Dialogue Goal Optimization Apparatus and Method

Abstract

This invention relates to a dialogue goal optimization device, method, and program for optimizing goal achievement in user interactions. Based on user utterances and actions, it estimates user states, including emotions, resistance, comprehension, and signs of detachment, as well as goal progress indicating the degree of achievement of primary and sub-goals, and maps these to a state space. After restricting the set of candidate response strategies based on constraints such as compliance and cost applied to the dialogue, it selects a response strategy that maximizes the objective function. Furthermore, it performs a self-correction process that dynamically updates the parameters of the response strategy in response to signs of failure detected during the dialogue. This makes it possible to suppress user detachment while adhering to practical constraints and continuously improve the business goal achievement rate.

Description

Technical Field

[0001] The present invention relates to a dialogue goal optimization device that optimizes goal achievement in dialogue by dynamically selecting and self-correcting response strategies according to user status and constraint conditions.

Background Art

[0002] Many conventional dialogue systems such as chatbots are based on fixed rules or general reinforcement learning, and the efficiency for achieving business goals has not been optimized. According to Patent Document 2, dialogue agents using multiple learning engines have been proposed, but these mainly focus on emotional communication with users and have problems in decision-making under strict constraint conditions. In particular, a structure for detecting failure signs such as signs of user departure and self-correcting response strategies in real time while managing compliance and dialogue costs has not been established, and it has been difficult to maximize the goal achievement rate according to dynamic situation changes.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the prior art such as the dialogue system described in Patent Document 2, the main focus is on improving the accuracy of response selection, and the perspective of maximizing the efficiency of achieving business goals is lacking. In addition, it is difficult to perform control that explicitly considers practical constraints such as compliance and costs, and constructing a mechanism for pursuing results while complying with strict operation rules in the real world has been a major issue.

[0005] Furthermore, when user resistance occurs during a conversation, the system lacks the ability to grasp the situation and modify the strategy under uncertain circumstances, as described in Patent Document 1. It fails to prevent users from abandoning the conversation due to continued inappropriate responses. A key technical challenge is achieving advanced self-correction capabilities that autonomously update conversation parameters based on the detection of signs of failure and flexibly guide the user towards the goal according to the situation. [Means for solving the problem]

[0006] The dialogue goal optimization device according to the present invention comprises a state estimation unit that maps the user state and goal progress in a dialogue with the user to a state space, a constraint management unit that manages constraints related to the dialogue, a strategy selection unit that selects the optimal response strategy from a set of candidate strategies to maximize the objective function, and a self-correction unit that detects signs of failure during the dialogue and dynamically updates the parameters of the response strategy. Under constraints such as compliance and cost, this device enables the achievement rate of business goals based on the objective function while preventing dialogue dropouts by activating a self-correction loop that dynamically switches strategies in response to the user's emotions and signs of abandonment. [Effects of the Invention]

[0007] According to the present invention, by integrating dynamic strategy selection and constraint management based on an objective function, and a self-correction function based on signs of failure, it is possible to achieve both compliance with practical constraints such as compliance and user churn prevention, thereby maximizing the achievement rate of business goals. Furthermore, since transparency of the dialogue process is ensured by recording the rationale for strategy selection, it has the excellent effect of significantly improving adaptability to actual business operations and economic convenience compared to conventional dialogue systems. [Brief explanation of the drawing]

[0008] [Figure 1] This system configuration diagram shows an example of the overall configuration of a system including the dialogue goal optimization device according to this embodiment. [Figure 2] This block diagram shows the main functional configuration of the dialogue goal optimization device. [Figure 3] This is an explanatory diagram illustrating the concept of the process in the state estimation unit that maps the user state and goal progress to the state space. [Figure 4] This figure shows an example of how to define the slot fulfillment rate for the set main goal and sub-goals. [Figure 5] This flowchart shows the procedure for selecting the optimal response strategy based on the objective function in the strategy selection unit. [Figure 6] This flowchart shows the process by which the self-correction unit detects signs of failure during dialogue and dynamically updates the parameters of the response strategy. [Figure 7] This diagram shows the configuration of compliance regulations and cost constraints that are dynamically applied by the constraint management unit. [Figure 8] This diagram shows the data structure of the audit log, including the rationale for selecting the response strategy output by the log generation unit. [Figure 9] This flowchart shows the overall processing procedure for the dialogue goal optimization method according to this embodiment. [Figure 10] This is a block diagram showing the computer hardware configuration for realizing the dialogue goal optimization device according to this embodiment. [Modes for carrying out the invention]

[0009] As shown in Figure 1, a dialogue goal optimization system according to one embodiment of the present invention comprises a dialogue goal optimization device 10 that manages the optimization of dialogue, a user terminal 20 operated by the user, and a network 30 that connects these to each other. The dialogue goal optimization device 10 communicates with one or more user terminals 20 via the network 30 and, as necessary, cooperates with an external system 800 to acquire information and perform processing. The external system 800 is, for example, a group of computers having functions such as a database server, an external reservation management system, or an information search engine.

[0010] The user terminal 20 is an information processing device such as a smartphone, personal computer, or tablet, which provides the user with a dialogue interface using voice or text. The network 30 includes wireless or wired communication paths, which are composed of a wide-area communication network such as the Internet, or a local area network. As shown in the overall processing procedure in Figure 9, the dialogue goal optimization device 10 generates an optimal response based on the user's speech data and operation input transmitted from the user terminal 20 and sends it back to the user terminal 20.

[0011] Next, with reference to Figure 10, the hardware configuration for realizing the dialogue goal optimization device 10 according to this embodiment will be described. The dialogue goal optimization device 10 has a basic configuration as a computer and includes a CPU 900 for arithmetic processing, a memory 910 as main memory, and a storage 920 as auxiliary memory. These components are interconnected via a bus or the like, enabling the exchange of information.

[0012] The CPU 900 is a central processing unit that realizes various functional units in the dialogue goal optimization device 10 by loading predetermined programs stored in the storage 920, etc., onto the memory 910 and executing them. The memory 910 is a volatile memory such as RAM, and functions as a workspace for the CPU 900, as well as temporarily holding various data to be processed. The storage 920 is a non-volatile storage device such as an HDD or SSD, and holds various data necessary for dialogue control, trained models, and executable programs.

[0013] Figure 2 is a block diagram showing the main functional configuration of the dialogue goal optimization device 10. The dialogue goal optimization device 10 functionally includes a state estimation unit 100, a constraint management unit 110, a strategy selection unit 120, a self-correction unit 130 shown in Figure 6, and a log generation unit 140 shown in Figure 8. Each of these units is realized as a means for hardware resources and software to work together by the aforementioned CPU 900 executing a specific program.

[0014] In the interaction with the user via the user terminal 20, the state estimation unit 100 executes a process of analyzing in real time the user's state and the progress of the interaction that change moment by moment. Specifically, the state estimation unit 100 extracts a user state 200 estimated from the interaction history and acoustic feature amounts, and a goal progress 210 indicating the degree of achievement of the interaction goal, and performs a physical arithmetic process of mapping these to an abstract multi-dimensional state space 220.

[0015] The user state 200 includes a plurality of psychological and situational indicators estimated from the user's utterance content, voice tone, intonation, response timing, etc. In the present embodiment, the state estimation unit 100 has a function of estimating at least one of the user's emotion, the sense of resistance to the content proposed by the system, the degree of understanding of the interaction content, and the sign of trying to disengage from the interaction as the user state 200. Thereby, not only the literal meaning but also the user's potential intention can be grasped.

[0016] The goal progress 210 is an index that quantitatively indicates how much a goal (for example, purchase of a product, confirmation of a reservation, listening of specific information, etc.) defined in advance in the interaction has been achieved at the current time. The state estimation unit 100 determines at which stage the current interaction is in toward the final goal for each exchange with the user or for each predetermined segment, and updates the goal progress 210.

[0017] FIG. 3 shows the concept of the process of mapping the user state 200 and the goal progress 210 to the state space 220. The state space 220 is configured as a multi-dimensional vector space with numerically quantified plural parameters as axes. The state estimation unit 100 integrates each component of the estimated user state 200 and the value of the current goal progress 210, and arranges them in the state space 220 as one state vector.

[0018] This mapping process is physically performed by the CPU 900 through matrix operations and neural network forward propagation. The state estimation unit 100 normalizes the features extracted from the input dialogue data and generates a vector with these features as components, thereby identifying the characteristics of the current situation within the vast number of dialogue patterns as coordinates in the state space 220.

[0019] In user state 200, treating emotions, resistance, comprehension, and signs of detachment as independent dimensions allows for capturing subtle changes in the user's responses as the dialogue progresses. For example, even if the content of the utterance is positive, if the tone of voice indicates high resistance, the coordinates in state space 220 are controlled to move to a specific area. This enables state management that goes beyond superficial responses and is tailored to the user's psychological state.

[0020] Furthermore, the state estimation unit 100 continuously monitors the goal progress 210 along the time axis and stores it as a trajectory in the state space 220. By analyzing this trajectory, it statistically determines whether the dialogue is progressing smoothly or is stagnating. Specific transition patterns in the state space 220, such as an increase in signs of withdrawal without improvement in goal progress 210, serve as important triggers for modifying the dialogue strategy in subsequent processing.

[0021] In this way, by employing a physical data processing structure—a mapping to the state space 220—qualitative information such as user emotions and the context of the dialogue can be converted into a format that can be handled by a computer. This conversion process makes it possible to evaluate diverse user characteristics and complex dialogue scenarios using common metrics, forming a foundation for sophisticated decision-making adapted to individual dialogue situations.

[0022] The information mapped to the state space 220 is temporarily recorded in a predetermined area on the memory 910 and becomes immediately accessible from other functional units, such as the constraint management unit 110 shown in Figure 7 and the strategy selection unit 120 shown in Figure 5. In this way, by centrally managing the state of the dialogue with a single spatial model, the data processing efficiency of the entire device is improved, and real-time dialogue response with low latency is realized.

[0023] The definition of goal progress 210 in the dialogue goal optimization device 10 will now be explained. In this embodiment, the progress of the dialogue is managed from multiple perspectives, not only by a single outcome indicator, but also by the degree of achievement of multiple structured goals.

[0024] As shown in Figure 4, the goals have a hierarchical or parallel structure and consist of a main goal 300, which is the final destination of the dialogue, and one or more sub-goals 310 that should be achieved in the process of reaching the main goal 300.

[0025] The main goal 300 defines the business outcomes that the dialogue system should ultimately aim for, such as closing a service deal, concluding a contract, or solving a specific complex problem.

[0026] Subgoal 310 is set as a guideline for collecting the information necessary to achieve the main goal 300 in stages. For example, confirming individual items such as the user's name, contact information, or specific requests corresponds to each subgoal 310.

[0027] Progress toward these main goals 300 and subgoals 310 is quantitatively defined using the slot completion rate 320. The slot completion rate 320 represents the percentage of slots—which are pre-defined necessary information items for achieving each goal—that have already been identified through actual dialogue.

[0028] Specifically, as illustrated in Figure 4, by calculating the ratio of the number of completed slots to the total number of slots associated with a particular subgoal 310, it becomes possible to grasp the current progress as a numerical value between 0 and 1. This allows for an objective and precise evaluation of the dialogue's progress.

[0029] Next, we will describe the dynamic constraint management process performed by the constraint management unit 110 via the network 30 by the dialogue goal optimization device 10 shown in Figures 1 and 2. In each turn of the dialogue, the constraint management unit 110 refers to multiple constraint conditions 600 and performs strict control to ensure that the system's output and actions remain within acceptable limits.

[0030] As shown in Figure 7, the constraints 600 consist of items related to compliance regulations, prohibited expressions, dialogue time, and cost. The scope and priority of these conditions are dynamically changed depending on the context of the dialogue and changes in the user's state.

[0031] Compliance regulations and restrictions on prohibited expressions define what the system should not say and what important notices must be mentioned, based on legal regulations, ethical standards, or specific internal organizational rules. The Restriction Management Unit 110 continuously evaluates whether the generated response candidates violate these regulations.

[0032] Constraints on dialogue time manage the total duration allowed for a single dialogue session and the acceptable waiting time for individual responses. Constraints on cost are used to keep the costs associated with maintaining the dialogue, such as API charges for using external systems 800 and the consumption of computing resources, below predetermined thresholds.

[0033] The constraint management unit 110 executes a filtering mechanism to limit the search range of candidate strategies based on the constraint conditions 600, prior to the selection of a strategy. This mechanism preemptively excludes actions that may deviate from the constraints under the current circumstances from the strategic options.

[0034] By introducing this filtering mechanism, the overall security of the system is ensured, and response speed is improved by eliminating unnecessary strategy searches. For example, if there is little time remaining for the conversation, strategies involving lengthy explanations are filtered out, and the system is limited to strategies that only require concise consent confirmation.

[0035] Furthermore, the constraint management logic for constraints 600 also adjusts the strength of the constraints according to the phase of the dialogue. The constraint management unit 110 performs a dynamic application process in which constraints that reduce the user's psychological burden are emphasized in the introductory part of the dialogue, and compliance constraints that ensure the accuracy of the contract are applied more strictly in the closing part.

[0036] The strategy selection unit 120 determines the most appropriate behavior at each point in the dialogue, following a series of processing steps illustrated in Figure 5. Specifically, based on the current dialogue situation mapped to the state space 220 by the state estimation unit 100 shown in Figure 3, it extracts the action the system should take from a predefined set of candidate strategies 410.

[0037] The candidate strategy set 410 consists of multiple response options, such as asking the user a question, summarizing information, making specific suggestions, expressing empathy, or executing an external system 800. These options may be dynamically filtered depending on the context of the dialogue and constraints 600.

[0038] In selecting the optimal strategy, the objective function 400 is calculated. The objective function 400 is an evaluation metric used to quantify the impact of the dialogue's progress on business value and user experience, and is defined as a formula that integrates multiple independent variables.

[0039] One of the main independent variables that make up the objective function 400 is the goal progress 210, which indicates the feasibility of completing the main goal 300 and achieving the sub-goals 310, as shown in Figure 4. This index is evaluated based on the slot completion rate 320, etc., and is designed so that the higher this value, the greater the overall value of the objective function 400 increases in the positive direction.

[0040] Furthermore, the objective function 400 incorporates user detachment and discomfort, which are included in the user state 200, as independent variables. These act as negative rewards, and when the self-correction unit 130 detects signs of failure 500, such as increased user resistance or silence, it updates the strategy parameter 510 in the procedure shown in Figure 6, thereby reducing the value of the objective function 400 and avoiding the undesirable state.

[0041] Furthermore, dialogue cost is also considered as an independent variable of the objective function 400. Dialogue cost includes the number of utterances, time, and computational resources required to achieve the goal. This prioritizes strategies that eliminate redundant dialogue and reach the goal in the shortest and most efficient way possible.

[0042] The formula for calculating the objective function 400 is derived by assigning appropriate weighting coefficients to these variables and then adding or subtracting them. These weighting coefficients can be dynamically adjusted according to the type and purpose of the interaction, or specific business requirements.

[0043] The strategy selection unit 120 employs a reinforcement learning algorithm that aims to maximize the expected reward obtained over time, i.e., the cumulative value of the objective function 400, following the overall processing procedure shown in Figure 9. This pursues not only immediate rewards for single responses, but also the long-term and optimal goal achievement throughout the entire dialogue.

[0044] In the strategy decision process, the strategy selection unit 120 sequentially calculates the probability of transitioning to the next state and the expected reward obtained when each element of the candidate strategy set 410 is selected. This search process identifies the statistically most promising action, even in a dialogue environment with high uncertainty.

[0045] The final identified action is output as a response strategy 420. The response strategy 420 not only generates text or voice response messages to the user terminal 20, but also has the properties of outputting an audit log 700 including the selection basis 710 by the log generation unit 140 shown in Figure 8, and a control command for calling functions of an external system 800.

[0046] An important element of the response strategy 420 shown in Figure 5 is cooperation with the external system 800. The strategy selection unit 120 controls the external system 800 to execute specific processes if it determines that it is necessary to achieve the goal of the dialogue.

[0047] Specific examples of controlling the external system 800 include searching for information, booking facilities and services, or executing API calls to complete specific business processes. For example, if the data necessary to satisfy a user's request is not present internally, the system can supplement the information by selecting a search strategy.

[0048] Furthermore, it is possible to select a strategy in which, once the conditions are met during the dialogue, reservation commands and API calls for data linkage are automatically sent to the external system 800. By incorporating external linkage into the response strategy 420, practical actions can be automated on the hardware configuration shown in Figure 10.

[0049] Details of the processing performed by the self-correcting unit 130 will be explained (see Figure 2). The self-correcting unit 130 has the function of monitoring the user's response in real time during the conversation and detecting failure indicators 500 that indicate the conversation is heading in an undesirable direction. This makes it possible to deal with changes in the user's psychology that are difficult to handle with fixed scenarios.

[0050] The failure indicators 500 include prolonged silence in user responses, a sudden increase in negative expressions, repeated requests for the same content, or specific keywords suggesting the termination of the dialogue. These are determined based on the amount of change and the temporal progression of the user state 200 output from the state estimation unit 100, according to the flowchart shown in Figure 6.

[0051] When the self-correction unit 130 detects a failure indicator 500, it dynamically updates the strategy parameters 510 referenced by the strategy selection unit 120. The strategy parameters 510 include weighting coefficients for controlling the selection probability of candidate response strategies 420, and definition information for the set of strategies that should be prioritized in the current context.

[0052] As shown in Figure 6, the self-correction unit 130 adjusts the strategy parameters 510 according to the type and intensity of the failure indicators 500. For example, if it is determined that user resistance has increased, the unit performs an update process that reduces the weight of strategies that prioritize providing information and encouraging action, and increases the weight of strategies that express empathy and re-explain the objective.

[0053] This dynamic parameter update allows the strategy selection unit 120 to immediately select a new response strategy 420 that minimizes the risk of failure in the previous dialogue, following the procedure shown in Figure 5. This means that the dialogue trajectory is corrected in response to the user's current reaction, rather than branching based on past static condition judgments.

[0054] Next, the process of generating audit logs 700 by the log generation unit 140 will be described. The log generation unit 140 generates the content of the response strategy 420 selected at each step of the dialogue, along with the selection rationale 710 which shows the logical basis for that selection, and records it as audit logs 700.

[0055] Figure 8 shows an example of the data structure of the audit log 700. The audit log 700 stores, in an interconnected manner, the timestamp when the interaction was performed, the estimated user state 200, the applied constraints 600 as shown in Figure 7, the calculated values ​​of each variable in the objective function 400, and the information of the response strategy 420 that was ultimately adopted.

[0056] Selection rationale 710 is metadata that explains why a particular strategy was preferred over other candidates. This includes specific parameters that contributed to maximizing the objective function 400, as well as a list of alternative strategies excluded due to compliance constraints. This ensures transparency and accountability in the conversational AI's decision-making process.

[0057] Furthermore, the device according to this embodiment includes a function to automatically extract the success factors and failure factors of a dialogue. Based on outcome information such as whether the dialogue ultimately achieved the main goal 300 or whether a dropout occurred along the way, the log generation unit 140 evaluates the effectiveness of each strategy in the dialogue. In this process, as shown in Figures 3 and 4, changes in the goal progress 210 and slot occupancy rate 320 in the state space 220 are referenced.

[0058] The accumulated audit logs 700 and evaluation results are stored as training data and returned to the system through the overall processing procedure shown in Figure 9. This cyclical process improves the accuracy of the state estimation unit 100 during the next interaction and improves the optimization logic of the objective function 400 in the strategy selection unit 120.

[0059] Specifically, by using contextual information from past instances where failure indicators 500 were detected as training data, the state estimation unit 100 can predict user discomfort and signs of churn at an earlier stage. Furthermore, the self-correction unit 130 can calculate more appropriate updated values ​​for the strategic parameters 510 based on past successful correction cases.

[0060] With the above configuration, it is possible to apply self-correction in real time to unexpected stagnation during dialogue and psychological barriers to the user, while accumulating the entire decision-making process as an audit log 700, leading to continuous quality improvement. In this embodiment, as shown in Figure 1, log generation is performed by a log generation unit 140 inside the device, but it is also possible to outsource some processing to an external system 800 and accumulate learning data in a distributed manner. Furthermore, as shown in Figure 10, the dialogue goal optimization device 10 is realized by a computer equipped with a CPU 900, memory 910, and storage 920. [Explanation of Symbols]

[0061] 10 Dialogue Goal Optimization Device 20 User Terminals 30 Networks 100 State Estimation Unit 110 Constraint Management Department 120 Strategic Selection Department 130 Self-correction section 140 Log generation unit 200 User Status 210 Goal Progress 220 State Space 300 Main Goals 310 Sub-goals 320 slots filled 400 Objective Function 410 Candidate Strategy Set 420 Response Strategy 500 Signs of Failure 510 Strategy Parameters 600 Constraints 700 audit logs 710 Selection Basis 800 External Systems 900 CPU 910 memory 920 storage

Claims

1. A dialogue goal optimization device that generates and outputs a response based on dialogue data in a user interaction, A state estimation unit that maps the user state and goal progress to a state space, A constraint management unit manages constraints including at least one of compliance regulations, prohibited expressions, constraints on dialogue time including the total duration of dialogue sessions and response waiting time, and constraints on costs including API charges and computing resource consumption associated with the use of external systems, and, prior to the selection of a response strategy by the strategy selection unit, performs filtering to exclude response strategies that violate the constraints from the set of candidate response strategies based on the said constraints to limit the search scope, A strategy selection unit selects the optimal response strategy from the aforementioned limited set of candidates to maximize an objective function that includes at least two independent variables: goal achievement rate, user drop-off, discomfort, and interaction cost. A self-correcting unit that detects signs of failure during the dialogue and dynamically updates the weight coefficients of the objective function and / or the strategy parameters that define the selection probability of the response strategy in accordance with the detected signs of failure, A log generation unit stores the selected response strategy, the constraints, the calculated values ​​of each independent variable of the objective function, and the response strategy in relation to each other as an audit log, and also stores the selection basis, including at least one of the applicable rule ID, score, and judgment criterion, in the audit log. Equipped with, The aforementioned signs of failure include at least one of the following keywords that suggest prolonged silence in user responses, a surge in negative language, repeated requests for the same content, or forced termination of the conversation: The basis for the selection includes a list of alternative strategies excluded by the constraints, Dialogue goal optimization device.

2. The interactive goal optimization device according to claim 1, wherein the goal progress is defined by the slot completion rate for the set main goal and sub-goals.

3. The dialogue goal optimization apparatus according to claim 1, wherein the objective function is derived by a calculation formula that assigns weight coefficients to the independent variables and adds or subtracts them, and the weight coefficients are dynamically adjustable according to the type, purpose, or business requirements of the dialogue.

4. The dialogue goal optimization device according to claim 1, wherein the state estimation unit estimates at least one of the user's emotions, resistance, level of understanding, and signs of detachment as the user state.

5. The strategy selection unit selects a search, reservation, or API call using an external system as an element of the response strategy. The dialogue goal optimization device according to claim 1, wherein the constraint management unit excludes the response strategy by filtering so that the API charge amount associated with the use of the external system is below a predetermined threshold.

6. The dialogue goal optimization apparatus according to claim 1, wherein the audit log includes a timestamp, an estimated user state, applied constraints, calculated values ​​of each independent variable of the objective function, a response strategy adopted, and the rationale for the selection.

7. The dialogue goal optimization device according to claim 1, wherein the constraint management unit changes the scope of application or priority of the constraint conditions according to the dialogue phase.

8. Computers A state estimation process that maps the user state and goal progress in the user interaction to a state space, A constraint management step involves managing constraints, including at least one of compliance regulations, prohibited expressions, constraints on dialogue time including the total duration of dialogue sessions and response waiting time, and constraints on costs including API charges and computing resource consumption associated with the use of external systems, and prior to the strategy selection step, performing filtering to exclude response strategies that violate the constraints from the set of candidate strategies based on the said constraints to limit the search scope. A strategy selection step involves selecting the optimal response strategy from the aforementioned set of candidate strategies with a limited search range so as to maximize an objective function that includes at least two independent variables: goal achievement rate, user drop-off, discomfort, and interaction cost. A self-correction process that detects signs of failure during the dialogue and dynamically updates the weight coefficients of the objective function and / or the strategy parameters that define the selection probability of the response strategy in accordance with the detected signs of failure, A log generation step includes storing the selected response strategy, the constraints, the calculated values ​​of each independent variable of the objective function, and the response strategy in relation to each other as an audit log, as well as storing the selection rationale, which includes at least one of the applicable rule ID, score, and decision criterion, in the audit log, and including a list of alternative strategies excluded by the constraints in the selection rationale, A method for optimizing dialogue goals to execute them.

9. The dialogue goal optimization method according to claim 8, wherein in the state estimation step, the goal progress is calculated using the slot fulfillment rate for the main goal and sub-goals, and updated in real time during the dialogue.

10. The dialogue goal optimization method according to claim 8, further comprising the steps of accumulating the audit log as learning data including success factors or failure factors of the dialogue, and using it for the state estimation step or updating the strategy parameters in subsequent dialogues.

Images