case of two consecutive times, a method may be used in which the R component is decreased from the current point by the next trial and the via point is set to the same value as the current point for the θ component.
(3) In the above embodiment, the learning data is held in the data format of the sequence fθ {θ1, θ2,..., Θn} and the sequence fR {R1, R2,. May be in other data formats.

For example, plotting via points in past episodes on Cartesian coordinates with the horizontal axis representing the number of trials t and the vertical axis representing θ or R, and obtaining an approximate line θ = at + b or R = at + b from these by the least square method Can do. Using this f (t) = at + b as learning data, the tth θ and R may be determined.
As another example, the inclination a of the above approximate line may be used as learning data. In this case, a function point f (t) = at + θ1 and f (t) = at + R1 with the current point of the first trial as an intercept can be generated to determine the via point. Here, θ1 is the θ component of the current point in the first trial, and R1 is the R component of the current point in the first trial.

  (4) A knowledge database may be provided outside the dialogue processing apparatus. For example, a server device connected to the Internet may function as a knowledge database, and a communication interface may be provided in the dialog processing device as shown in FIG. 16 to communicate with the knowledge database via a network such as a wireless LAN. In such a modification, it becomes easy to update the collected vocabulary of the knowledge database in a timely manner via the Internet.

(5) In the above embodiment, the current point is obtained by combining the output of ST emotion on the polar coordinates. However, the method of expressing the current point of user emotion is not necessarily limited to polar coordinates.
For example, by adding a z-axis corresponding to the value of “excitement” among the output values of ST emotion to the polar coordinates described in the above embodiment, the current point of user emotion is expressed in spatial coordinates instead of plane coordinates You can also

  (6) Each embodiment and modification may be combined.

  INDUSTRIAL APPLICABILITY The present invention is useful for toys, robot devices, and the like that interact with a user.

DESCRIPTION OF SYMBOLS 1 Dialogue processing apparatus 10 Microcomputer system 11 Voice recognition part 12 Emotion recognition part 13 Emotion expression control part 14 Knowledge database 15 Response sentence creation part 16 Speech synthesis part 20 Microphone 30 Speaker 101 Coordinate conversion part 102 Via-point calculation part 103 Learning data management part 104 Vocabulary selector 105 Mahalanobis table holder

Claims (7)

A dialogue processing device that executes recognition of a user's emotion and emotion expression for the user using a knowledge database,
From the recognition result of the user emotion, obtain the current point in the coordinate system with the emotion type and emotion strength as components, and determine the via point closer to the target emotion convergence state than the current point in this coordinate system Including processing,
The knowledge database shows each of a plurality of vocabularies and fixed phrases in association with parameters according to emotion type and emotion strength,
The emotional expression for the user is obtained by obtaining at least one of a vocabulary and a fixed sentence to be used for dialogue by searching a knowledge database with a parameter according to the type of emotion at the waypoint and the strength of the emotion. An interactive processing device characterized in that it is made by creating a response sentence using. The coordinate system is a polar coordinate system in which an emotion type is assigned to each region of the θ component, and the strength of the emotion is an R component,
The parameter is a Mahalanobis distance using a vocabulary and a fixed sentence recorded in a knowledge database as a population in an orthogonal coordinate system in which an emotion type and an emotion strength are XY components. 1. The dialogue processing apparatus according to 1. The types of emotion indicated by the θ component of the polar coordinate system are the emotions of cheap, good, happy, surprise, jealousy, anger, jealousy, sadness, fear and shame in the polar coordinate system in a clockwise direction. The dialogue processing apparatus according to claim 2. The waypoint is determined
A first process for setting a value closer to the region to which the emotion type of cheap, good, and joy is assigned than the θ component at the current point as the θ component at the via point, and the strength of emotion than the R component at the current point 4. The dialogue processing apparatus according to claim 3, wherein the dialogue processing apparatus is configured to execute at least one of the second processing for setting a value obtained by weakening the value as the R component of the waypoint. The dialogue processing device has a learning function that reinforces and learns a waypoint determination method by presenting the created response sentence to the user and repeating an attempt to newly recognize the user emotion after the presentation. The dialogue processing apparatus according to claim 2. The learning function
The difference between the θ component of the via point and the θ component of the new current point related to the newly recognized user emotion, and the difference between the R component of the via point and the R component of the new current point are within the specified value. In some cases, the waypoints are stored as temporary learning data,
If the number of via points stored as temporary learning data is less than the existing learning result when the recognized user emotion becomes the target emotion convergence state, the temporary learning data becomes a new learning result. The interactive processing device according to claim 5, wherein the interactive processing device is set. The learning result is stored as a function θ = f (t) and a function R = f (t) that can be derived from the number of trials t,
The dialog processing apparatus according to claim 6, wherein the determination of the waypoint is performed by determining θ and R based on these functions.

Images