Action determination device, action determination method, and action determination program

Abstract

To improve the efficiency in consensus building.SOLUTION: An action determination device executes: search processing which searches for a plurality of simulation conversation routes from the start to the end of simulation conversation, by executing the simulation conversation between a first agent which simulates a facilitator on the basis of a language model, and a second agent which simulates a participant on the basis of the language model; calculation processing which calculates an evaluation value which evaluates utterance of a simulation response sentence in the simulation conversation generated from the second agent in the plurality of simulation conversation routes, on the basis of an inner state indicating group perception with respect to a participant group of the participant in the second agent; extraction processing which extracts a specific simulation conversation candidate sentence from among the plurality of simulation conversation candidate sentences at a start time point of the simulation conversation, on the basis of the evaluation value; and output processing which outputs the specific simulation conversation candidate sentence.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an action determination device, an action determination method, and an action determination program for determining actions.

Background Art

[0002] The following Non-Patent Document 1 discloses a two-layer facilitation agent designed to model a dynamic discussion process as a Markov Decision Process (MDP) and learn an optimal facilitation policy for multiple rounds of discussions.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the above-mentioned conventional technology, since the facilitation agent learns the utterances that support the conversation through reinforcement learning, it is necessary to collect and learn a large amount of facilitation action data as learning data during execution, and efficient consensus formation is not considered.

Means for Solving the Problems

[0005] An action decision device which is an aspect of the invention disclosed in the present application is an action decision device having a processor that executes a program and a storage device that stores the program, wherein the processor includes a first agent that simulates a facilitator who supports consensus formation in conversations within a group of participants based on a language model, and a second agent that simulates a participant who participates in the conversation based on the language model. By executing a simulated conversation between the first agent and the second agent, a search process is performed to search for a plurality of simulated conversation paths from the start to the end of the simulated conversation. Based on the internal state indicating the group awareness of the participant in the second agent with respect to the group of participants, a calculation process is performed to calculate an evaluation value for evaluating the utterance of the simulated response sentence in the simulated conversation generated from the second agent in the plurality of simulated conversation paths searched by the search process. Based on the evaluation value calculated by the calculation process, an extraction process is performed to extract a specific simulated utterance candidate sentence from among a plurality of simulated utterance candidate sentences from the first agent to the second agent at the start point of the simulated conversation. And an output process is performed to output the specific simulated utterance candidate sentence extracted by the extraction process.

Effect of the Invention

[0006] According to a typical embodiment of the present invention, the efficiency of consensus formation can be improved. Problems, configurations, and effects other than those described above will be clarified by the description of the following examples.

Brief Description of the Drawings

[0007]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

MODE FOR CARRYING OUT THE INVENTION

[0008] In each of the following embodiments, a behavior decision-making system that utilizes a Large Language Model (LLM) to assist in human relationships and consensus formation is provided. The behavior decision-making system incorporates, in particular, an agent for alleviating conflicts in scenarios where human-to-human "conflicts" can be a problem. Specifically, for example, the behavior decision-making system is realized by incorporating a free energy principle and a collective cognitive model into active inference, and agentification (autonomization) by incorporating active inference into the LLM.

[0009] Application scenarios of the behavior decision-making system include 1. Meeting facilitation (promoting consensus formation among multiple participants) 2. Management of Social Networking Service (SNS) or online bulletin boards (suppressing inappropriate interactions among participants such as violent posts) There are. These will be described in detail in each embodiment below.

Embodiment

[0010] <System configuration example> FIG. 1 is an explanatory diagram showing a system configuration example 1 of the behavior decision-making system. The behavior decision-making system 100 includes a behavior decision-making device 101 serving as a server and one or more (4 in FIG. 1 as an example) terminals 102. The behavior decision-making device 101 and the terminals 102 are communicably connected by a network 103 such as the Internet, a Local Area Network (LAN), or a Wide Area Network (WAN). Also, an external device 104 is communicably connected to the behavior decision-making device 101 and the terminals 102 via the network 103.

[0011] The behavior decision-making device 101 and the external device 104 have an LLM 110. When using the LLM 110 of the external device 104, it is not necessary to install the LLM 110 in the behavior decision-making device 101.

[0012] LLM110 is a conversational language model that is deeply learned using a learning dataset related to a large number of conversations, and can perform various tasks such as responses to questions, correction and summarization of texts, translation of texts, and generation of texts. The LLM110 of the action decision device 101 is implemented using, for example, Open Source Software. Examples of the LLM110 of the external device 104 include BERT and ChatGPT.

[0013] The terminal 102 is assigned to participants h1, h2, …. When not distinguishing between participants h1, h2, …, they are denoted as participant h. Participant h is, for example, a user of the terminal 102 who participates in conversations such as meetings and chats. In the example of FIG. 1, since there are 4 terminals 102, there are 4 participants h. Therefore, there are as many terminals 102 as the number of participants h. The terminal 102 performs voice input / output, character string input / output, and speech recognition.

[0014] The action decision device 101 uses the LLM110 to generate a facilitator agent FA, which is the first agent, and participant agents PA1 to PA4, which are the second agents. When not distinguishing between the participant agents PA1 to PA4, they are simply denoted as participant agent PA. When not distinguishing between the facilitator agent FA and the participant agent PA, they are denoted as agent. An agent is an instance that converts the personal information and speech history of each participant h into a prompt and inputs it to the LLM110 to simulate the speech of the facilitator or participant h.

[0015] The facilitator agent FA is an agent that simulates the speech of the facilitator. The facilitator is a virtual progress manager of the conversation who makes speeches to assist in consensus formation in the conversation within the group of participants h. The participant agent PA is an agent that simulates the speech of the participant h.

[0016] In addition, since the LLM 110 has high versatility, without re-training for each agent, by providing the personal information for each agent as input information, it is possible to simulate the speech for each agent. In this case, the facilitator agent FA and the participant agent PA can simulate the speech of the facilitator or the participant h without re-training, by appropriately changing the input information (prompt) corresponding to the instruction to the model. For example, by inputting a prompt such as ""You are a college student in the Faculty of Engineering. What kind of career are you aiming for in the future?"" to the LLM 110, the LLM 110 generates, as a response, a string for each user enclosed in double quotes, such as ""a college student in the Faculty of Engineering"", from the table of personal information.

[0017] The actual situation of a program that generates a prompt for the LLM 110 from personal information and the like, inputs it to the LLM 110, and generates a response sentence is called an instance. Note that since the LLM 110 itself has no state, it can be shared among all agents. Therefore, the LLM 110 itself is not included in the instance. In fact, since the LLM 110 consumes a large amount of capacity, it is not instantiated for each agent. The processing for the LLM 110 is instructed to the shared LLM instance.

[0018] Personal information refers to information specific to a person, such as age, gender, occupation, hobbies, and values. Specifically, for example, the facilitator agent FA is an agent to which personal information about the facilitator is assigned, and the participant agent PA is an agent to which personal information about the participant h is assigned. The agent is generated within the action decision device 101, whether in the case of using the LLM 110 of the action decision device 101 or in the case of using the LLM 110 of the external device 104.

[0019] That is, the personal information and conversation history indicating the state of the agent are input into the LLM110 as a prompt each time. For this reason, there is only one LLM110 itself, and the LLM110 may be implemented either inside or outside the action decision device 101. When using the LLM110 implemented in the external device 104, the state of the LLM110 is not stored in the external device 104 and is input as a prompt from the action decision device 101 to the external device 104.

[0020] In this example, there is no facilitator as a human, but by giving certain personal information to the LLM110 as a prompt, the action decision device 101 acts as a facilitator agent FA as if a facilitator were participating in the conversation. Also, although the participant h is participating in the conversation, by giving certain personal information to the LLM110 as a prompt, the action decision device 101 acts as a participant agent PA as if it were the participant h.

[0021] Specifically, for example, in a conversation in which the participant h participates, the facilitator agent FA actively determines the utterances that the facilitator should make for consensus formation and speaks to the participant h. This utterance is called the utterance in the current conversation state. The utterances in the current conversation state by the facilitator agent FA are prepared in multiple based on a preset topic at the beginning of the conversation, and are prepared in multiple based on the utterances of the facilitator agent FA in response to the utterances of the participant h during the conversation. Each of these multiple prepared utterances of the facilitator agent FA is referred to as a simulated utterance candidate sentence. Note that in order for the facilitator to actively determine the utterances that should be made for consensus formation, the participant agent PA responds to the utterances in the current conversation state determined in the facilitator agent FA on behalf of the participant h.

[0022] The action decision device 101 generates a search tree by continuously conducting a simulated conversation, which is a virtual conversation with multiple patterns that branch out from a plurality of simulated utterance candidate sentences starting from the current conversation state, between the facilitator agent FA and the participant agent PA until the discussion converges. In the search path where the utterance with the highest evaluation exists, the action decision device 101 determines a simulated utterance candidate sentence based on the current conversation state as the utterance sentence (hereinafter referred to as the utterance result sentence) that the facilitator should make for promoting the conversation. The action decision device 101 transmits the utterance result sentence to the terminal 102. The terminal 102 outputs it in a visible manner to the participant h or outputs it as voice by reading it aloud.

[0023] After that, when the participant h is prompted by the utterance of the facilitator agent FA to make an utterance, the action decision device 101 causes the facilitator agent FA to respond to the utterance of the participant h and updates the current conversation state of the facilitator agent FA. The action decision device 101 re-executes the generation and search of the search tree, the identification and output of the utterance result sentence by generating a plurality of simulated utterance candidate sentences based on the updated new current conversation state, and waits for the utterance of the participant h. Through this repetition, the conversation is promoted towards reaching an agreement.

[0024] <Hardware configuration example of the computer (action decision device 101, terminal 102, external device 104)> FIG. 2 is a block diagram showing an example of the hardware configuration of a computer. The computer 200 includes a processor 201, a storage device 202, an input device 203, an output device 204, and a communication interface (communication IF) 205. The processor 201, the storage device 202, the input device 203, the output device 204, and the communication IF 205 are connected by a bus 206. The processor 201 controls the computer 200. The storage device 202 serves as a working area for the processor 201. Also, the storage device 202 is a non-temporary or temporary recording medium that stores various programs and data. Examples of the storage device 202 include a ROM (Read Only Memory), a RAM (Random Access Memory), an HDD (Hard Disk Drive), and a flash memory. The input device 203 inputs data. Examples of the input device 203 include a keyboard, a mouse, a touch panel, a numeric keypad, a scanner, a microphone, and a sensor. The output device 204 outputs data. Examples of the output device 204 include a display, a printer, and a speaker. The communication IF 205 is connected to the network 103 and transmits and receives data.

[0025] <Person Information Table> FIG. 3 is an explanatory diagram showing an example of a person information table. The person information table 300 is stored, for example, in the storage device 202 of the action determination device 101. The person information table 300 has, as fields, a person ID 301, an age 302, an occupation 303, a personality 304, interests 305, a participant flag 306, and a facilitator flag 307. A combination of the values of the fields in the same row becomes an entry indicating person information that identifies a single person. Note that the person information table 300 is not limited to these and may include other fields such as date of birth, gender, and hobbies. Also, the person in the person information may be a real person or a fictional person, but since there is a participant h in the conversation, the person information of a real person is essential.

[0026] The person ID 301 is identification information that uniquely identifies a person. For example, the values h1 to h4 of the person ID 301 correspond to the participants h1 to h4 respectively.

[0027] The age 302 is a count value of the number of years elapsed since the date of birth of the person identified by the person ID 301. The age 302 is updated based on the current time measured by the clock in the action determination device 101 with reference to the date of birth.

[0028] The occupation 303 is the type of current job the person identified by the person ID 301 is engaged in or a position equivalent thereto. The occupation 303 can be updated, for example, by directly operating the action determination device 101 or by indirectly operating the action determination device 101 from the terminal 102.

[0029] The personality 304 is a combination of mental and moral qualities and characteristics peculiar to the person identified by the person ID 301. The personality 304 can be updated, for example, by directly operating the action determination device 101 or by indirectly operating the action determination device 101 from the terminal 102.

[0030] The concern 305 is a word or sentence indicating something the person identified by the person ID 301 is interested in. The concern 305 can be updated, for example, by directly operating the action determination device 101 or by indirectly operating the action determination device 101 from the terminal 102.

[0031] The participant flag 306 is an identifier indicating whether the person identified by the person ID 301 participates in the conversation. "1" indicates participation, and "0" indicates non-participation. The participant flag 306 can be updated, for example, by directly operating the action determination device 101 or by indirectly operating the action determination device 101 from the terminal 102.

[0032] The facilitator flag 307 is an identifier indicating whether the person specified by the person ID 301 is the facilitator of the conversation. If "1" is set, the person specified by the person ID 301 becomes the facilitator. That is, a facilitator agent FA is generated that converts the person information of the entry and the topic information described later into input information (prompt) to the LLM 110. The facilitator flag 307 can be updated, for example, when the terminal 102 accesses the action decision device 101. The facilitator flag 307 can be updated by directly operating the action decision device 101 or by indirectly operating the action decision device 101 from the terminal 102.

[0033] <Topic Information Table> Figure 4 is an explanatory diagram showing an example of a topic information table. The topic information table 400 is stored, for example, in the storage device 202 of the action decision device 101. The topic information table 400 has, as fields, a person ID 301, a topic 401, and a keyword 402. The combination of the values of each field in the same row becomes an entry indicating topic information.

[0034] The topic 401 is a sentence indicating the topic of the conversation provided when the person specified by the person ID 301 is the facilitator. The keyword 402 is a word related to the topic 401. For example, the facilitator agent FA instructs the LLM 110 to "generate a sentence to propose the topic 401 (= How about ramen for lunch?) to the participant h in consideration of the keyword 402 (= price, distance, congestion level)." and generates a simulated utterance candidate sentence.

[0035] The action decision device 101 generates a facilitator agent FA that converts the topic information of the person ID 301 for which the facilitator flag 307 is "1" into input information (prompt) to the LLM 110 together with the person information.

[0036] <Dialogue Target Information Table> FIG. 5 is an explanatory diagram showing an example of an interaction target information table. The interaction target information table 500 is stored, for example, in the storage device 202 of the action determination device 101. The interaction target information table 500 has, as fields, an interaction target 501 and a score 502. A combination of the values of the fields in the same row becomes an entry indicating interaction target information.

[0037] The interaction target 501 is a word or sentence indicating a criterion for evaluation in a conversation. The score 502 is an evaluation value of the interaction target 501. For example, when the goal is to reach an agreement, the score 502 is set as follows. For example, when the interaction target 501 indicates agreement with an opinion, the value of the score 502 is set high, and when it indicates disagreement with an opinion, the value of the score 502 is set lower than when it indicates agreement. Also, when the interaction target 501 indicates a friendly utterance to another participant h, the value of the score 502 is set high, and when it indicates an aggressive utterance to another participant h, the value of the score 502 is set lower than when it indicates a friendly utterance.

[0038] <Search control information table> FIG. 6 is an explanatory diagram showing an example of a search control information table. The search control information table 600 is stored, for example, in the storage device 202 of the action determination device 101. The search control information table 600 has, as fields, a search control parameter 601 and a parameter value 602. A combination of the values of the fields in the same row becomes an entry indicating search control information.

[0039] The search control parameter 601 is a parameter for controlling a search tree. The search tree is generated by continuing a plurality of patterns of simulated conversations until the discussion converges through the simulated conversations between agents. The search control parameter 601 includes, for example, an upper limit of the number of searches and a temperature. The upper limit of the number of searches is the upper limit value of the number of searches of the search tree. The temperature is not an index indicating the degree of warmth or cold in the real environment, but an internal parameter for converging the simulated conversation.

[0040] The parameter value 602 is the value set as the search control parameter 601. Since the parameter value 602 for the maximum number of searches is "100", the search tree is used to perform 100 searches. Also, the parameter value 602 for temperature is "1.0". For example, at the start of a conversation, a predetermined temperature is set (hereinafter referred to as the start temperature), and each time a predetermined time elapses during the conversation, the temperature decreases by "1.0" from the previous temperature. When the updated temperature becomes equal to or lower than a predetermined threshold value, it is forcibly determined that the conversation has converged. If the start temperature is 40 degrees, the temperature decreases by the parameter value 602 of the temperature every time 1 minute elapses, and the threshold value is 10 degrees, it is forcibly determined that the conversation has converged in 30 minutes.

[0041] <Action decision processing procedure> FIG. 7 is a flowchart showing an example of an action decision processing procedure by the action decision device 101.

[0042] (Step S701) The action decision device 101 reads person information, topic information, dialogue target information, and search control information. That is, the action decision device 101 reads each entry of the person information table 300, the topic information table 400, the dialogue target information table 500, and the search control information table 600.

[0043] (Step S702) The action decision device 101 generates an agent for each person. Specifically, for example, the action decision device 101 generates input information (prompt) to the LLM 110 from the person information in which the facilitator flag 307 is "1" (hereinafter referred to as facilitator information) among the person information acquired in step S701, and inputs the generated prompt to the LLM 110, thereby generating an instance of an agent that simulates a facilitator as the facilitator agent FA.

[0044] Similarly, the action decision device 101 generates input information (prompt) to the LLM 110 from each piece of person information in which the participant flag 306 is "1" among the person information acquired in step S701 (hereinafter referred to as participant information), and inputs the generated prompt to the LLM 110, thereby generating an instance of an agent that simulates each participant h as a participant agent PA.

[0045] Thereby, a multi-agent system composed of the facilitator agent FA and the participant agent PA is constructed.

[0046] (Step S703) The action decision device 101 executes a speech information search process. The speech information search process (step S703) is a process of generating a root node in which the topic 401 with the same person ID 301 as the facilitator information is set to the current conversation state, growing a search tree from the root node in the simulated conversation between agents, and searching the grown search tree. Specifically, for example, the action decision device 101 executes a Monte Carlo tree search process of growing a Monte Carlo tree from the root node in the simulated conversation between agents and searching the grown Monte Carlo tree. The details of the speech information search process (step S703) will be described later with reference to FIG. 10. Although various configurations of Monte Carlo tree search are conceivable, FIG. 10 shows an example.

[0047] (Step S704) The action decision device 101 outputs the speech result sentence obtained in the speech information search process (step S703) to the terminal 102 of the participant h as the speech of the facilitator agent FA so that the participant h can audition it. That is, the action decision device 101 transmits the voice data of the speech result sentence to the terminal 102 or transmits the text data of the speech result to the terminal 102. Thereby, the terminal 102 outputs the speech result sentence as voice or displays the text data indicating the speech result sentence.

[0048] (Step S705) The action decision device 101 determines whether or not the simulated conversation meets the end condition. If it does not meet the end condition (step S705: No), it proceeds to step S706. If it meets the end condition (step S705: Yes), the action decision device 101 ends the action decision process.

[0049] The end condition is a condition for ending the simulated conversation. For example, when the preset end time has elapsed, the action decision device 101 determines that it meets the end condition (step S705: Yes). Also, when the simulated conversation meets the convergence condition, for example, when the evaluation value of the simulated conversation becomes equal to or greater than the threshold value, the action decision device 101 determines that the conversation has converged (for example, an agreement has been formed) and thus meets the end condition (step S705: Yes). Further, when no utterance sentence is input for a certain period of time in step S705 or when the input of utterance sentences after a certain period of time continues and is detected a predetermined number of times, the action decision device 101 may also determine that it meets the end condition (step S705: Yes) on the grounds that it is no longer possible to continue the simulated conversation.

[0050] (Step S706) The action decision device 101 receives the input of the utterance sentence of participant h. Specifically, for example, the terminal 102 converts the utterance of participant h who has viewed the utterance result sentence into an utterance sentence which is text data by voice recognition, or receives an utterance sentence which is text data by the operation input of that participant h, and transmits the utterance sentence to the action decision device 101. The action decision device 101 receives the utterance sentence transmitted from the terminal 102. Then, the action decision device 101 inputs the received utterance sentence to the facilitator agent FA to output its response sentence and sets it as the agreement support sentence.

[0051] Therefore, in the utterance information search process (step S703), the search tree 1100 is generated and searched using the agreement support sentence set in step S706 as the root node N0 indicating the current conversation state.

[0052] <Internal Model of Agent> Next, the internal model of the agent will be described. The agent is constructed based on the free energy principle. The free energy principle is a hypothesis of a unified explanatory principle for various cognitive functions of autonomous agents such as living organisms. According to the free energy principle, the agent internally holds a model of the environment (a group of participants h) (hereinafter, the internal model), and the agent's perception, learning, and action planning are performed by minimizing the free energy, which is a quantity representing the uncertainty of the model of the environment. That is, the free energy principle is a hypothesis that attempts to uniformly explain various cognitive processes of humans, including perception and action, by minimizing free energy. Therefore, the agent becomes an internal model that models the process by which a living organism subjectively predicts the future of the environment and acts based on that subjective prediction.

[0053] LLM110 is used as the internal model of the free energy principle. Mathematically, the free energy principle is formulated as variational Bayes inference, and the internal model is modeled as an instance of LLM110.

[0054] When the observed data (utterance in this example) from the environment is o and the latent state of the environment is s, the generative model of the environment is represented by P(s, o). The latent state corresponds to, for example, a set of internal states (emotions, cognitive states towards others) of the participant h or something called the atmosphere of the place. An agent following the free energy principle tries to approximate the internal model to the generative model P(s, o) of the environment. In variational Bayes inference, this problem is assumed to be the problem of finding the probability distribution Q θ (s) that minimizes the free energy F defined by the following equation (1).

[0055]

Equation

[0056] That is, the free energy F is determined by the generative model P(s, o) of the latent state s and the observed data o of the environment, and the internal model Q approximated by the parameter θ of the generative model P(s, o) of the latent state s of the environment.θ It is defined by formula (1) using (s) and the observed data o. KL [ ] is the Kullback–Leibler divergence, and E[ ] is the expected value. Thus, the free energy F can be calculated from the current probability distribution Q and the observed data o.

[0057] The following formula (2) shows the free energy F obtained by extending the free energy principle in formula (1) to a set including other participants h except for a specific participant h. we is shown.

[0058]

Number

[0059] In the above formula (2), F my is the free energy of a certain participant h, and F oth i is the free energy of other participants h. i is an integer of 1 or more and is an index for specifying other participants h.

[0060] In the above formula (2), the integrated free energy F is defined by the following formula (3) using the weight w indicating the sensitivity to others. we is defined by the following formula (3).

[0061]

Number

[0062] In the above formula (3), w my is the weight indicating the sensitivity of the free energy F my and w oth i is the weight indicating the sensitivity of the free energy F oth i If w my and w oth i are not distinguished, they are simply denoted as w. w is the free energy F my and the free energy F othi It is adjusted according to each change amount. For example, in the case of the same group (in-group) as oneself, an adjustment is made to strengthen w in the positive direction, and conversely, in the case of a different group (out-group) from oneself, an adjustment is made to strengthen w in the negative direction.

[0063] z my and z oth i are the inner states output from the encoder 811. The inner state z is a latent vector indicating the group perception of the group of participants including the participant h himself / herself. "z my ·z oth i " in the above formula (3) is the inner product of z my and z oth i and z. The smaller this inner product is, the more similar the opinions among the participants h are.

[0064] Also, in the above formula (3), when each of all the participants h is set as "a certain participant h", the integrated free energy F we-all is defined by the following formula (4).

[0065]

Equation

[0066] j is an integer of 1 or more and is an index for identifying each of the participants h. Returning to the explanation of the internal model.

[0067] [Internal Model of Facilitator Agent FA] FIG. 8 is an explanatory diagram showing an example of the internal model of the facilitator agent FA generated in step S702. The internal model 800 of the facilitator agent FA is an LLM 110 having an encoder 811 and a decoder 812. The facilitator information 801 is vectorized person information in which the facilitator flag 307 is "1". When the facilitator information 801 is input, the facilitator agent FA is generated as a facilitator-specific instance.

[0068] The internal model 800 has a probabilistic sampling method (Reparameterization Trick) used in the variational autoencoder (VAE) and stochastically samples the latent state s. That is, the internal model 800 is a model that stochastically outputs the future latent state s from the current latent state s. The same applies to the internal model 800 of the participant agent PA.

[0069] In the utterance information search process (step S703), the facilitator agent FA generates a prompt from the consensus support sentence (initially the topic 401), the participant information 802, the inner state zp of each participant agent PA, and the random information 803, and inputs it to the internal model 800. The consensus support sentence is a sentence that supports the formation of consensus among the participant agents PA. The inner state zp of the participant agent PA corresponds to the inner state z my and the inner state z oth i corresponds to.

[0070] The facilitator agent FA outputs its own inner state zx from the encoder 811. The facilitator agent FA inputs the inner state zx to the decoder 812 and outputs a simulated utterance candidate sentence. When not distinguishing between the inner states zp and zx, it is denoted as z. The inner state z is vector information indicating the inner self of the person simulated by the agent.

[0071] The inner state zp is used to identify the utterance partner to whom the facilitator agent FA speaks the simulated utterance candidate sentence, that is, the participating agent PA who is the input destination of the simulated utterance candidate sentence. For example, among the participant agents PA, the two participant agents PA with the longest vector distance between the inner states zp are set as the utterance partners who speak the simulated utterance candidate sentence. Thereby, it is possible to induce a simulated response sentence from the participating agent PA of the utterance partner so that the vector distance between the inner states zp is short.

[0072] In addition, the action decision device 101 may narrow down the speaker of the simulated utterance candidate sentence by referring to the words or keywords 402 within the topic 401. For example, when a certain simulated utterance candidate sentence contains "swimming", the action decision device 101 identifies "h2" from the topic information table 400 as the person ID 301 that includes "swimming" in the keyword 402. Also, the action decision device 101 identifies "h1", "h3", and "h4" from the topic information table 400 as the person ID 301 that does not include "swimming" in the keyword 402.

[0073] The action decision device 101 sets, as the speaker of the simulated utterance candidate sentence, the combination of participant agents PA (for example, participant agents PA2 and PA4) that results in the longest vector distance among the vector distances between the internal state zp2 output by the participant agent PA2 of participant h2, the internal state zp1 output by the participant agent PA1 of participant h1, the vector distance between the internal state zp2 and the internal state zp3 output by the participant agent PA3 of participant h3, and the vector distance between the internal state zp2 and the internal state zp4 output by the participant agent PA4 of participant h4.

[0074] Thereby, it is possible to efficiently induce a simulated response sentence from the participant agent PA of the speaker so that the vector distance of the internal state zp becomes shorter.

[0075] The participant information 802 is the personal information of the participant agent PA who is the speaker of the simulated utterance candidate sentence. One or more combinations of speakers of the simulated utterance candidate sentence can be set. The facilitator agent FA generates a simulated utterance candidate sentence for each combination of speakers of the simulated utterance candidate sentence.

[0076] Note that the internal state zx may be included in the internal state zp. In this case, the facilitator agent FA sets, as the speaker of the simulated utterance candidate sentence, the participant agent PA that output the internal state zp with the longest distance from the internal state zx.

[0077] The random information 803 is a parameter for controlling the randomness of speech. Specifically, for example, the random information 803 is a parameter that gives a random adjustment to the generation of the consensus support sentence by the facilitator agent FA. For example, there are "temperature" and "top P". For example, in the case of "temperature", the random information 803 takes a value in the range of 0.0 to 2.0. The smaller the value of the random information 803, the smaller the randomness of the simulated speech candidate sentence output from the decoder 812.

[0078] The larger the value of the random information 803, the greater the randomness of the simulated speech candidate sentence output from the decoder 812, and the greater the change in the expression of the consensus support sentence input to the encoder 811. The value of the random information 803 can be set to one or more, and is set before the start of the matters in the speech information search process (step S703). The facilitator agent FA generates simulated speech candidate sentences for each value of the random information 803.

[0079] Also, by setting a plurality of values of the random information 803 for each combination of speech partners who speak the simulated speech candidate sentence, the facilitator agent FA generates simulated speech candidate sentences for each combination of speech partners who speak the simulated speech candidate sentence for each value of the random information 803. Also, not only the temperature, but also the simulated speech candidate sentence is generated according to a predetermined generation rule. For example, it is generated according to the generation rule of "the opinion that all members most agree with for the combination of speech partners".

[0080] [Internal model of participant agent PA] FIG. 9 is an explanatory diagram showing an example of the internal model of the participant agent PA generated in step S702. The internal model 900 of the participant agent PA is an LLM110 having an encoder 811 and a decoder 812, similar to the internal model 800 of the facilitator agent FA. The participant information 802 is vectorized person information with the participant flag 306 being "1". By inputting the participant information 802, the participant agent PA is generated as an instance unique to the participant h.

[0081] In the speech information search process (step S703), the participant agent PA inputs the simulated speech candidate sentence from the facilitator agent FA. The participant agent PA outputs its internal state zp from the encoder 811. The participant agent PA inputs the internal state zp to the decoder 812 and outputs a simulated response sentence for the simulated speech candidate sentence.

[0082] When the agent is executed using the LLM110 of the action decision device 101, the action decision device 101 acquires the internal state z from the encoder 811 of the agent. When the agent is executed using the LLM110 of the external device 104, the action decision device 101 requests the internal state z from each agent of the external device 104, and the external device 104 acquires the internal state z output from the encoder 811 of each agent via the network 103 in response to the request.

[0083] Unless retraining is performed for each agent, the internal models 800, 900 (LLM110) themselves are actually shared by all agents.

[0084] <Speech Information Search Process (Step S703)> FIG. 10 is a flowchart showing a detailed processing procedure example of the speech information search process (step S703). FIG. 11 is an explanatory diagram showing an example of a search tree.

[0085] (Step S1001) The action decision device 101 initializes the search count m to m = 1 and initializes the level k of the search tree 1100 to be generated and searched to k = 1. m is an integer of 1 or more, and k is an integer of 0 or more.

[0086] (Step S1002) The action decision device 101 generates a node N0 indicating the conversation state at layer k = 0. When layer k = 0, the conversation state indicated by node N0 is the current conversation state. Node N0 is referred to as the root node of the search tree 1100. In the root node N0, the topic "How about having ramen for lunch?" of the facilitator (person ID 301 = hx) obtained from the topic information table 400 in step S701 is set as the consensus support sentence.

[0087] (Step S1003) In the currently generated search tree 1100, the action decision device 101 follows the link Lk from the root node N0 to the node Nk with the highest evaluation value Vk at each layer k until the terminal node Nk (leaf), and updates based on the layer of the leaf reached at the current layer k.

[0088] For example, in the initial state, since there is only the root node N0, the root node N0 itself becomes the reached leaf, the layer k = 0 is incremented to layer k = 1, and the process proceeds to step S1004.

[0089] Also, for example, as shown in the search tree 1100 of FIG. 11, when node N4c is not generated and all other nodes are already generated, assuming that the node N3f is reached as a result of following the node with the highest evaluation value at each layer. In this case, node N3f is the terminal node and its layer k is k = 3, so the layer k = 3 is incremented and the layer k is updated to k = 4. Also in this case, in step S1004, the link L4c of the updated layer k = 4 is generated, and in step S1005, the node N4c of the updated layer k = 4 is generated.

[0090] In addition, the number of visits is set for all nodes Nk of the search tree 1100 currently being generated. The initial value of the number of visits is 0. The number of visits of the nodes Nk on the path to the leaf node Nk via the node Nk with the highest evaluation value Vk increases by 1. The number of visits of the node Nk is used for correcting the evaluation value Vk. Note that the direction in which the hierarchy k increases is the direction in which the links Lk and the nodes Nk are generated and the search tree 1100 grows. That is, it is the direction in which the pseudo-conversation between the facilitator agent FA (simulated utterance candidate sentence) and the participant agent PA (simulated response sentence to the simulated utterance candidate sentence) progresses.

[0091] (Step S1004: Generation of Link Lk) The action determination device 101 sets the leaf node traced in step S1003 as the node Nk, and generates one or more links Lk connected to the node Nk. Taking the case of hierarchy k = 1 as an example, the action determination device 101 selects the interlocutor of the simulated utterance candidate sentence of the facilitator agent FA from among the group of participant agents PA.

[0092] At hierarchy k = 1, each participant agent PA has not yet calculated the internal state zp. Therefore, the action determination device 101 determines the interlocutor of the simulated utterance candidate sentence as one or more combinations of participant agents PA set in advance (for example, all participant agents PA1 to PAn). Here, for example, the values of the random information 803 are four values of "0.1", "0.5", "1.0", and "2.0", and the number of combinations of interlocutors of the simulated utterance candidate sentence is one of "all participant agents PA1 to PA4".

[0093] The number of links L(k + 1) output from a certain node Nk is the number of values of the random information 803 × the number of combinations of interlocutors of the simulated utterance candidate sentence. For example, in the case of the link L1, when the values of the random information 803 are four values of "0.1", "0.5", "1.0", and "2.0", and the number of combinations of interlocutors of the simulated utterance candidate sentence is one of "all participant agents PA1 to PA4", the number of the link L1 is four (hereinafter referred to as links L1a, L1b, L1c, and L1d).

[0094] The action decision device 101 sets the random information 803 and the combination of the interlocutors of the simulated utterance candidate sentences for the facilitator agent FA. When the action decision device 101 inputs "How about ramen for lunch?" of the topic 401 to the facilitator agent FA, the facilitator agent FA generates simulated utterance candidate sentences for each of the links L1a, L1b, L1c, and L1d to the "all participant agents PA1 to PA4".

[0095] For example, the link L1a indicates a simulated utterance candidate sentence (for example, the topic 401 as it is) based on the value "0.0" of the random information 803. The link L1b indicates a simulated utterance candidate sentence (for example, "How about noodles for lunch?") based on the value "0.5" of the random information 803. The link L1c indicates a simulated utterance candidate sentence (for example, "What instant noodles do you like?") based on the value "1.0" of the random information 803. The link L1d indicates a simulated utterance candidate sentence (for example, "What's your favorite food?") based on the value "2.0" of the random information 803.

[0096] Also, when the hierarchy k ≥ 2, the participant agent PA has the latest internal state zp. In this case, as described above, for each simulated utterance candidate sentence, the action decision device 101 determines the interlocutor of the simulated utterance candidate sentence based on the internal state zp.

[0097] In the following example, as shown in FIG. 11, the case where the links L2a to L2h are generated at the hierarchy k = 2, the links L3a to L3g are generated at the hierarchy k = 3, and the links L4a to L4c are generated at the hierarchy k = 4 will be described. Note that the link Lk is not generated under certain conditions, for example, when there is no participant agent PA that becomes a simulated utterance candidate sentence, or when the evaluation value Vk of the node exceeds a threshold value.

[0098] (Step S1005: Generation of node Nk) The action decision device 101 generates a node Nk indicating the conversation state at layer k for the leaf node traced in step S1003. The number of visits to the node Nk at the time of generation is 0. Specifically, for example, the action decision device 101 inputs the simulated utterance candidate sentence from the facilitator agent FA to the participant agent PA of the speaker of the simulated utterance candidate sentence. This participant agent PA generates a simulated response sentence for the simulated utterance candidate sentence. Taking the link L1 as an example, the action decision device 101 inputs the simulated utterance candidate sentences of each link L1a~L1d to the participant agent PA of the speaker of the simulated utterance candidate sentence.

[0099] For example, in the above example, the speakers of the simulated utterance candidate sentences are "all participant agents PA1~PA4". Therefore, the action decision device 101 inputs the simulated utterance candidate sentence of link L1a to each of the participant agents PA1~PA4, inputs the simulated utterance candidate sentence of link L1b to each of the participant agents PA1~PA4, inputs the simulated utterance candidate sentence of link L1c to each of the participant agents PA1~PA4, and inputs the simulated utterance candidate sentence of link L1d to each of the participant agents PA1~PA4.

[0100] The action decision device 101 acquires the simulated response sentence and the internal state zp output from the participant agent PA. For example, in the above example, in response to the simulated utterance candidate sentence of link L1a, each of the participant agents PA1~PA4 generates a simulated response sentence and acquires the internal states zp1~zp4 output from the encoders 811 of each of the participant agents PA1~PA4. These four simulated response sentences and the internal states zp1~zp4 become the node N1a connected to the link L1a.

[0101] Similarly for the simulated utterance candidate sentences of links L1b~L1d, each of the participant agents PA1~PA4 generates a simulated response sentence and acquires the internal states zp1~zp4 output from the encoders 811 of each of the participant agents PA1~PA4. Thereby, the nodes N1b~N1d connected to the links L1b~L1d are generated.

[0102] As shown in FIG. 11, at layer k = 2, nodes N2a to N2h connected to links L2a to L2h are generated, at layer k = 3, nodes N3a to N3g connected to links L3a to L3g are generated, and at layer k = 4, nodes N4a to N4c connected to links L4a to L4c are generated.

[0103] In the search tree 1100, the paths from the root node N0 indicating the current conversation state to each of the terminal nodes N2a, N2b, N3a, N3b, N3c, N2e, N4a, N4b, N3e, N4c, N3g are referred to as simulated conversation paths.

[0104] (Step S1006: Update of the inner state zp) The action decision device 101 updates the inner state z of layer k - 1 to the inner state z of layer k. When k = 1, since there is no inner state z of layer k = 0, the inner state z of layer k = 1 is set. Specifically, for example, at layer k = 1, the action decision device 101 sets the inner states zp1 to zp4 generated by link L1a to the newly generated node N1a of layer k = 1. Similarly for nodes N1b to N1d, the action decision device 101 sets the inner states zp1 to zp4 generated by each of links L1b to L1d.

[0105] As shown in FIG. 11, at layer k = 2, the inner states zp1 to zp4 of each of nodes N2a to N2h are updated, at layer k = 3, the inner states zp1 to zp4 of each of nodes N3a to N3g connected to links L3a to L3g are updated, and at layer k = 4, the inner states zp1 to zp4 of each of nodes N4a to N4c connected to links L4a to L4c are updated.

[0106] (Step S1007: Calculation of the evaluation value Vk) The action decision device 101 calculates the evaluation value Vk of the node Nk. The evaluation value Vk is an index indicating whether the conversation state indicated by the node Nk is heading towards consensus formation. The higher the value, the more the conversation state is heading towards consensus formation. Specifically, for example, at layer k = 1, the action decision device 101 calculates the evaluation values V1a to V1d for each of the nodes N1a to N1d. At layer k = 2, the action decision device 101 calculates the evaluation values V2a to V2h for each of the nodes N2a to N2h. At layer k = 3, the action decision device 101 calculates the evaluation values V3a to V3g for each of the nodes N3a to N3g. At layer k = 4, the action decision device 101 calculates the evaluation values V4a to V4c for each of the nodes N4a to N4c. The calculation of the evaluation value Vk will be described later.

[0107] (Step S1008) The action decision device 101 determines whether the search count m has reached the upper limit search count M. If the search count m has not reached the upper limit search count M (Step S1008: No), it proceeds to Step S1009. If the search count m has reached the upper limit search count M (Step S1008: Yes), it proceeds to Step S1010.

[0108] (Step S1009) The action decision device 101 increments the search count m and returns to Step S1003.

[0109] (Step S1010) The action decision device 101 selects the node Nmax with the maximum evaluation value Vmax from the search tree 1100. In the example of FIG. 11, assume that the evaluation value V3d of the node N3d is the maximum evaluation value Vmax. Note that the action decision device 101 is not limited to selecting the node Nmax with the maximum evaluation value Vmax from the search tree 1100. It may also select the top n nodes Nk (n is an integer of 1 or more) in descending order of the evaluation value Vk. Further, the action decision device 101 may select the nodes Nk whose evaluation value Vk is greater than or equal to the threshold value. In any case, if a plurality of nodes Nk are selected, the action decision device 101 selects any one of the selected plurality of nodes Nk.

[0110] (Step S1011) The action determination device 101 identifies the link L1 for transitioning from the root node N0 to the node N1 in the path Pmax from the root node N0 to the node Nmax with the maximum evaluation value Vmax. In the case of the search tree 1100 in FIG. 11, the path Pmax is the path passing through the root node N0, the link L1c, the node N1c, the link L2f, the node N2f, the link L3d, and the node N3d. Among these, the link L1 for transitioning from the root node N0 to the node N1c is the link L1c.

[0111] (Step S1012) The action determination device 101 extracts the simulated utterance candidate sentence of the link L1c identified in Step S1011 as the utterance result sentence and proceeds to Step S704. As a result, the action determination device 101 can output the simulated utterance candidate sentence of the link L1c extracted in Step S1012 to the terminal 102 so that it can be viewed as the utterance result sentence (Step S704). Specifically, for example, the terminal 102 can display the simulated utterance candidate sentence on the display screen or read out the simulated utterance candidate sentence and output it as voice. The terminal 102 performs at least one of display and voice output.

[0112] After this, as described above, the action determination device 101 receives the input of the utterance sentences from the participants h1 to h4 who have viewed the utterance result sentence from the terminal 102 (Step S705) and re-executes the utterance information search process (Step S703). In this case, the facilitator agent FA inputs the utterance sentences from the participants h1 to h4 and outputs a response sentence for supporting consensus formation as the consensus support sentence. The action determination device 101 generates this response sentence instead of the topic 401 as the root node N0 indicating the current conversation state at the hierarchy k = 0 (Step S1002).

[0113] As a result, the action decision device 101 constructs multi-agents (facilitator agent FA and participant agent PA), and using the virtual simulated utterance candidate sentences generated from the participant agent PA, realizes the determination of the active utterance result sentences of the facilitator agent FA that supports consensus formation.

[0114] <Calculation of evaluation value Vk (step S1007)> Next, the calculation of the evaluation value Vk (step S1007) will be specifically described. The action decision device 101 uses the integrated free energy F of the above-described formula (4) we-all and the dialogue target information table 500 to calculate the evaluation value Vk. Note that the participant agent PA for which the evaluation value Vk is calculated may assume all participants h, or may be limited to the participant agent PA that is the interlocutor of the simulated utterance candidate sentence specified by the node Nk. Also, the subsequent evaluation value Vk is corrected according to the number of visits to the node. For example, when the number of visits to the node is n, it is corrected by formula (5).

[0115] [Number]

[0116] [When using the integrated free energy F we-all When the LLM110 is implemented in the external device 104, for example, the action decision device 101 calculates the evaluation value Vk for each node Nk (excluding the root node N0) based on the integrated free energy F we-all . Since the smaller the value of the integrated free energy F we-all , the smaller the prediction uncertainty, the evaluation value Vk becomes higher.

[0117] In this way, the action decision device 101 can promote conversations with less prediction uncertainty by using the integrated free energy F we-all .

[0118] [When using the internal state zp between participant agents PA]​ The action decision device 101 may calculate the evaluation value Vk using the internal state zp among the participant agents PA without using the integrated free energy F. we-all

[0119] FIG. 12 is an explanatory diagram showing an example graph of the internal state zp. In FIG. 12, for convenience of explanation, the internal states zp1 to zp3 of the participant agents PA1 to PA3 are shown in the vector space 1200. In this example, the smaller the sum of the vector distances d12 between the vectors of the internal states zp1 and zp2, the vector distance d23 between the vectors of the internal states zp2 and zp3, and the vector distance d13 between the vectors of the internal states zp1 and zp3, the closer the internal states of each participant h are, and the higher the possibility of reaching an agreement is considered. Therefore, the action decision device 101 increases the evaluation value Vk as the sum of the vector distances decreases. For example, the action decision device 101 uses the reciprocal of the sum of the vector distances (a positive value may be added so that the denominator is always a value greater than 0) as the evaluation value Vk.

[0120] FIG. 13 is an explanatory diagram showing an example graph of the internal state z. In FIG. 13, for convenience of explanation, the internal states zp1 and zp2 of the participant agents PA1 and PA2 are shown in the vector space 1200. The internal state zp1(k - 1) is the internal state zp1 output from the participant agent PA1 at the node N(k - 1) in the layer k - 1 of the search tree 1100, and the internal state zp1(k) is the internal state zp1 output from the participant agent PA1 at the node N(k) in the layer k of the search tree 1100. The same applies to the internal states zp2(k - 1) and zp2(k).

[0121] The change amounts regarding the progress direction of the simulated conversation from layer k - 1 to layer k are represented by d1(k) and d2(k). The change amount d1(k) is the vector distance between the internal state zp1(k - 1) and the internal state zp1(k), and the change amount d2(k) is the vector distance between the internal state zp2(k - 1) and the internal state zp2(k). That is, the change amounts d1(k) and d2(k) are the vector distances of the internal state z before and after the update in step S1006.

[0122] ​ The larger the sum of the change amounts d1(k) and d2(k), the more the inner state changes, and it is considered that there is a high possibility that the discussion is activated. Therefore, the action determination device 101 increases the evaluation value Vk as the sum of the inter-vector distances increases. For example, the action determination device 101 sets the sum of the inter-vector distances as the evaluation value Vk.

[0123] In this way, the action determination device 101 can support consensus formation according to the uncertainty of prediction, the proximity of the inner states among the participants h, or the activity level of the discussion state by using the inner state z.

[0124] Note that in the determination of the speaker of the simulated utterance candidate sentence and the calculation of the evaluation value Vk, the inner state zp of the participant agent PA was applied, but the inner state zx of the facilitator agent FA may also be included. For example, in the determination of the speaker of the simulated utterance candidate sentence, the action determination device 101 may determine the participant agent PA with the longest vector distance between the inner state zx and the inner state zp as the speaker of the simulated utterance candidate sentence. Thereby, considering the position of the facilitator agent FA, the determination of the speaker of the simulated utterance candidate sentence and the calculation of the evaluation value Vk can be executed, and consensus formation becomes more efficient compared to the case where the inner state zx of the facilitator agent FA is not included.

[0125] Also, in both the case of using the above-described integrated free energy F we-all and the case of using the inner state z among all the participant agents PA, the action determination device 101 may adjust the evaluation value Vk by using the dialogue target information table 500.

[0126] For example, the action determination device 101 vectorizes the simulated response sentence of the participant agent PA of the node Nk and the dialogue target 501 by a vectorization method such as doc2vec, for example, and calculates the inter-vector distance. The action determination device 101 calculates a weighted linear sum obtained by multiplying each inter-vector distance by the score 502 corresponding to the dialogue target 501, and adds the weighted linear sum to the evaluation value Vk.

[0127] In this way, by adding the weighted linear sum using the vector distance from the dialogue target 501 to the evaluation value Vk, it is possible to promote a conversation with less prediction uncertainty so as to approach the dialogue target 501.

[0128] <Example of screen display> Next, an example of the screen display of the utterance result sentence of the facilitator agent FA output in the utterance information search process (step S703) will be described.

[0129] FIG. 14 is an explanatory diagram showing an example 1 of the screen display of the utterance result sentence of the facilitator agent FA. The display device 1400 is an example of the output device 204 of the terminal 102. The display device 1400 displays, on its display screen 1401, an avatar 1402 of a facilitator virtually existing by the facilitator agent FA and an utterance result sentence 1403 of the facilitator agent FA output in the utterance information search process (step S703). The action determination device 101 may output the utterance result sentence 1403 as voice from a speaker which is an example of the output device 204.

[0130] FIG. 15 is an explanatory diagram showing an example 2 of the screen display of the utterance result sentence of the facilitator agent FA. The display device 1400 displays a bulletin board 1500 on its display screen 1401. On the bulletin board 1500, an utterance result sentence 1501 of the facilitator agent FA output in the utterance information search process (step S703), an utterance sentence 1502 of the participant h in response to the utterance result sentence 1501 in step S705, and an utterance result sentence 1503 of the facilitator agent FA output again in the utterance information search process (step S703) in response to the utterance sentence 1502 are displayed in chronological order. The action determination device 101 may output the utterance result sentence 1501, the utterance sentence 1502, and the utterance result sentence 1503 as voice from a speaker which is an example of the output device 204.

[0131] As described above, according to the first embodiment, in a conversation without a facilitator, the facilitator agent FA can determine, without the need for learning using facilitation action data collected at the time of execution, the utterances that the facilitator should make to promote the conversation in a conversation in which the participant h participates, and can utter to the participant h. Therefore, the efficiency of consensus formation can be improved.

Embodiment

[0132] The second embodiment will be described. In the first embodiment described above, the system configuration in which each participant h uses his or her own terminal 102 is taken as an example. In the second embodiment, the case where the terminal 102 does not exist will be described as an example. In the second embodiment, in order to mainly describe the differences from the first embodiment, the same reference numerals are given to the parts common to the first embodiment, and the description thereof is omitted.

[0133] FIG. 16 is an explanatory diagram showing a second example of the system configuration of the action determination system. The action determination system 100 includes an action determination device 101, a microphone 1601, and a display device 1602. The microphone 1601 and the display device 1602 are connected to the action determination device 101.

[0134] The microphone 1601 is an example of the input device 203 of the action determination device 101. The microphone 1601 is disposed, for example, on a table 1603 surrounded by the participants h, inputs the speech voice from the participants h, and outputs the voice data to the action determination device 101. The display device 1602 is an example of the output device 204 of the action determination device 101. The display device 1602 performs the same display as the display device 1400 of the terminal 102 shown in FIGS. 14 and 15. Further, as described with reference to FIGS. 14 and 15, the action determination device 101 may output the speech of the display content as voice.

[0135] The action decision-making device 101 holds in advance the sample voice data of the participant h, and based on the voice data from the microphone 1601 and the sample voice data by existing voice recognition, recognizes which participant h is speaking. Also, the microphone 1601 may be prepared for each participant h. In this case, since each participant h corresponds to each microphone, it becomes easy to recognize which participant h is speaking.

[0136] Thus, also in the system configuration shown in FIG. 16, similar to Example 1, in a conversation without a facilitator, the facilitator agent FA can actively determine the utterances that the facilitator should make for promoting the conversation in the conversation in which the participant h participates, and can speak to the participant h. Therefore, the efficiency of consensus formation can be improved.

Example

[0137] Example 3 will be described. In the above-described Examples 1 and 2, the case where there is one facilitator agent FA for all the participant agents PA was taken as an example, but in Example 3, the facilitator agent FA is implemented for each participant agent PA. In Example 3, in order to mainly explain the differences from Examples 1 and 2, the parts common to Examples 1 and 2 are given the same reference numerals and their explanations are omitted.

[0138] FIG. 17 is an explanatory diagram showing a system configuration example 3 of the action decision-making system 100. In the action decision-making device 101, facilitator agents FA1 to FA4 are respectively implemented for each of the participant agents PA1 to PA4. In this case, the action decision-making device 101 executes the processes shown in FIGS. 7 and 10 for each of the facilitator agents FA1 to FA4.

[0139] That is, the action decision device 101 repeatedly executes a speech information search process (step S703) between the facilitator agent FA1 and the participant agent PA1, outputs the speech result sentence so that the participant h1 can view it, and inputs the speech sentence from the participant h1 (step S705). Similarly, the action decision device 101 repeatedly executes a speech information search process (step S703) between the facilitator agent FA2 and the participant agent PA2, outputs the speech result sentence so that the participant h2 can view it, and inputs the speech sentence from the participant h2 (step S705). The same applies to between the facilitator agent FA3 and the participant agent PA3, and between the facilitator agent FA4 and the participant agent PA4.

[0140] Thus, in the third embodiment, since the facilitator agent FA is implemented for each participant agent PA, each facilitator agent FA can actively make a speech that promotes the speech of the participant h.

[0141] Also, although one facilitator agent FA is implemented for one participant agent PA, one facilitator agent FA may be implemented for a plurality of participant agents PA. For example, the facilitator agent FA1 may be implemented for the participant agents PA1 to PA3, and the facilitator agent FA4 may be implemented for the participant agent PA4.

[0142] Also, in FIG. 17, the system configuration of FIG. 1 in the first embodiment is taken as an example for explanation, but the same applies to the system configuration of FIG. 16 in the second embodiment.

[0143] Note that the present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the scope of the appended claims. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and the present invention is not necessarily limited to those having all the configurations described. Also, a part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Further, the configuration of another embodiment may be added to the configuration of one embodiment. Also, for a part of the configuration of each embodiment, addition, deletion, or replacement with other configurations may be made.

[0144] In addition, each of the above-described configurations, functions, processing units, processing means, etc. may be realized in hardware, for example, by designing part or all of them with an integrated circuit, or may be realized in software by a processor interpreting and executing a program for realizing each function.

[0145] Information such as programs, tables, files, etc. for realizing each function can be stored in a storage device such as a memory, a hard disk, an SSD (Solid State Drive), or a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0146] Also, the control lines and information lines shown are those considered necessary for explanation, and do not necessarily show all the control lines and information lines necessary for implementation. In practice, it may be considered that almost all configurations are interconnected.

Explanation of Reference Numerals

[0147] 100 Action Decision System 101 Action Decision Device 102 Terminal 103 Network 104 External Device 110 LLM 201 Processor 202 Storage Device 203 Input Device 204 Output device 300 Person information table 400 Topic information table 500 Dialogue target information table 600 Exploration control information table 800 Internal model 801 Facilitator information 802 Participant information 900 Internal model 1100 Exploration tree Nk Node Lk Link

Claims

1. An action decision device having a processor that executes a program and a storage device that stores the program, wherein the processor performs a search process of searching for a plurality of simulated conversation routes from the start to the end of the simulated conversation by executing a simulated conversation between a first agent that simulates a facilitator who supports consensus formation in a conversation within a group of participants based on a language model and a second agent that simulates a participant who participates in the conversation based on the language model; a calculation process of calculating an evaluation value for evaluating the utterance of a simulated response sentence in the simulated conversation generated from the second agent in the plurality of simulated conversation routes searched by the search process based on an internal state indicating the group awareness of the participant in the group of participants in the second agent; an extraction process of extracting a specific simulated utterance candidate sentence from a plurality of simulated utterance candidate sentences from the first agent to the second agent at the start point of the simulated conversation based on the evaluation value calculated by the calculation process; an output process of outputting the specific simulated utterance candidate sentence extracted by the extraction process; and executes the above, characterized in that it is an action decision device.

2. The action decision device according to claim 1, wherein in the search process, the processor repeatedly executes a first generation of inputting a consensus support sentence that supports the consensus formation of the conversation as the simulated conversation to the first agent to generate a plurality of simulated utterance candidate sentences, and a second generation of inputting the simulated utterance candidate sentences to the second agent to generate a simulated response sentence to the simulated utterance candidate sentences as the consensus support sentence. characterized in that it is an action decision device.

3. The action decision device according to claim 2, wherein the storage device stores the language model and person information indicating the characteristics of a person, the processor executes a generation process of generating the first agent and the second agent based on the language model and the person information, and in the search process, the processor searches for the plurality of simulated conversation routes by repeatedly executing the first generation and the second generation as the simulated conversation between the first agent and the second agent generated by the generation process. characterized in that it is an action decision device.

4. The action decision device according to claim 2, wherein It is possible to access an external device that stores the language model capable of communicating with the action decision-making device, The storage device stores person information indicating the characteristics of a person, The processor, Sends a generation request for the first agent and the second agent including the person information to the external device, and executes an agent generation process for causing the external device to generate the first agent and the second agent based on the person information, In the search process, the processor repeats the first generation and the second generation as the simulated conversation between the first agent and the second agent generated by the agent generation process, thereby searching for the plurality of simulated conversation paths. An action decision-making device characterized by the above.

5. The action decision-making device according to claim 2, In the search process, the processor determines the second agent that is the input destination of the simulated utterance candidate sentence based on the vector distance between the internal state vectors of the second agents, and inputs the simulated utterance candidate sentence to the determined second agent to generate the simulated response sentence as the consensus support sentence. An action decision-making device characterized by the above.

6. The action decision-making device according to claim 2, In the search process, the processor determines the second agent that is the input destination of the simulated utterance candidate sentence based on the vector distance between the internal state of the first agent and the internal state of the second agent, and inputs the simulated utterance candidate sentence to the determined second agent to generate the simulated response sentence as the consensus support sentence. An action decision-making device characterized by the above.

7. The action decision-making device according to claim 1, The processor, Calculates the evaluation value based on the integrated free energy obtained by integrating the free energy related to the collective cognition of the participants in the free energy principle. An action decision-making device characterized by the above.

8. The action decision-making device according to claim 1, In the calculation process, the processor calculates the evaluation value based on the sum of the vector distances between the internal states of the second agents. An action decision-making device characterized by the above.

9. The action decision-making device according to claim 1, In the calculation process, the processor calculates the evaluation value based on the sum of the vector distances of the inner surface states in the progress direction of the simulated conversation for each of the second agents. An action determination device characterized by the above.

10. The action determination device according to claim 2, wherein the storage device stores dialogue target information in which a dialogue target and a score for the dialogue target are associated with each other. In the calculation process, the processor adjusts the evaluation value based on the vector distance between the dialogue target and the simulated response sentence generated by the second generation, and the score. An action determination device characterized by the above.

11. The action determination device according to claim 2, wherein the processor performs setting processing of inputting a speech sentence from the participant to the first agent as a result of outputting the specific simulated speech candidate sentence by the output processing, and setting the sentence output from the first agent as the consensus support sentence in the first generation. An action determination device characterized by executing the above.

12. An action determination method executed by an action determination device having a processor that executes a program and a storage device that stores the program, wherein the processor executes a search process of searching for a plurality of simulated conversation routes from the start to the end of the simulated conversation by executing a simulated conversation between a first agent that simulates a facilitator who supports consensus formation in a conversation within a participant group based on a language model and a second agent that simulates a participant who participates in the conversation based on the language model. a calculation process of calculating an evaluation value for evaluating the utterance of the simulated response sentence in the simulated conversation generated by the second agent in the plurality of simulated conversation routes searched by the search process based on the inner surface state indicating the group awareness of the participant in the participant group in the second agent; an extraction process of extracting a specific simulated speech candidate sentence from a plurality of simulated speech candidate sentences from the first agent to the second agent at the start point of the simulated conversation based on the evaluation value calculated by the calculation process; an output process of outputting the specific simulated speech candidate sentence extracted by the extraction process. An action determination method characterized by executing the above.

13. To the processor A first agent that simulates a facilitator who supports consensus formation in conversations within a group of participants based on a language model, and a second agent that simulates a participant who participates in the conversation based on the language model, repeatedly execute a simulated conversation between them to perform a search process for exploring multiple simulated conversation paths from the start to the end of the simulated conversation. Based on the internal state indicating the group awareness of the participant in the second agent with respect to the group of participants, calculate an evaluation value for evaluating the utterance of the simulated response sentence in the simulated conversation generated from the second agent in the multiple simulated conversation paths explored by the search process. Based on the evaluation value calculated by the calculation process, perform an extraction process of extracting a specific simulated utterance candidate sentence from multiple simulated utterance candidate sentences from the first agent to the second agent at the start point of the simulated conversation. Perform an output process of outputting the specific simulated utterance candidate sentence extracted by the extraction process. A behavior decision program characterized by causing the above to be executed.

Images