KR102586289B1 - Conversational agent system and method using predict conversation prediction - Google Patents

Abstract

This technology discloses a conversation agent system and method using conversation prediction. A specific example of implementation of this technology is to reflect the cumulative reward value of each of the multiple response candidates for each of the multiple query candidates in the future dialog step based on the current dialog step, thereby generating the multiple response candidates in the current dialog step. By deriving one response candidate and uttering the derived response candidate, a natural conversation can be continued in real time on the open domain, and by deleting a malicious response candidate among multiple response candidates, malicious conversation can be prevented in advance.

Description

Conversation agent system and method using conversation prediction {CONVERSATIONAL AGENT SYSTEM AND METHOD USING PREDICT CONVERSATION PREDICTION}

The present invention relates to a conversation agent system and method using conversation prediction. More specifically, the present invention relates to a conversation agent system and method using conversation prediction. More specifically, the present invention relates to a user's query in the current conversation step by reflecting the cumulative reward value of multiple response candidates for the query of the next conversation step in an open domain. It is about technology that allows conversations to proceed naturally in real time as responses are uttered.

Existing research on interactive systems focused on providing answers to sentences uttered without considering the user's emotions, but recently, research on developing interactive systems that include emotions is actively underway.

The deep learning encoder applied to these interactive systems uses deep learning technology to express variable-length documents as fixed-length document vectors, and can show excellent performance in the field of emotion classification.

However, in the case of an LSTM (Long Short Term Memory) encoding device that considers the last output of the entire document sequence as a sentence vector, the recognition rate of the initially input pattern decreases rapidly as the input becomes longer, making it unsuitable as an encoding device for long sentences. There is a problem that was not done.

1. Korean Patent Publication No. 2013-013557 (Chain conversation pattern-based conversation system and method)

2. Jang, Youngsoo, Jongmin Lee, and Kee-Eung Kim. “Bayes-Adaptive Monte-Carlo Planning and Learning for Goal-Oriented Dialogues.” AAAI. 2020.

Therefore, the present invention estimates the conversation of the next conversation step based on the cumulative reward value of a plurality of response candidates to the inquiry of the next conversation step, and reflects the query of the estimated next conversation step to answer the user's query in the current conversation step. The goal is to provide a conversation agent system and method using conversation prediction that can naturally continue a conversation in real time on an open domain as a response is uttered.

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

A conversation agent system using conversation prediction according to an embodiment of the present invention,

A conversation learning model building unit that collects a plurality of conversation data and previous conversation data and builds a conversation learning model to estimate a number of response candidates for the query of the input current conversation step;

If the number of dialogue steps for a query in the current dialogue step entered reaches the predetermined maximum number of dialogue steps, or the number of searches for a query in the current dialogue step entered is not greater than the estimated number of response candidates, a search response candidate deriving unit that derives one response candidate to be searched among the estimated plurality of response candidates; and

Regarding the search response candidate, based on the current conversation step, one response candidate among the plurality of response candidates of the current conversation step is derived by reflecting the cumulative reward value of the plurality of response candidates for each query candidate in the next conversation step. One feature is that it includes a response speech unit that utters the derived response candidate.

Preferably the response speech part is:

a current compensation value calculation module that calculates the compensation value of the search response candidate and outputs a current compensation value;

a future cumulative reward value calculation module for deriving a cumulative reward value of a plurality of response candidates for a plurality of query candidates in a next conversation step based on the current conversation step for the search response candidate;

a search response candidate reward value calculation module that derives a reward value of the search response candidate from the sum of the future cumulative reward value and the current reward value; and

An average reward value derivation module that derives the average of search response candidate reward values for a plurality of response candidates; and

It may include an utterance module that derives one response among the plurality of response candidates by averaging the compensation values of all search response candidates for the plurality of response candidates and utters the derived single response.

Preferably, the current compensation value calculation module is:

If the calculated current compensation value is smaller than the predetermined minimum allowable compensation value, the search response candidate may be determined to be a malicious response candidate and may further include a filter module for deleting the search response candidate from the plurality of response candidates.

Preferably, the system:

If the number of dialogue steps for the query of the input current dialogue step does not reach the predetermined maximum number of dialogue steps and the number of searches for the query of the input current dialogue step is greater than the estimated number of response candidates, a new response candidate It may further include an additional response candidate generator that additionally estimates .

Preferably, the cumulative compensation value is

It can be equipped with a Generalized Advantage Estimation (GAE) model that optimizes the cumulative reward value using a policy gradient technique.

According to another embodiment of the present invention, a conversation agent method using conversation prediction,

A conversation learning model building step of collecting a plurality of conversation data and previous conversation data to build a conversation learning model that estimates a number of response candidates for the query of the current conversation step input;

If the number of dialogue steps for a query in the current dialogue step entered reaches the predetermined maximum number of dialogue steps, or the number of searches for a query in the current dialogue step entered is not greater than the estimated number of response candidates, the number of search steps in the current dialogue step is A search response candidate derivation step of deriving one response candidate to be searched from among the estimated plurality of response candidates; and

Regarding the search response candidate, based on the current conversation step, one response candidate among the plurality of response candidates of the current conversation step is derived by reflecting the cumulative reward value of the plurality of response candidates for each query candidate in the next conversation step. and a response utterance step of uttering the derived response candidate.

Preferably, the response speech step is,

calculating a compensation value of the search response candidate and outputting a current compensation value;

Deriving a cumulative reward value of a plurality of response candidates for a plurality of query candidates in a next dialogue step based on the current dialogue step for the search response candidate;

Deriving a reward value of the search response candidate from the sum of the future accumulated reward value and the current reward value;

Deriving an average of search response candidate reward values for a plurality of response candidates; and

It may include deriving one response among the plurality of response candidates by averaging the reward values of all search response candidates for the plurality of response candidates and uttering the derived single response.

Preferably, the step of calculating the current compensation value is,

If the calculated current compensation value is less than a predetermined minimum allowable compensation value, the step of determining the search response candidate as a malicious response candidate and deleting the search response candidate from the plurality of response candidates may be further included.

Preferably, the method:

If the number of dialogue steps for the query of the input current dialogue step does not reach the predetermined maximum number of dialogue steps and the number of searches for the query of the input current dialogue step is greater than the estimated number of response candidates, a new response candidate It may further include an additional response candidate generation step of additionally estimating .

According to one embodiment, based on the current dialogue step, one of the plurality of response candidates in the current dialogue step is reflected by reflecting the cumulative reward value of each of the plurality of response candidates for each of the plurality of query candidates in the future dialogue step. By deriving response candidates and uttering the derived response candidates, a natural conversation can be continued in real time on the open domain.

In addition, according to one embodiment, by deleting a malicious response candidate from a plurality of response candidates, there is an effect of preventing emotional damage caused by a malicious conversation in advance.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and together with the detailed description of the invention described later, serve to further understand the technical idea of the present invention. Therefore, the present invention includes the matters described in such drawings. It should not be interpreted as limited to only .
1 is a configuration diagram of a conversation agent system using conversation prediction according to an embodiment.
Figure 2 is a detailed configuration diagram of the response speech unit of the system of one embodiment.
Figure 3 is an overall flow diagram of a conversation agent process in another embodiment.

Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be modified and implemented in various forms. Accordingly, the embodiments are not limited to the specific disclosed form, and the scope of the present specification includes changes, equivalents, or substitutes included in the technical spirit.

Terms such as first or second may be used to describe various components, but these terms should be interpreted only for the purpose of distinguishing one component from another component. For example, a first component may be named a second component, and similarly, the second component may also be named a first component.

When a component is referred to as being “connected” to another component, it should be understood that it may be directly connected or connected to the other component, but that other components may exist in between.

Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate the presence of the described features, numbers, steps, operations, components, parts, or combinations thereof, but are not intended to indicate the presence of one or more other features or numbers, It should be understood that this does not exclude in advance the possibility of the presence or addition of steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the relevant technical field. Terms as defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings they have in the context of the related technology, and unless clearly defined in this specification, should not be interpreted in an idealized or overly formal sense. No.

Hereinafter, the present invention will be described in detail by explaining preferred embodiments of the present invention with reference to the attached drawings.

Figure 1 is a configuration diagram of a conversation agent system using conversation prediction according to an embodiment, and Figure 2 is a detailed configuration diagram of the future conversation estimation unit of Figure 1.

Referring to Figures 1 and 2, a conversation agent system using conversation prediction according to an embodiment accumulates reward values for each of a plurality of response candidates for each of a plurality of query candidates in a future conversation step based on the current conversation step. It is equipped to reflect the reward value to derive one response candidate among the plurality of response candidates in the current conversation step and utter the derived response candidate. Accordingly, the system includes a conversation learning model construction unit 100 and a search response candidate derivation unit. It may include at least one of (200), a response speech unit 300, and an additional response candidate generation unit 400.

Here, the conversation learning model building unit 100 can collect a large number of conversation data such as conversation data of movie and drama subtitles, social media, etc., and collect a number of previous conversation data between the user and the conversation agent system and collect the collected conversations. Save the data in the form of learning data.

In response to the user's inquiry in the current input dialogue step, the system performs learning based on a pre-built dialogue learning model to derive a plurality of response candidates, and the derived plurality of response candidates are delivered to the search response candidate derivation unit 200. do. _Multiple response candidates are derived from the probability of a response candidate

The search response candidate derivation unit 200 determines whether the depth of the number of conversation steps for the user's query in the inputted current conversation step reaches the predetermined maximum number of conversation steps Max_Depth or the number of searches for the query in the current input conversation step N ( If ht) is not greater than the number children(ht) of the estimated multiple response candidates, the search response candidate xt is derived from the multiple response candidates estimated in the current conversation step. For example, a conversation step is defined as one response sentence to one query sentence in a conversation between a user and the system, and the conversation between the user and the system progresses through multiple conversation steps. Here, the maximum conversation step is predetermined.

Here, the search response candidate x _t can be expressed as the following equation 1.

[Equation 1]

Here, Q(h _t , x) is the reward value for search response candidate x among the multiple response candidates estimated for query h _t at the current stage , and N(h _t ) is the multiple response candidate for query h _t This is the number of searches among response candidates, and N(h _t , x) is the number of searches among multiple response candidates at the current conversation step. Here, the number of searches is the number of response candidates uttered as reward values among a plurality of response candidates. For example, if the number of response candidates for the query h _t in the current conversation step is 10 and the number of response candidates uttered with the calculated compensation value is 5, N(h _t ) is 5.

And the search response candidate x _t is delivered to the response speech unit 300. For the _search response candidate It is configured to derive one response candidate among a plurality of response candidates and utter the derived response candidate, and the response utterance unit 300 includes a current compensation value calculation module 310, a future accumulated compensation value calculation module 320, It may include a final compensation value calculation module 330, an average compensation value derivation module 340, and an ignition module 350.

Here, the current compensation value operation module 310 learns the received search response candidate based on a pre-built compensation operation model to derive the current compensation value, and the derived current compensation value is transmitted to the final compensation value operation module 330. do. Here, deriving the reward value for the search response candidate based on the reward calculation model has already been applied in the conversation agent system, and although the process of deriving the reward value for the search response candidate is not specifically specified in this specification, those skilled in the art It must be understood at this level.

Meanwhile, the future cumulative reward value calculation module 320 accumulates the reward values of multiple response candidates for each of the multiple query candidates at the next conversation step t+1 based on the current conversation step t for the search response candidate x _t . This derives the cumulative compensation value. Here, the cumulative reward value is derived based on a learning model built with the Generalized Advantage Estimation (GAE) algorithm that optimizes the reward value of each response candidate using a policy gradient technique. In other words, the model built using the GAE technique is a policy gradient method that optimizes cumulative reward values. It uses the advantage function of the exponential weight estimator to reduce the variance of the policy gradient estimate, and uses the value function to reduce the bias ( Perform trust region optimization by adjusting bias, variance, and tradeoff. Accordingly, based on the GAE algorithm, the compensation value of multiple response candidates for each query candidate at the next search time t+1 for the input search response candidate is estimated, and the estimated reward values are accumulated to output the cumulative reward value.

And the output future cumulative compensation value is transmitted to the final compensation value calculation module 330.

The final compensation value calculation module 330 derives the final compensation value as the sum of the future cumulative compensation value and the current compensation value of multiple response candidates for one query, and the derived final compensation value is calculated from the average compensation value derivation module 340. is passed on.

The average reward value derivation module 340 derives the average reward value for multiple response candidates for all query candidates in the next dialogue step t+1 as the average value of the accumulated reward values of multiple response candidates for one query, and The derived average compensation value is transmitted to the speech module 350.

The speech module 350 derives a response candidate for the current conversation step using the average compensation value and utters the derived response candidate. Here, a series of processes for deriving one response sign among multiple response candidates using the average reward value has already been applied in the conversation agent system, and in this specification, one response candidate is selected by reflecting the derived reward value and the selected response candidate is selected. Although the process of uttering a response candidate is not specifically specified, it should be understood at the level of those skilled in the art.

Meanwhile, the response utterance unit 300 further includes a filter module 360, where the filter module 360 compares the current compensation value with a predetermined minimum allowable compensation value, and as a result of the comparison, the current compensation value is smaller than the minimum allowable compensation value. In this case, the derived search response candidate is determined to be a malicious response candidate and the search response candidate is removed from the majority of response candidates. As a result, malicious conversations are deleted to prevent emotional damage between users and the system.

In addition, the system of one embodiment further includes an additional response candidate generating unit 400, and the additional response candidate generating unit 400 generates a dialogue step for the query of the current dialogue step input from the search response candidate generating unit 200. If the number does not reach the predetermined maximum number of conversation steps and the number of searches for the query of the current conversation step input is greater than the estimated number of response candidates, new response candidates are additionally estimated.

Accordingly, in one embodiment, based on the current dialogue step, the accumulated compensation value for each of the plurality of response candidates for each of the plurality of query candidates in the future dialogue step is reflected, and one of the plurality of response candidates in the current dialogue step is reflected. By deriving response candidates and uttering the derived response candidates, a natural conversation can be continued in real time on the open domain, and by deleting malicious response candidates among multiple response candidates, malicious conversations can be prevented in advance.

FIG. 3 is an overall flowchart showing the operation process of the conversation agent system using conversation prediction according to one embodiment. Referring to FIGS. 1 to 3, the conversation agent method using conversation prediction according to another embodiment of the present invention is explained.

In one embodiment, the conversation learning model building unit 100 collects a lot of conversation data and previous conversation data from movie and drama subtitles, social media, etc., and collects a conversation learning model to estimate a number of response candidates for the input query. Let's assume it has been built.

In step S11, the search response candidate derivation unit 200 of one embodiment derives a plurality of response candidates for the query input in the current conversation step and stores the number of conversation step depth for the query input in the current conversation step. When the set maximum number of conversation steps Max_depth is reached (step S13) or the search number N(h _t ) of multiple response candidates for the query is not greater than the estimated number children(h _t ) of multiple response candidates (S15) Search response candidates are derived from among the response candidates (S17).

And in step S19, the response utterance unit 300 of one embodiment derives a current compensation value by learning based on a pre-built compensation calculation model for the derived search response candidate.

Meanwhile, in steps S21 and S23, the response utterance unit 300 of one embodiment compares the current compensation value r with a predetermined minimum allowable compensation value MIN_R and determines that the current compensation value r is smaller than the predetermined minimum allowable compensation value MIN_R. In this case, it is judged to be a malicious conversation and the derived search response candidate is deleted.

If the current compensation value r is not smaller than the predetermined minimum allowable compensation value MIN_R, in step S25, the response speech unit 300 of one embodiment selects the next dialogue step for the search response candidate x _t based on the current dialogue step t. The future cumulative reward value is derived by accumulating the reward values of the multiple response candidates for each of the multiple query candidates at t+1.

Then, in step S27, the response speech unit 300 of one embodiment adds the current reward value and the future cumulative reward value to derive the final reward value of multiple response candidates for one query candidate and stores the result.

In step S29, the response speech unit 300 of one embodiment determines whether one query candidate is the last query candidate among the estimated plurality of query candidates, and if the determination result is the last query candidate, in step S31, The response speech unit 300 of the embodiment calculates the average value of the final reward values of multiple response candidates for one derived query candidate and derives the average reward value.

Then, in step S33, the response speech unit 300 of one embodiment derives a response candidate for the user's inquiry of the current conversation step by reflecting the average compensation value and utters the derived response candidate.

Meanwhile, in steps S13 and S15, the number of dialogue steps for the query of the current dialogue step input does not reach the predetermined maximum number of dialogue steps, and the number of searches for the query of the current dialogue step input is reduced to the estimated number of responses. If the number of candidates is greater than the number of candidates, in step S35, the additional response candidate generator 400 of one embodiment adds a response candidate to the user inquiry of the current conversation step.

Accordingly, according to one embodiment, based on the current conversation step, the cumulative reward value for each of the plurality of response candidates for each of the plurality of question candidates in the future dialogue step is reflected and the cumulative compensation value of each of the plurality of response candidates in the current dialogue step is reflected. By deriving one response candidate and uttering the derived response candidate, a natural conversation can be continued in real time on the open domain, and by deleting a malicious response candidate among multiple response candidates, malicious conversation can be prevented in advance.

The embodiments described above may be implemented with hardware components, software components, and/or a combination of hardware components and software components. For example, the devices, methods, and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, and a field programmable gate (FPGA). It may be implemented using one or more general-purpose or special-purpose computers, such as an array, programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. A processing device may execute an operating system (OS) and one or more software applications that run on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include a plurality of processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. Software and/or data may be used on any type of machine, component, physical device, virtual equipment, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. Computer-readable media may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on a computer-readable medium may be specially designed and configured for an embodiment or may be known and usable by those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. Includes magneto-optical media and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Although the embodiments have been described with limited drawings as described above, those skilled in the art can apply various technical modifications and variations based on the above. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

100: Dialogue learning model construction unit
200: Search response candidate derivation unit
300: response firing unit
310: Current compensation value calculation module
320: Future accumulated compensation value calculation module
330: Final compensation value calculation module
340: Average compensation value derivation module
350: Ignition module
400: Additional response candidate generation unit

Claims (10)

A conversation learning model building unit that collects a plurality of conversation data and previous conversation data and builds a conversation learning model to estimate a number of response candidates for the query of the input current conversation step;
If the number of dialogue steps for a query in the current dialogue step entered reaches the predetermined maximum number of dialogue steps, or the number of searches for a query in the current dialogue step entered is not greater than the estimated number of response candidates, a search response candidate deriving unit that derives one response candidate to be searched among the estimated plurality of response candidates; and
Based on the current dialogue step for the search response candidate, the cumulative reward value of the plurality of response candidates for each query candidate is reflected as a learning result of the dialogue learning model constructed in the next dialogue step, and the plurality of the current dialogue step is calculated. A conversation agent system using conversation prediction, characterized by including a response speech unit that derives one response candidate among response candidates and utters the derived response candidate. The method of claim 1, wherein the response speech unit,
a current compensation value calculation module that calculates the compensation value of the search response candidate and outputs a current compensation value;
a future cumulative reward value calculation module for deriving a cumulative reward value of a plurality of response candidates for a plurality of query candidates in a next conversation step based on the current conversation step for the search response candidate;
a search response candidate reward value calculation module that derives a reward value of the search response candidate from the sum of the future cumulative reward value and the current reward value; and
An average reward value derivation module that derives the average of search response candidate reward values for a plurality of response candidates; and
A conversation agent system using conversation prediction, characterized by including a speech module that derives one response among the plurality of response candidates by averaging the reward values of all search response candidates for the plurality of response candidates and utters the derived one response. . The method of claim 2, wherein the current compensation value calculation module,
Conversation prediction further comprising a filter module that determines the search response candidate as a malicious response candidate and deletes the search response candidate from the plurality of response candidates when the calculated current compensation value is smaller than a predetermined minimum allowable compensation value. A conversation agent system using . The system of claim 1, wherein:
If the number of dialogue steps for the query of the input current dialogue step does not reach the predetermined maximum number of dialogue steps and the number of searches for the query of the input current dialogue step is greater than the estimated number of response candidates, a new response candidate A conversation agent system using conversation prediction, further comprising an additional response candidate generator that additionally estimates . The method of claim 2, wherein the cumulative compensation value is
A conversation agent system using conversation prediction, characterized by being derived by the GAE (Generalized Advantage Estimation) algorithm that optimizes the cumulative reward value using a policy gradient technique. In the conversation agent method using conversation prediction, which is performed based on a conversation agent system using conversation prediction, including a conversation learning model building unit, a search response candidate derivation unit, and a response speech unit,
A conversation learning model building step of collecting a plurality of conversation data and previous conversation data to build a conversation learning model that estimates a number of response candidates for the query of the current conversation step input;
If the number of dialogue steps for a query in the current dialogue step entered reaches the predetermined maximum number of dialogue steps, or the number of searches for a query in the current dialogue step entered is not greater than the estimated number of response candidates, A search response candidate derivation step of deriving one response candidate to be searched from among the estimated plurality of response candidates; and
Based on the current dialogue step for the search response candidate, the cumulative reward value of the plurality of response candidates for each query candidate is reflected as a learning result of the dialogue learning model constructed in the next dialogue step, and the plurality of the current dialogue step is calculated. A conversation agent method using conversation prediction, comprising a response utterance step of deriving one response candidate among response candidates and uttering the derived response candidate. The method of claim 6, wherein the response utterance step is,
calculating a current compensation value by calculating a compensation value of the search response candidate;
Deriving a future cumulative reward value of a plurality of response candidates for a plurality of query candidates in a next dialogue step based on the current dialogue step for the search response candidate;
Deriving a reward value of the search response candidate from the sum of the future accumulated reward value and the current reward value;
Deriving an average of search response candidate reward values for a plurality of response candidates; and
A conversation agent method using conversation prediction, comprising the step of deriving one response among the plurality of response candidates by averaging the reward values of all search response candidates for the plurality of response candidates and uttering the derived one response. The method of claim 7, wherein calculating the current compensation value includes:
Conversation prediction further comprising the step of determining the search response candidate as a malicious response candidate and deleting the search response candidate from the plurality of response candidates when the calculated current reward value is less than a predetermined minimum allowable reward value. Conversational agent method used. The method of claim 6, wherein
If the number of dialogue steps for the query of the input current dialogue step does not reach the predetermined maximum number of dialogue steps and the number of searches for the query of the input current dialogue step is greater than the estimated number of response candidates, a new response candidate A conversation agent method using conversation prediction, further comprising an additional response candidate generation step of additionally estimating . A computer-readable recording medium on which a program for executing the conversation agent method using conversation prediction according to any one of claims 6 to 9 is recorded.