KR20200111595A - Conversation agent system and method using emotional history - Google Patents

Abstract

The present technique relates to a conversation agent system and method using an utterance emotion history. According to a specific embodiment of the present technique, learning materials are derived by reflecting a cumulative discount reward value to induce a change in emotion for many collected original materials and a conversation model is constructed by learning based on a learning technique set for the derived learning materials. By generating a response sentence and response emotion for each of the utterance sentences and utterance emotion inputted using the constructed conversation model, converting the generated response emotion and response sentence into natural language forms, and combining the converted forms, the accuracy for emotion of a speaker can be improved as the speaker utters and an effective conversation reflecting the emotion can be conducted.

Description

Conversation agent system and method using speech emotion history {CONVERSATION AGENT SYSTEM AND METHOD USING EMOTIONAL HISTORY}

The present invention relates to a dialogue agent system and method using a speech emotion history, and more particularly, a dialogue in which the emotion of the talker is actively reflected by generating a response sentence by reflecting a discount cumulative compensation value for inducing a change in the emotion of the talker. It is about the technology that has made it possible to perform.

In the existing interactive system research, it was urgent to answer the uttered sentences without considering the user's emotions, but in recent years, research to develop an interactive system including emotions has been actively conducted.

The deep learning encoder applied to such an interactive system is a method of expressing a variable length document as a fixed length document vector using deep learning technology, and can show excellent performance in the field of emotion classification.

However, in the case of an LSTM (Long Short Term Momory) encoding device that considers the last output of the entire document sequence as a document vector, the recognition rate of the initially input pattern rapidly decreases as the input lengthens, making it suitable as an encoding device for long documents. There is a problem that has not been done.

The present invention can improve the accuracy of conversation by generating and uttering response sentences and response emotions reflecting the discount cumulative compensation value from which the speaker's emotion change is induced, and using a speech emotion history capable of performing an effective dialogue reflecting emotions. It is to provide a conversation agent system and method.

The object of the present invention is not limited to the above-mentioned object, and other objects and advantages of the present invention that are not mentioned can be understood by the following description, and will be more clearly understood by examples of the present invention. In addition, it will be easily understood that the objects and advantages of the present invention can be realized by means and combinations thereof indicated in the claims.

The conversation agent system using the speech emotion history according to an embodiment

A conversation model building device for calculating a discount cumulative reward value for inducing a change in the talker's emotion for the collected original data and constructing a dialog model by reflecting the calculated discount cumulative reward value;

A receiving device for receiving sentences and emotions included in the speech;

A preprocessing device for preprocessing the extracted sentences and emotions of the receiving device into one learning material;

A response generating device for deriving a response sentence and an emotion, respectively, based on a dialogue model built for one learning material of the preprocessor; And

It may be characterized in that it comprises an output device for converting each of the derived response sentences and emotions into a natural language form and then combining them to combine the response emotions for the uttered emotions and the response sentences for the uttered sentences for speech.

Preferably the conversation model building device,

An original data collection module for collecting a plurality of original data;

A discount accumulation compensation value calculation module for calculating a discount accumulation compensation value for each of the original data;

A learning module that generates learning materials by reflecting the calculated discount cumulative compensation value on each of the original materials, and then performs learning on the generated learning materials based on a predetermined learning algorithm; And

It may include a dialog model building module for building a dialog model based on the learning result.

Preferably the discount cumulative compensation value calculation module,

The discount calculation reward value is initially set to reward=0,

If index+2n+2< 2, it is determined whether the emotion of the episode of x[index +2n+2] is happiness, and if it is happiness, the current discount calculation compensation value = previous discount calculation compensation value + r ⁿ is set,

It may be provided to increase to n=n+1 and then repeat it for all original data.

Here, r is the discount rate, x[index+2n+2] is the episode of one original material with sentences and emotions, index is the identification information of the original material, and n is the learning material and original corrected by the discount cumulative reward value. This is the distance information from the data.

Preferably, the cumulative discount compensation value is increased based on the previous discount cumulative compensation value when the same sentiment is uttered immediately following the response of a random response sentence of the original data, and a predetermined number is a response of a random response sentence of the original data. If the sentence of the same emotion is uttered after the passing of, it may be provided to decrease based on the previous discount accumulated compensation value.

Preferably, the dialogue agent system using the speech emotion history

A model updating device that updates the conversation model by learning about the input speech sentences and emotions and the output response sentences and emotions using a policy gradient training method based on a predetermined reinforcement learning policy It may further include.

A conversation agent method using a speech emotion history according to an embodiment,

A dialogue model construction step of calculating a discount cumulative reward value for inducing a change in the talker's emotion for the collected original data and constructing a dialogue model by reflecting the calculated discount cumulative reward value;

A receiving step of receiving sentences and emotions included in the speech;

A preprocessing step of preprocessing the extracted sentences and emotions into one learning material;

A response generation step of deriving a response sentence and an emotion, respectively, based on a dialogue model built for the one learning material; And

By converting each of the derived response sentences and emotions into a natural language form and combining them to combine the response emotions for the uttered emotions and the response sentences for the uttered sentences, the dialogue agent method using the speech emotion history includes an output step. A conversation agent method using the speech emotion history is provided.

Preferably, the dialogue model building step,

Collect a number of original data,

Calculate the discount cumulative compensation value for each of the above original data,

The learning materials are generated by reflecting the calculated discount cumulative compensation value on each of the original materials, and then learning is performed on the generated learning materials based on a predetermined learning algorithm,

It may be provided to build a conversation model based on the learning result.

Preferably the discount cumulative compensation value,

Initially set to reward=0,

If index+2n+2< 2, it is determined whether the emotion of the episode of x[index +2n+2] is happiness, and if it is happiness, the current discount calculation compensation value = previous discount calculation compensation value + r ⁿ is set,

It may be provided to increase to n=n+1 and then repeat it for all original data.

Here, r is the discount rate, x[index+2n+2] is the episode of one original material with sentences and emotions, index is the identification information of the original material, and n is the learning material and original corrected by the discount cumulative reward value. This is the distance information from the data.

The cumulative discount reward value increases based on the previous discount cumulative reward value when a sentence of the same sentiment is uttered immediately following the response of a random response sentence of the original data. When the sentence of the same emotion is uttered afterwards, it may be provided to decrease based on the previous discount accumulated compensation value.

Preferably after the output step

A model update step of updating the dialogue model by learning the input speech sentences and emotions and the output response sentences and emotions using a policy gradient training according to a predetermined reinforcement learning policy It may further include.

According to an embodiment, learning materials are derived by reflecting the accumulated discount compensation value for inducing emotional change for a plurality of collected original materials, and learning is performed on the derived learning materials based on a set learning technique to rescue a dialogue model. And, by using the rescued dialogue model, a response sentence and a response emotion for each of the input speech sentence and speech emotion are generated, and the generated response emotion and response sentence are converted into natural language form, and then combined and uttered, the emotion of the speaker It is possible to improve the accuracy of the person and conduct effective conversations that reflect emotions.

The following drawings appended in the present specification illustrate preferred embodiments of the present invention, and serve to further understand the technical idea of the present invention together with the detailed description of the present invention to be described later, so the present invention is described in such drawings. It is limited only to and should not be interpreted.
1 is a system configuration diagram according to an embodiment.
2 is a detailed configuration diagram of an apparatus for constructing a conversation model of a system according to an embodiment.
3 is an exemplary diagram showing original materials and learning materials according to an embodiment.
4 is an overall flowchart illustrating an operation process of a system according to an exemplary embodiment.
5 is a detailed flowchart of a conversation model building process according to an embodiment.

The present invention is applied to an interactive system. However, the present invention is not limited thereto, and can be applied to all interactive communication systems and methods to which the technical idea of the present invention can be applied.

It should be noted that the technical terms used in the present specification are only used to describe specific embodiments, and are not intended to limit the present invention. In addition, the technical terms used in the present specification should be interpreted as generally understood by those of ordinary skill in the technical field to which the present invention belongs, unless otherwise defined in the present specification, and excessively comprehensive It should not be construed as a human meaning or an excessively reduced meaning. In addition, when a technical term used in the present specification is an incorrect technical term that does not accurately express the spirit of the present invention, it should be replaced with a technical term that can be correctly understood by those skilled in the art. In addition, general terms used in the present invention should be interpreted as defined in the dictionary or according to the context before and after, and should not be interpreted as an excessively reduced meaning.

In addition, the singular expression used in the present specification includes a plurality of expressions unless the context clearly indicates otherwise. In the present application, terms such as "consist of" or "include" should not be construed as necessarily including all of the various elements or various steps described in the specification, and some of the elements or some steps It may not be included, or it should be interpreted that it may further include additional elements or steps.

In addition, terms including ordinal numbers such as first and second used herein may be used to describe various elements, but the elements should not be limited by the terms. These terms are used only for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may be referred to as a first component.

When a component is referred to as being "connected" or "connected" to another component, it may be directly connected or connected to the other component, but another component may exist in the middle. On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that no other component exists in the middle.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings, but the same or similar components are assigned the same reference numerals regardless of the reference numerals, and redundant descriptions thereof will be omitted. In addition, in describing the present invention, when it is determined that a detailed description of a related known technology may obscure the subject matter of the present invention, a detailed description thereof will be omitted. In addition, it should be noted that the accompanying drawings are only intended to facilitate understanding of the spirit of the present invention, and should not be construed as limiting the spirit of the present invention by the accompanying drawings. The spirit of the present invention should be construed as extending to all changes, equivalents, or substitutes in addition to the accompanying drawings.

One embodiment generates and generates response words and response emotions for each of the words and emotions included in the speaker's sentence using a dialogue model built with learning data reflecting the accumulated discount reward values derived based on the original data between the talker and the talker. Each of the response emotions and response words is converted into natural language, and a combined response sentence is uttered.

1 is a diagram showing a configuration of a conversation agent system using an emotion history according to an embodiment, FIG. 2 is a diagram showing a detailed configuration of the conversation model building apparatus 100 shown in FIG. 1, and FIG. It is a diagram showing the original data and learning data collected to build the illustrated dialogue model.

1 to 3, a system according to an embodiment includes a conversation model building apparatus 100, a receiving apparatus 200, a preprocessing apparatus 300, a response generating apparatus 400, and an output apparatus 500. It may include.

The conversation model building apparatus 100 derives learning data by reflecting the accumulated discount reward value for inducing a change in the speaker's emotions with respect to the collected original data, and builds a dialog model through learning the derived learning materials. In this case, as shown in FIG. 2, the conversation model building apparatus 100 includes an original data collection module 111, a discount accumulated compensation value calculation module 112, a learning module 113, and a conversation model building module ( 114) may be included.

The original data collection module 111 may collect a plurality of original data, and the collected original data may be transferred to the discount cumulative compensation value calculation module 112. Here, as shown in Fig. 3(a), one original material is expressed as a sentence and emotion included in the spoken sentence between the interlocutors.

The discount cumulative compensation value calculation module 112 may calculate a discount cumulative compensation value based on a predetermined loss function from at least one sentence and emotion following the utterance of the collected original data.

The accumulated discount reward can be calculated by deriving the solution of the loss function with the following algorithm.

(1) Initially set as reward=0 for discount calculation reward value

(2) If index+2n+2< 2, it is determined whether the emotion of the episode of x[index +2n+2] is happiness, and if it is happiness, the current discount calculation reward value = previous discount calculation reward value + r ⁿ

(3) After increasing n=n+1, proceed to step (2)

Here, r is the discount rate and has a value less than 1, and x[index+2n+2] is a single original data as sentences and emotions. Here, if x is an odd number, it is the original data uttered by the talker 1 as an odd-numbered sentence, and if x is an even number, it is the original data uttered by the talker 2.

Also, the index is an index of the original data, and a discount cumulative reward value can be derived for any two consecutive episodes. Thus, episode x indicates the number of sentences spoken in the conversation. For example, since the values of index and n of the first speech sentence are 0, the first episode x[index+2n+2] for deriving the discount cumulative reward value is x[2]. When the emotion of happiness is included in the episode of x[2], the accumulated discount compensation value is the previous discount calculation compensation value + r ⁿ . Where n has a value of 0.

That is, n is a variable for calculating the discount cumulative reward value and represents the distance from the speech sentence to the emotion for calculating the discount cumulative reward value. For example, the discount cumulative reward value for the emotion in the first speech sentence and the first response sentence The distance to the first sentence compensated by is n = 0, the distance from the third sentence and the third reply sentence to the second sentence compensated with the cumulative reward value for the emotion is n = 1, and in the fifth sentence and the fifth reply sentence The distance n=2 to the second sentence compensated by the cumulative reward value of the discount on emotion.

In addition, if the emotion included in the immediately next sentence is happiness as a response of a random sentence of the original data, the accumulated discount reward value may be increased based on the previous discount accumulated reward value, and the response of the random sentence If the emotion included in the sentence is happiness, the accumulated discount compensation value may be increased based on the previous discount accumulated compensation value.

The learning material of FIG. 3B in which the calculated discount cumulative compensation value is reflected is transferred to the learning module 313, and the learning module 113 may perform learning on the learning material on which the discount cumulative compensation value is reflected. Here, in an embodiment, learning may be performed in various forms, such as machine learning or deep learning, but the present invention is not limited thereto.

In addition, the learning result of the learning module 113 is transmitted to the dialog model building module 114, and the dialog model building module 114 may build a dialog model based on the learning result. Such a dialog model is a recurrent neural network encoder and decoder model, and a series of processes for constructing a recurrent neural network encoder and decoder model have been previously filed by the present applicant. Accordingly, the dialogue model building module 114 derives discounted cumulative compensation for each input sentence and output sentence, multiplies the discounted cumulative compensation by a weighting factor of a predetermined loss function to derive a final loss function, and the final loss A final dialogue model can be constructed by learning the original data input by the gradient descent algorithm using the function, and the response word and response emotion for each of the speech sentences and emotions input based on the constructed final dialogue model Can be derived.

On the other hand, the receiving device 200 performs a function of receiving a sentence of the talker, and, for example, when the talker says "I am so happy", the voice of the talker is converted into a sentence and transmitted to the preprocessor 300. .

Here, in an embodiment, for convenience of explanation, the receiving device 200 describes as an example a voice converter that simply converts the talker's voice into a sentence, but extracts emotions in various forms described above and transmits them to the preprocessor 300. I can.

In addition, the receiving device 200 may extract the speaker's emotion through recognition of the speaker's facial expressions or actions, and also recognize the speaker's voice strength and low level, etc., in which the emotion included in the received word is reflected, Can be extracted.

That is, using the talker's facial expression recognition algorithm, fast signals such as facial shapes that change according to the movement of facial muscles, changes in eyes, nose, mouth, and temporary wrinkles are extracted, and the talker's emotions are derived from the extracted fast signals. . Here, emotion means surprise, fear, disgust, anger, happiness, and sadness. In other words, surprise has the shortest duration, fear is felt before taking damage, and disgust appears as an act of antipathy to something. Anger is the most dangerous emotion, caused by frustration, threats, or stimulation. On the other hand, happiness is the most positive emotion, and sadness is characterized by a long duration due to loss. The emotion extracted by these features is transmitted to the preprocessor 300.

On the other hand, the receiving device 200 extracts the talker's emotion by estimating the prosody boundary of the talker's voice using at least one of HMM (Hidden Markov Models), CART (Classification and Regression Trees), and SSL (Stacked Sequential Learning) method. The emotion of the talker is extracted based on the result of analyzing the frequency domain and the size of each emotion and transmitted to the preprocessor 300.

Hereinafter, in the present embodiment, for convenience of explanation, a speech-to-text converter is used to convert the speaker's voice into a word form, and then, the converted word and the emotion included in the word are transferred to the preprocessor 300 as an example. do.

Accordingly, the preprocessor 300 performs a function of separating the received speaker's sentence into morpheme units and then outputting a morpheme type word and an emotion included in the separated word. For example, the preprocessor 200 outputs words of "I", "too", and "I'm happy" (x ₁ to x ₄ ,) and an emotion (e_x) of "happiness", which is a user's emotion.

In addition, the emotion is derived using emotional TOBI (Tones and Breaking Indices), etc., and the derived emotion (e_x) is added to the tail of the word representing the emotion and transmitted to the response generating device 400.

For example, the response generation device 400 performs a function of generating a response sentence and a response emotion based on the above-described final dialogue model for the speech sentence and emotion processed based on the preprocessor 300. 400) outputs a response emotion for each speech emotion and a response sentence for the speech sentence. Each of these response emotions and response sentences is transmitted to the output device 500.

The output device 500 converts the received response sentiment and the response sentence into natural language forms, respectively, and combines them to generate a response sentence and utters the generated response sentence. For example, various response sentences such as "I'm happy too" and response emotions are output, which is uttered by the speaker.

In addition, the output device 500 may output a response sentence and emotion in which emotion is reflected in various forms. For example, the output unit 500 may output and/or utter a response sentence to a character such as an avatar with a facial expression and/or behavior matched with a response emotion and adjusted voice strength and low level.

Accordingly, in an embodiment, the learning material is derived by reflecting the cumulative discount compensation value for inducing emotional change for a plurality of collected original materials, and learning is performed on the derived learning material based on a set learning technique to establish a dialogue model. , Using the constructed dialogue model, a response sentence and a response emotion for each of the input speech sentence and speech emotion are generated, and the generated response emotion and response sentence are converted into natural language form, and then combined and uttered, the emotion of the speaker Accuracy can be improved, and effective conversation can be conducted that reflects emotions.

Meanwhile, the conversation agent system using the emotion history according to an embodiment may further include a model update device 600. The model update device 600, which received the sentences and emotions from the preprocessor 300, is a policy gradient training method based on a reinforcement learning policy. The conversation model of the apparatus 100 for building a conversation model may be updated by learning about emotions.

Accordingly, the conversation model building device 100 learns the sentences and emotions of the preprocessor 200 based on the policy change learning method, and updates the learning results to the established conversation model. Can increase.

In one embodiment, the emotion is described as an example of happiness, but it is applicable to a plurality of emotions, and the present invention is not limited thereto.

FIG. 4 is a flowchart illustrating an operation of a conversation agent system using an emotion history in response to an utterance, and FIG. 5 is a detailed flowchart of the conversation model building apparatus 100 of FIG. 4. A recording medium readable by a computer in which a program for executing a conversation agent method using an emotion history according to an embodiment is recorded may be provided. The program may include at least one of an application program storing an item recommendation method, a device driver, firmware, middleware, a dynamic link library (DLL), and an applet. The dialogue agent system using the emotion history includes a processor, and the processor reads a recording medium on which the dialogue agent method using the emotion history is recorded, thereby executing the dialogue agent method using the emotion history.

4 and 5, in step 10, the conversation model building apparatus 100 of the conversation agent system using the speech emotion history is a discount cumulative compensation value for inducing a change in the talker's emotion with respect to the collected original data. Can be derived.

For example, the model building apparatus 100 may collect a plurality of original data in step 10 and then calculate a discount cumulative compensation value for the collected original data.

That is, the model building apparatus 100 collects a plurality of original data in step 11, and then sets the initial discount cumulative compensation value reward to 0 for each original data in step 12, and then step 13 If index+2n+2<2 is determined in step (13), and index+2n+2<2 is satisfied in step (13), the emotion of episode x[index +2n+2] in step (14) is happy Judging recognition.

In addition, when the emotion of the episode of x[index +2n+2] in step 14 is happiness, the conversation model building apparatus 100 in step 15 includes the current discount calculation compensation value = previous discount calculation compensation value + r ⁿ Is set to and then increases to n=n+1 in step (16), and then proceeds to step (13) to determine whether index+2n+2<2.

Accordingly, learning materials are derived by reflecting the discount cumulative compensation value for the original data input in step (17), and learning is performed using a preset machine learning technique for the learning materials derived in step (18), and ), you can build a dialogue model based on the learning results.

Meanwhile, in step 20, the reception apparatus 200 of the conversation agent system using the speech emotion history may receive sentences and emotions included in the speech. At this time, the sentence may be derived based on the voice converter, and the emotion may be extracted through the emotion analyzer.

And in step 30, the preprocessor 300 may preprocess the extracted sentences and emotions into one learning material.

Meanwhile, in step 40, the response generating apparatus 400 may derive a response sentence and an emotion, respectively, based on a conversation model built for one learning material.

In step 50, the output device 500 converts each of the derived response sentences and emotions into a natural language form, and then combines them to combine and output the response emotions for the speech emotions and the response sentences for the speech sentences.

On the other hand, in step 60, the model update device 600 learns about the input speech sentence and emotion and the output response sentence and emotion by the policy gradient training according to the reinforcement learning policy. By doing so, the conversation model of the apparatus 100 for constructing a conversation model may be updated.

Conversation model is constructed by deriving learning materials by reflecting the accumulated discount compensation value to induce emotional change for a large number of collected original materials, and performing learning based on the set learning technique for the derived learning materials. By generating response sentences and response emotions for each of the input speech sentences and speech emotions, the generated response emotions and response sentences are converted into natural language form, and then combined and uttered, improving the accuracy of the speaker’s emotions. And, you can conduct effective conversations that reflect your emotions.

Although the above has been described with reference to the preferred embodiments of the present invention, those skilled in the art will be able to variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the following claims. You will understand that you can.

100: conversation model building device
111: original data collection module
112: discount accumulated compensation value calculation module
113: learning module
114: dialog model building module
200: receiving device
300: pretreatment device
400: response generating device
500: output device
600: model update device

Claims (11)

A conversation model building device for calculating a discount cumulative reward value for inducing a change in the talker's emotion for the collected original data and constructing a dialog model by reflecting the calculated discount cumulative reward value;
A receiving device for receiving sentences and emotions included in the speech;
A preprocessing device for preprocessing the extracted sentences and emotions of the receiving device into one learning material;
A response generating device for deriving a response sentence and an emotion, respectively, based on a dialogue model built for one learning material of the preprocessor; And
Conversation using speech emotion history, characterized in that it includes an output device that converts each of the derived response sentences and emotions into natural language form and combines them and combines the response emotions for the speech emotions and the response sentences for the speech emotions. Agent system. The method of claim 1, wherein the conversation model building apparatus,
An original data collection module for collecting a plurality of original data;
A discount cumulative compensation value calculating module for calculating a discount cumulative compensation value for each of the original data;
A learning module that generates learning materials by reflecting the calculated discount cumulative compensation value on each of the original materials, and then performs learning on the generated learning materials based on a predetermined learning algorithm; And
And a dialogue model building module for constructing a dialogue model based on the learning result. The method of claim 2, wherein the discount cumulative compensation value calculation module,
The discount calculation reward value is initially set to reward=0,
If index+2n+2< 2, it is determined whether the emotion of the episode of x[index +2n+2] is happiness, and if it is happiness, the current discount calculation compensation value = previous discount calculation compensation value + r ⁿ is set,
It is provided to repeat n=n+1 and then repeat for all original data,
Here, r is the discount rate, x[index+2n+2] is the episode of one original material with sentences and emotions, index is the identification information of the original material, and n is the learning material and original corrected by the discount cumulative reward value. Conversation agent system using a history of speech emotions, characterized in that distance information from data. The method of claim 1, wherein the cumulative discount compensation value increases based on a previous discount cumulative compensation value when a sentence of the same emotion is uttered immediately following a response of an arbitrary response sentence of the original data,
A conversation agent system using a speech emotion history, characterized in that it is provided to decrease based on a previous discount cumulative reward value when the same emotion sentence is uttered after a predetermined number of response sentences in response to a random response sentence of the original data. The method of claim 1, wherein the conversation agent system using the speech emotion history
A model updating device that updates the conversation model by learning about the input speech sentences and emotions and the output response sentences and emotions using a policy gradient training method based on a predetermined reinforcement learning policy Conversation agent system using a speech emotion history, characterized in that it further comprises. A dialogue model construction step of calculating a discount cumulative reward value for inducing a change in the talker's emotion for the collected original data and constructing a dialogue model by reflecting the calculated discount cumulative reward value;
A receiving step of receiving sentences and emotions included in the speech;
A preprocessing step of preprocessing the extracted sentences and emotions into one learning material;
A response generation step of deriving a response sentence and an emotion, respectively, based on a dialogue model built for the one learning material; And
A conversation agent method using a speech emotion history comprising an output step of converting each of the derived response sentences and emotions into a natural language form and then combining them to combine the response emotions for the speech emotions and the response sentences for the speech emotions. The method of claim 6, wherein the building of the dialogue model comprises:
Collect a number of original data,
Calculate the discount cumulative compensation value for each of the above original data,
The learning materials are generated by reflecting the calculated discount cumulative compensation value for each of the original materials, and then learning is performed on the generated learning materials based on a predetermined learning algorithm,
A conversation agent method using a speech emotion history, which is provided to construct a conversation model based on the learning result. The method of claim 1, wherein the discount cumulative compensation value is
Initially set to reward=0,
If index+2n+2< 2, it is determined whether the emotion of the episode of x[index +2n+2] is happiness, and if it is happiness, the current discount calculation compensation value = previous discount calculation compensation value + r ⁿ is set,
It is provided to repeat n=n+1 and then repeat for all original data
Here, r is the discount rate, x[index+2n+2] is the episode of one original material with sentences and emotions, index is the identification information of the original material, and n is the learning material and original corrected by the discount cumulative reward value. A conversation agent method using a history of speech emotions, which is distance information from data. The method of claim 8, wherein the discount cumulative compensation value is
If a sentence of the same emotion is uttered immediately following the response of a random response sentence in the original data, it increases based on the previous discount cumulative compensation value,
A conversation agent method using a speech emotion history, which is provided to decrease based on a previous discount cumulative reward value when the same emotion sentence is uttered after a predetermined number of responses to a random response sentence of the original data. The method of claim 6, after the output step
A model update step of updating the dialogue model by learning the input speech sentences and emotions and the output response sentences and emotions using a policy gradient training according to a predetermined reinforcement learning policy Conversation agent method using a speech emotion history further comprising a. A computer-readable medium for performing the conversation agent method using the speech emotion history of claim 6.

Images