Information processing systems, programs, and information processing methods

Abstract

This invention provides an information processing system, program, and method that can adapt the responses of an interactive artificial intelligence system to user input. [Solution] The system comprises: an input acquisition unit 11 that acquires user input directed to an interactive artificial intelligence system; a feature calculation unit 13 that calculates a plurality of feature quantities including the level of abstraction of the user input and the degree of inference required for the user input based on the acquired user input; a distribution generation unit 14 that generates a cognitive depth probability distribution across multiple depth levels based on the calculated plurality of feature quantities; a depth determination unit 15 that determines the response depth corresponding to the user input based on the generated cognitive depth probability distribution; a control information generation unit 16 that generates control information, which is information for causing the interactive artificial intelligence system to generate a natural language response that is suitable for the user input, based on the determined response depth; and an output unit 17 that outputs the user input and control information to the interactive artificial intelligence system.

Description

Technical Field

[0001] The present invention relates to an information processing system, a program, and an information processing method for adapting a response from an interactive artificial intelligence system to a user input.

Background Art

[0002] In recent years, interactive artificial intelligence systems, particularly dialogue services using large-scale language models, have been widely used. Such an interactive artificial intelligence system can generate a natural language response corresponding to the content of an input natural language sentence input by a user.

[0003] Interactive artificial intelligence systems are used for various applications such as question answering, document creation support, learning support, business support, and consultation support. In addition, users provide inputs of various natures to the interactive artificial intelligence system, such as simple fact checks, specific work instructions, situation descriptions, emotional expressions, abstract consultations, and queries involving value judgments.

[0004] Regarding such interactive artificial intelligence systems, large-scale language models that generate natural language responses in response to prompts given by users have been disclosed (for example, see Non-Patent Document 1). Furthermore, techniques have also been disclosed that generate responses involving reasoning such as arithmetic reasoning, commonsense reasoning, or symbolic reasoning by providing examples including reasoning processes to large-scale language models (for example, see Non-Patent Document 2).

[0005] In general, an interactive artificial intelligence system generates a response based on words, context, or topics included in a user input. However, the interactive artificial intelligence system may generate a uniform response without sufficiently distinguishing the depth, hierarchy, explanation granularity, or degree of reasoning of the response required by the user input.

[0006] For example, an interactive artificial intelligence system may respond with the same response format, level of explanation, or degree of reasoning whether the user input is merely fact-checking or involves consultation with emotions or value judgments. In this case, the response may correspond to the superficial content of the user input, but may not adequately address the level of cognitive depth required by that user input. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] OpenAI, “GPT-4 Technical Report”, p.1 Abstract, p.1 Introduction, arXiv:2303.08774, March 27, 2023, OpenAI, [online], [Retrieved June 8, 2026], Internet<URL: https: / / cdn.openai.com / papers / gpt-4.pdf> [Non-Patent Document 2] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou, “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”, p.1 Abstract, p.2 Section 1, p.3 Section 2, 36th Conference on Neural Information Processing Systems (NeurIPS 2022), 2022, Google Research, Brain Team, [online], [Retrieved June 8, 2026], Internet <url: https: openreview.net pdf?id="_VjQlMeSB_J"> [Overview of the project] [Problems that the invention aims to solve]

[0008] However, the required level of cognitive depth for user input varies depending on the content of the input. For example, if the user input is a confirmation of specific facts, a concise and specific response may be appropriate. On the other hand, if the user input includes emotions, situations, background circumstances, causal relationships, value judgments, or self-perception, a response that includes summarizing the situation, explaining the underlying structure, multi-stage reasoning, or presenting an abstract framework may be appropriate.

[0009] Therefore, when conversational artificial intelligence systems generate uniform responses, they may produce responses that are too shallow, too deep, too abstract, or too specific to user input, resulting in responses that do not adequately match the user input.

[0010] Therefore, the object of the present invention is to provide an information processing system, program, and information processing method that can adapt the response of an interactive artificial intelligence system to user input. [Means for solving the problem]

[0011] To achieve the above objective, the information processing system described in claim 1 of the present invention is: An input acquisition unit (11) that acquires user input directed to an interactive artificial intelligence system, A feature calculation unit (13) calculates a plurality of features including the abstraction level of the user input and the inference requirement level of the user input based on the acquired user input, A distribution generation unit (14) generates a cognitive depth probability distribution across multiple depth levels based on the multiple features calculated above, A depth determination unit (15) determines the response depth corresponding to the user input based on the cognitive depth probability distribution generated above, A control information generation unit (16) generates control information, which is information for causing the conversational artificial intelligence system to generate a natural language response that is suitable for the user input, based on the response depth determined above. An output unit (17) that outputs the user input and the control information to the interactive artificial intelligence system, It is characterized by being equipped with [the following features].

[0012] Furthermore, the information processing system described in claim 2 is The plurality of features are characterized in that they include at least one of the information density of the user input and the degree of metacognitive representation included in the user input.

[0013] Furthermore, the information processing system described in claim 3 is The distribution generation unit (14) is characterized by weighting the calculated features to calculate an overall score and generating the cognitive depth probability distribution based on the calculated overall score.

[0014] Furthermore, the information processing system described in claim 4 is The depth determination unit (15) is characterized by calculating a representative depth based on the cognitive depth probability distribution.

[0015] Furthermore, the information processing system described in claim 5 is The depth determination unit (15) is characterized by determining the response depth based on the calculated representative depth.

[0016] Furthermore, the information processing system described in claim 6 is The representative depth is characterized by being an expected value calculated by summing the products of numerical values ​​corresponding to the multiple depth levels and the probabilities of each depth level in the cognitive depth probability distribution.

[0017] Furthermore, the information processing system described in claim 7 is The feature quantity calculation unit (13) outputs a feature quantity calculation prompt including the user input to a large language model, and calculates at least a part of the plurality of feature quantities based on a numerical evaluation result returned from the large language model.

[0018] Further, the information processing system according to claim 8 is characterized in that The feature quantity calculation unit (13) calculates at least a part of the plurality of feature quantities based on words, symbols, phrases, semantic units included in the user input, or the number of characters of the user input.

[0019] Further, the information processing system according to claim 9 is characterized in that The feature quantity calculation unit (13) outputs a feature quantity calculation prompt including the user input to a large language model, and calculates at least a part of the plurality of feature quantities based on the numerical evaluation result returned from the large language model, and the words, symbols, phrases, semantic units included in the user input, or the number of characters of the user input.

[0020] Further, the program according to claim 10 causes a computer to execute an input acquisition process of acquiring a user input directed to a conversational artificial intelligence system, a feature quantity calculation process of calculating a plurality of feature quantities including the abstraction level of the user input and the inference request level of the user input based on the acquired user input, a distribution generation process of generating a cognitive depth probability distribution over a plurality of depth levels based on the plurality of calculated feature quantities, a depth determination process of determining a response depth corresponding to the user input based on the generated cognitive depth probability distribution, a control information generation process of generating control information, which is information for causing the conversational artificial intelligence system to generate a natural language response suitable for the user input, based on the determined response depth, and an output process of outputting the user input and the control information to the conversational artificial intelligence system.​

[0021] Furthermore, the information processing method described in claim 11 is Computers An input acquisition step that obtains user input directed at an interactive artificial intelligence system, A feature calculation step that calculates multiple features including the level of abstraction of the user input and the degree of inference required for the user input based on the acquired user input, A distribution generation step that generates a cognitive depth probability distribution across multiple depth levels based on the multiple features calculated above, A depth determination step in which the response depth corresponding to the user input is determined based on the cognitive depth probability distribution generated above, A control information generation step, which generates control information that causes the conversational artificial intelligence system to generate a natural language response that is suitable for the user input, based on the response depth determined above, The system is characterized by performing an output step of outputting the user input and the control information to the interactive artificial intelligence system.

[0022] Here, the symbols in parentheses indicate the corresponding elements or items shown in the drawings and the embodiments for carrying out the invention described later. [Effects of the Invention]

[0023] According to the present invention, the system includes an input acquisition unit that acquires user input directed to an interactive artificial intelligence system; a feature calculation unit that calculates a plurality of feature quantities, including the level of abstraction of the user input and the degree of inference required for the user input, based on the acquired user input; a distribution generation unit that generates a cognitive depth probability distribution based on the calculated plurality of feature quantities; a depth determination unit that determines the response depth based on the generated cognitive depth probability distribution; a control information generation unit that generates control information based on the determined response depth; and an output unit that outputs the user input and control information to the interactive artificial intelligence system. As a result, the interactive artificial intelligence system can generate natural language responses based on a response depth that reflects the level of abstraction and degree of inference required for the user input.

[0024] Furthermore, according to the present invention, since the multiple feature quantities further include at least one of the information density of the user input and the degree of metacognitive expression contained in the user input, the amount of information in the user input or the degree of self-aware expression can be reflected in the determination of the response depth, in addition to the level of abstraction and the degree of inference requirement. This makes it possible to better adapt the response of the conversational artificial intelligence system to the user input.

[0025] Furthermore, according to the present invention, the distribution generation unit calculates an overall score by weighting multiple calculated features and generates a cognitive depth probability distribution based on the calculated overall score. Therefore, it is possible to determine the response depth not only by treating multiple features independently, but also by using an integrated index that reflects the contribution of each feature. This makes it possible to adapt the response of the interactive artificial intelligence system to user input.

[0026] Furthermore, according to the present invention, since the depth determination unit calculates a representative depth based on a cognitive depth probability distribution, a cognitive depth probability distribution spanning multiple depth levels can be treated as a representative value that is easy to use in determining the response depth. This makes it easier to adapt the response of the conversational artificial intelligence system to the user input.

[0027] Furthermore, according to the present invention, since the depth determination unit determines the response depth based on the calculated representative depth, a cognitive depth probability distribution spanning multiple depth levels can be converted into a response depth used to generate control information for the conversational artificial intelligence system. This makes it possible to reflect the probabilistically estimated cognitive depth of user input as a response depth for controlling the explanatory granularity, degree of structuring, or number of inference stages of the natural language response.

[0028] Furthermore, according to the present invention, since the representative depth is an expected value calculated by summing the products of numerical values ​​corresponding to multiple depth levels and the probabilities of each depth level in the cognitive depth probability distribution, it is possible to reflect not only a single maximum probability depth level but also a probabilistic spread across multiple depth levels in determining the response depth. This makes it possible to further adapt the response of the conversational artificial intelligence system to the user input.

[0029] Furthermore, according to the present invention, the feature calculation unit outputs a feature calculation prompt including user input to a large-scale language model, and calculates at least a portion of multiple features based on the numerical evaluation results returned from the large-scale language model. This makes it possible to quantify features that are difficult to calculate by simple character count or word matching alone, such as the level of abstraction and inference requirement of the user input. As a result, the response of the interactive artificial intelligence system can be better adapted to the user input.

[0030] Furthermore, according to the present invention, the feature calculation unit calculates at least a portion of multiple features based on the words, symbols, phrases, semantic units, or the number of characters in the user input. Therefore, it is possible to calculate features corresponding to user input without relying on the output of a large-scale language model, or while suppressing the use of a large-scale language model. This makes it possible to perform the process of adapting the response of the conversational artificial intelligence system to the user input in a lightweight or stable manner.

[0031] Furthermore, according to the present invention, the feature calculation unit calculates at least a portion of multiple features based on the numerical evaluation results returned from the large-scale language model and the words, symbols, phrases, semantic units, or the number of characters in the user input. This allows for the combined use of evaluation results from the large-scale language model and rule-based or dictionary-based evaluation results. This enhances the flexibility and stability of feature calculation, allowing the response of the interactive artificial intelligence system to be more closely matched to the user input.

[0032] Furthermore, according to the present invention, the computer is made to perform an input acquisition process to acquire user input directed to an interactive artificial intelligence system, a feature calculation process to calculate multiple feature quantities including the level of abstraction and the degree of inference requirement, a distribution generation process to generate a cognitive depth probability distribution, a depth determination process to determine the response depth, a control information generation process to generate control information, and an output process to output user input and control information. Thus, a function to adapt the response of the interactive artificial intelligence system to the user input can be realized as a program.

[0033] Furthermore, according to the present invention, the computer performs an input acquisition step of acquiring user input directed to an interactive artificial intelligence system, a feature calculation step of calculating a plurality of features including the level of abstraction and the degree of inference requirement, a distribution generation step of generating a cognitive depth probability distribution, a depth determination step of determining the response depth, a control information generation step of generating control information, and an output step of outputting user input and control information. Thus, an information processing method can be realized that adapts the response of the interactive artificial intelligence system to the user input.

[0034] Furthermore, the fact that the information processing system, program, and information processing method of the present invention calculate multiple feature quantities including the level of abstraction and the degree of inference requirement of user input, generate a cognitive depth probability distribution across multiple depth levels based on these feature quantities, determine the response depth based on this cognitive depth probability distribution, and output control information based on this response depth to the interactive artificial intelligence system is not described at all in the aforementioned Non-Patent Documents 1 and 2. [Brief explanation of the drawing]

[0035] [Figure 1] This figure shows an information processing system according to the first embodiment of the present invention. [Figure 2] This figure shows the functional configuration of the management server in the information processing system according to the first embodiment of the present invention. [Figure 3] This is a detailed diagram of the feature calculation unit of the management server in the information processing system according to the first embodiment of the present invention. [Figure 4] This is a sequence diagram showing the processing of a program in an information processing system according to the first embodiment of the present invention. [Figure 5] This is a flowchart showing an example of the cognitive depth probability distribution generation and response depth determination process according to the first embodiment. [Figure 6] This is a correspondence table between depth levels and response specifications in an information processing system according to the first embodiment of the present invention. [Figure 7] This is a sequence diagram showing the processing of a program in an information processing system according to the second embodiment of the present invention. [Figure 8] This is a detailed diagram of the feature calculation unit of the management server in the information processing system according to the third embodiment of the present invention. [Modes for carrying out the invention]

[0036] (First Embodiment) An information processing system, program, and information processing method according to the first embodiment of the present invention will be described with reference to Figures 1 to 6. As an example of a program according to the present invention, a program executed in an information processing system is provided. The information processing method according to the present invention is realized by the execution of this program in the information processing system.

[0037] <1. Outline of the Information Processing System> As shown in Figure 1, the information processing system according to the first embodiment comprises a management server 10 and a first large-scale language model 20. The user terminal 1 and the second large-scale language model 2 are located outside the information processing system and are connected to the information processing system via a network so as to be able to communicate with it.

[0038] User terminal 1 is a terminal device used by the user. User terminal 1 may be, for example, a smartphone, tablet, personal computer, or other information processing terminal. The user uses user terminal 1 to input user input directed to the conversational artificial intelligence system. User input may be, for example, a question, request, consultation, instruction, explanation, or a combination thereof, written in natural language.

[0039] The management server 10 is an information processing device that obtains user input from the user terminal 1 and generates control information to be provided to the second large-scale language model 2 based on that user input. The management server 10 acts as an intermediary between the user terminal 1 and the second large-scale language model 2. Instead of directly passing user input to the second large-scale language model 2, it determines the response depth based on the user input and generates control information based on that response depth.

[0040] Control information is information that enables the second large-scale language model 2 to generate a natural language response that is adapted to user input. The control information includes, for example, a depth specification value indicating the response depth, a response specification corresponding to the response depth, or a prompt for response generation. The response specification is information that specifies, for example, at least one of the response's level of abstraction, descriptive granularity, number of inference stages, degree of structuring, or expertise.

[0041] The management server 10 calculates several features based on user input to estimate the cognitive depth of the user input in order to determine the response depth. In this embodiment, the multiple features include abstraction level A, inference requirement level I, information density D, and metacognitive words M.

[0042] Abstraction level A is a feature that indicates whether user input is closer to specific fact-finding or specific work instructions, or closer to abstract consultation, value judgment, or conceptual question. For example, inputs requesting specific numbers, dates, times, locations, procedures, or factual verification may be evaluated as inputs with a relatively low level of abstraction. On the other hand, inputs involving value judgments, conceptual clarification, self-understanding, policy consideration, or abstract consultations may be evaluated as inputs with a relatively high level of abstraction.

[0043] The inference requirement I is a feature that represents the degree to which background information, causal relationships, implicit assumptions, implications, or multi-step inference are required to generate an appropriate response to user input. For example, an input that only requires a single factual response may be evaluated as having a relatively low degree of inference requirement. On the other hand, an input that requires supplementing background information not explicitly stated in the user input, comparing multiple possibilities, and considering causal relationships or response strategies step by step may be evaluated as having a relatively high degree of inference requirement.

[0044] Information density D is a feature that represents the amount of information or the density of semantic units contained in user input. Information density may be calculated based, for example, on the number of words, phrases, clauses, semantic units, characters, named entities, conditional expressions, or combinations thereof included in user input. Alternatively, information density may be calculated based on the number of multiple circumstances, conditions, constraints, purposes, or sentimental expressions included in user input.

[0045] Metacognitive term M is a feature that represents the degree of metacognitive expression related to self-awareness, thought processes, judgments, confirmations, verifications, hypotheses, adjustments, or introspection contained in user input. Metacognitive terms may be calculated, for example, based on the degree of agreement between user input and words included in a metacognitive term dictionary stored in the memory unit 22. The metacognitive term dictionary may include words such as "notice," "overview," "organize," "intention," "inference," "verification," "hypothesis," "adjustment," "reflection," and "confirmation."

[0046] It should be noted that metacognitive terms are not limited to specific words or phrases included in user input. Metacognitive terms may also include expressions that objectify one's own state, understanding, judgment, feelings, behavior, or thought processes. For example, expressions such as "I don't even know why," "I want to organize what I should prioritize," and "I want to check if my thinking is biased" may be evaluated as metacognitive expressions.

[0047] The first large-scale language model 20 numerically evaluates at least some of the multiple features based on user input and feature calculation prompts received from the management server 10. In this embodiment, the first large-scale language model 20 numerically evaluates the base value of the level of abstraction and the base value of the inference requirement, and returns the numerical evaluation result, including the said base value, to the management server 10.

[0048] On the other hand, information density and metacognitive terms are calculated by local auxiliary analysis within the management server 10. For example, the management server 10 calculates information density and metacognitive terms based on the preprocessing results obtained by lightweight preprocessing and the dictionary, rules, thresholds, or correction rules stored in the memory unit 22. The management server 10 also calculates correction values ​​for the level of abstraction and the degree of inference requirement based on dictionary match rate, causal representation, conditional representation, number of issues, or number of semantic units.

[0049] The second large-scale language model 2 is an example of an interactive artificial intelligence system that generates natural language responses in response to user input. The second large-scale language model 2 generates natural language responses based on user input and control information output from the management server 10. The second large-scale language model 2 is, for example, an external interactive artificial intelligence service or a large-scale language model provided separately from the management server 10.

[0050] The first large-scale language model 20 and the second large-scale language model 2 may be the same model or they may be different models. For example, the first large-scale language model 20 and the second large-scale language model 2 may be provided by the same large-scale language model service. Furthermore, the first large-scale language model 20 is a model used on the management server 10 side to numerically evaluate the base values ​​of features, and the second large-scale language model 2 may be a different conversational artificial intelligence system for each user, each session, each application, or each contract type.

[0051] In this configuration, the management server 10 does not simply pass user input to the second large-scale language model 2, but rather determines the response depth based on the user input and generates control information based on that response depth. Specifically, the management server 10 determines the level of abstraction, the degree of inference requirement, the information density, and the metacognitive words of the user input based on the numerical evaluation results returned from the first large-scale language model 20 and the rule evaluation values ​​or correction values ​​calculated on the management server 10 side.

[0052] The second large-scale language model 2 generates a natural language response based on user input and control information received from the management server 10. The management server 10 may acquire the natural language response generated by the second large-scale language model 2 and output the acquired natural language response to the user terminal 1. In other embodiments, the second large-scale language model 2 or the external service providing it may directly output the natural language response to the user terminal 1.

[0053] <2. Functional configuration of management server 10> Next, the functional configuration of the management server 10 according to the first embodiment will be described. Figure 2 is a block diagram showing an example of the functional configuration of the management server 10 according to the first embodiment. As shown in Figure 2, the management server 10 comprises an input acquisition unit 11, a lightweight preprocessing unit 12, a feature calculation unit 13, a distribution generation unit 14, a depth determination unit 15, a control information generation unit 16, an output unit 17, a response acquisition unit 18, a response output unit 19, and a storage unit 22.

[0054] The input acquisition unit 11 acquires user input from the user terminal 1 directed to the conversational artificial intelligence system (second large-scale language model 2). User input is, for example, a natural language sentence entered by the user to request a response from the second large-scale language model 2. User input may include questions, requests, consultations, instructions, situational descriptions, sentimental statements, questions with value judgments, or combinations thereof.

[0055] The lightweight preprocessing unit 12 performs lightweight preprocessing on the user input acquired by the input acquisition unit 11. Lightweight preprocessing includes, for example, removing unnecessary whitespace, symbols, or fillers, splitting into sentence or phrase units, calculating character counts, extracting words and phrases, extracting segments, extracting semantic units, normalizing input formats such as unifying full-width and half-width characters, reducing stop word weights, or a combination thereof. The preprocessing results obtained by the lightweight preprocessing unit 12 are used for local auxiliary analysis within the management server 10. For example, the preprocessing results may be used to calculate information density, the number of words and phrases, the number of segments, the number of semantic units, or the number of characters, to determine agreement with abstract or concrete words, to determine agreement with a metacognitive word dictionary, or to calculate correction values.

[0056] The feature calculation unit 13 calculates multiple features based on user input to estimate the cognitive depth of the user input. These multiple features include abstraction level, inference requirement level, information density, and metacognitive terms. The feature calculation unit 13 outputs user input and feature calculation prompts to the first large-scale language model 20 and obtains the numerical evaluation results returned from the first large-scale language model 20. In this embodiment, the first large-scale language model 20 numerically evaluates at least the base value of the abstraction level and the base value of the inference requirement level.

[0057] On the other hand, in this embodiment, information density and metacognitive words are not subject to numerical evaluation by the first large-scale language model 20, but are calculated by local auxiliary analysis within the management server 10. The feature calculation unit 13 determines the final level of abstraction, inference requirement, information density, and metacognitive words based on the numerical evaluation results obtained from the first large-scale language model 20 and the auxiliary analysis results from the local side. The level of abstraction, inference requirement, information density, and metacognitive words determined by the feature calculation unit 13 are output to the distribution generation unit 14.

[0058] In the above embodiment, an example was described in which information density and metacognitive terms are not subject to numerical evaluation by the first large-scale language model 20. However, in another example, the first large-scale language model 20 may be made to numerically evaluate all of the abstraction level, inference requirement level, information density, and metacognitive terms. For example, the feature calculation unit 13 may generate a feature calculation prompt that instructs the model to return a numerical value within a predetermined range for each of the abstraction level, inference requirement level, information density, and metacognitive terms. In this case, the first large-scale language model 20 may return a numerical evaluation result including a base value for the abstraction level, a base value for the inference requirement level, a base value for the information density, and a base value for the metacognitive terms, based on the user input and the feature calculation prompt. The feature calculation unit 13 may then determine the abstraction level, inference requirement level, information density, and metacognitive terms based on the numerical evaluation result.

[0059] The distribution generation unit 14 generates cognitive depth probability distributions across multiple depth levels based on multiple features determined by the feature calculation unit 13. In the first embodiment, the distribution generation unit 14 calculates an overall score by weighting the determined level of abstraction, inference requirement, information density, and metacognitive words, and generates a cognitive depth probability distribution based on the overall score.

[0060] For example, the distribution generation unit 14 may calculate an overall score by referring to the weight coefficients stored in the memory unit 22 and applying the weight coefficients corresponding to the level of abstraction, the degree of inference requirement, the information density, and the metacognitive words. The weight coefficients are values ​​that indicate the degree to which each feature contributes to cognitive depth. The weight coefficients may be fixed values ​​set in advance, or they may be changed according to user attributes, purpose of use, type of conversational artificial intelligence system, or service settings.

[0061] The distribution generation unit 14 may calculate weights corresponding to each depth level based on the calculated overall score and the central value of each depth level stored in the memory unit 22. The distribution generation unit 14 may assign a larger weight to a depth level if the difference between the overall score and the central value of each depth level is small. The distribution generation unit 14 then generates a cognitive depth probability distribution by normalizing the weights of each depth level.

[0062] The depth determination unit 15 determines the response depth corresponding to user input based on the cognitive depth probability distribution generated by the distribution generation unit 14. In the first embodiment, the depth determination unit 15 calculates a representative depth based on the cognitive depth probability distribution and determines the response depth based on the calculated representative depth. The representative depth is, for example, an expected value calculated by summing the products of the numerical value corresponding to each depth level and the probability of each depth level.

[0063] The control information generation unit 16 generates control information to be provided to the second large-scale language model 2 based on the response depth determined by the depth determination unit 15. The control information is information that causes the second large-scale language model 2 to generate a natural language response that is suitable for user input. The control information includes, for example, a depth specification value indicating the response depth, a response specification corresponding to the response depth, a response generation prompt, or a combination thereof.

[0064] The response specification is information that specifies the manner of the natural language response to be generated by the second large-scale language model 2. The response specification specifies, for example, the granularity of the response's explanation, level of abstraction, degree of structuring, number of inference levels, order in which conclusions are presented, whether or not explanations of reasons are provided, whether or not examples are provided, the specificity of the advice, or a combination thereof. The control information generation unit 16 may refer to the depth correspondence table stored in the storage unit 22 and generate a response specification or response generation prompt corresponding to the response depth.

[0065] The output unit 17 outputs user input and control information to the second large-scale language model 2. The second large-scale language model 2 is an example of an interactive artificial intelligence system and generates natural language responses based on user input and control information. The output unit 17 may output user input and control information to the second large-scale language model 2 via, for example, a communication interface, an application programming interface (API), a browser extension, a plugin, or other means of cooperation.

[0066] The response acquisition unit 18 acquires the natural language response generated by the second large-scale language model 2. The second large-scale language model 2 generates a natural language response corresponding to the response depth based on the user input and control information received from the output unit 17. The response acquisition unit 18 acquires the natural language response returned from the second large-scale language model 2 and outputs the acquired natural language response to the response output unit 19.

[0067] The response output unit 19 outputs the natural language response acquired by the response acquisition unit 18 to the user terminal 1. This allows the user terminal 1 to present the natural language response generated by the second large-scale language model 2 to the user. In other embodiments, the second large-scale language model 2 or an external service providing it may directly output the natural language response to the user terminal 1.

[0068] The memory unit 22 stores various information used in each process performed by the management server 10. For example, the memory unit 22 stores prompts for feature calculation, an abstract word dictionary, a concrete word dictionary, a metacognitive word dictionary, rules, correction values, correction rules, weight coefficients for each feature, multiple depth levels, a central value corresponding to each depth level, normalization conditions, sensitivity coefficients, a formula for generating a cognitive depth probability distribution, a formula for calculating a representative depth, a depth correspondence table, a response specification, or a prompt for generating a response.

[0069] The management server 10 is implemented by a computer, for example, which includes a processor, memory, storage device, and communication interface. The processor may implement at least a portion of the input acquisition unit 11, lightweight preprocessing unit 12, feature calculation unit 13, distribution generation unit 14, depth determination unit 15, control information generation unit 16, output unit 17, response acquisition unit 18, and response output unit 19 by executing a program stored in the storage device or memory.

[0070] <3. Detailed configuration of the feature calculation unit 13> Next, the detailed configuration of the feature calculation unit 13 according to the first embodiment will be described. Figure 3 is a diagram showing an example of the detailed configuration of the feature calculation unit 13 according to the first embodiment. In the first embodiment, the feature calculation unit 13 is a hybrid type feature calculation unit 13 that calculates the features of user input by using in combination the numerical evaluation results from the first large-scale language model 20 and the local auxiliary analysis and rule evaluation performed on the management server 10 side.

[0071] As shown in Figure 3, the feature calculation unit 13 includes a feature calculation prompt generation unit 13a, an input analysis unit 13b, a rule evaluation unit 13c, a numerical evaluation result acquisition unit 13d, an integration and correction unit 13e, and a feature determination unit 13f.

[0072] The feature calculation prompt generation unit 13a generates a feature calculation prompt for the first large-scale language model 20 to numerically evaluate at least the abstraction level and inference requirement level from among the features of the user input. The feature calculation prompt includes user input and information instructing the model to return a numerical value within a predetermined range for each feature of abstraction level and inference requirement level. The predetermined range is, for example, 0.00 or more and 1.00 or less.

[0073] The feature calculation prompt generation unit 13a outputs the generated feature calculation prompt and the original input, which is the user input, to the first large-scale language model 20. Based on the feature calculation prompt and the user input, the first large-scale language model 20 numerically evaluates the base value of the level of abstraction and the base value of the inference requirement for the user input. The first large-scale language model 20 may, for example, return the numerical evaluation result, including the base value of the level of abstraction and the base value of the inference requirement, in JSON format or other machine-readable format.

[0074] The numerical evaluation result acquisition unit 13d acquires the numerical evaluation results returned from the first large-scale language model 20. The numerical evaluation results include, for example, a base value for the level of abstraction and a base value for the inference requirement. The numerical evaluation result acquisition unit 13d outputs the acquired numerical evaluation results to the integration and correction unit 13e.

[0075] In the first embodiment, information density and metacognitive terms are not subject to numerical evaluation by the first large-scale language model 20, but are calculated by local auxiliary analysis within the management server 10. That is, in the first embodiment, the numerical evaluation results returned from the first large-scale language model 20 include a base value for the level of abstraction and a base value for the level of inference requirement, while information density and metacognitive terms are calculated based on the analysis results by the input analysis unit 13b and the rule evaluation unit 13c, which will be described later.

[0076] The input analysis unit 13b performs auxiliary analysis of user input locally based on the preprocessing results obtained from the lightweight preprocessing unit 12. For example, the input analysis unit 13b extracts words, phrases, clauses, semantic units, character counts, conditional expressions, causal expressions, sentiment expressions, abstract words, concrete words, or metacognitive expressions contained in the user input. The auxiliary analysis results from the input analysis unit 13b are not transmitted to the first large-scale language model 20, but are used by the rule evaluation unit 13c to calculate information density, metacognitive words, rule evaluation values, or correction values.

[0077] The rule evaluation unit 13c calculates a rule evaluation value, information density, metacognitive words, or correction value for user input based on the analysis results obtained by the input analysis unit 13b and the dictionary, rules, or thresholds stored in the memory unit 22. For example, the rule evaluation unit 13c may refer to the abstract word dictionary and the concrete word dictionary and calculate a rule evaluation value or correction value related to the level of abstraction based on the frequency of occurrence of abstract words or concrete words included in the user input.

[0078] Furthermore, the rule evaluation unit 13c may calculate a rule evaluation value or correction value related to the degree of inference requirement based on the causal expression, conditional expression, expression indicating assumptions, or expression indicating multiple points of contention included in the user input. For example, if the user input requires unexpressed background circumstances, causal relationships, multiple possibilities, or step-by-step judgment, the rule evaluation unit 13c may calculate a relatively high rule evaluation value or correction value related to the degree of inference requirement.

[0079] Furthermore, the rule evaluation unit 13c calculates information density based on the number of words, phrases, clauses, semantic units, characters, named entities, conditional expressions, or combinations thereof included in the user input. In addition, the rule evaluation unit 13c refers to a metacognitive word dictionary and calculates metacognitive words based on the frequency of appearance of expressions related to self-awareness, thought processes, judgments, confirmations, verifications, hypotheses, adjustments, or introspection included in the user input.

[0080] For example, the rule evaluation unit 13c calculates the information density as a relatively high value if the number of semantic units included in the user input is equal to or greater than a predetermined semantic unit threshold. Alternatively, the rule evaluation unit 13c may calculate the information density by dividing the number of phrases or characters by a predetermined normalization coefficient and clipping it to a range of 0.00 to 1.00.

[0081] Furthermore, the rule evaluation unit 13c calculates metacognitive words as having a relatively high value if the number of occurrences of words or similar metacognitive expressions included in the metacognitive word dictionary is equal to or greater than a predetermined metacognitive word threshold. For example, if the user input includes expressions related to self-awareness, judgment, confirmation, or introspection such as "I don't know," "I want to organize my thoughts," "I feel like," "I want to confirm," or "I want to think about it," the rule evaluation unit 13c calculates metacognitive words in a way that increases their value.

[0082] The memory unit 22 stores information referenced by the rule evaluation unit 13c and the integration / correction unit 13e. The memory unit 22 stores, for example, an abstract word dictionary, a concrete word dictionary, a metacognitive word dictionary, rules, thresholds, correction values, correction rules, or combinations thereof. The correction values ​​or correction rules stored in the memory unit 22 are used to correct the base value of the level of abstraction or the base value of the inference requirement returned from the first large-scale language model 20, or to correct the information density or metacognitive words calculated by the local auxiliary analysis.

[0083] The integration and correction unit 13e calculates the corrected value of each feature based on the base value of abstraction level and the base value of inference requirement level obtained from the numerical evaluation result acquisition unit 13d, the rule evaluation value, information density, metacognitive word or correction value obtained from the rule evaluation unit 13c, and the correction value or correction rule stored in the memory unit 22.

[0084] For example, the integration and correction unit 13e may calculate the corrected value of the abstraction level by applying a correction value to the base value of the abstraction level according to the degree of agreement with the abstract word dictionary or the concrete word dictionary. For example, if the abstract word dictionary agreement rate is equal to or greater than a predetermined abstract word threshold, the integration and correction unit 13e applies a positive correction value to the base value of the abstraction level. Also, if the concrete word dictionary agreement rate is equal to or greater than a predetermined concrete word threshold, the integration and correction unit 13e applies a negative correction value to the base value of the abstraction level.

[0085] Furthermore, the integration and correction unit 13e may calculate a corrected value of the inference requirement by applying a correction value to the base value of the inference requirement, according to the causal relationship, implicit assumptions, number of conditions, number of logical steps, or amount of latent missing information. For example, if the number of causal expressions or conditional expressions included in the user input is greater than or equal to a predetermined number, the integration and correction unit 13e applies a positive correction value to the base value of the inference requirement. The amount of missing latent information refers to the amount of information that is not explicitly stated in the input text but needs to be supplemented in order to generate an appropriate response.

[0086] On the other hand, in the first embodiment, for information density and metacognitive terms, values ​​calculated by the rule evaluation unit 13c are used instead of base values ​​returned from the first large-scale language model 20. The integration and correction unit 13e may use the information density and metacognitive terms calculated by the rule evaluation unit 13c as the corrected value for information density and the corrected value for metacognitive terms, respectively. Alternatively, the integration and correction unit 13e may calculate the corrected value for information density or metacognitive terms by applying a correction value based on a correction rule.

[0087] The integration and correction unit 13e may calculate the corrected value, for example, by adding the base value and the correction value for the level of abstraction and the degree of inference requirement. Alternatively, the integration and correction unit 13e may calculate the corrected value by weighting and integrating the base value, the rule evaluation value, and the correction value. Furthermore, the integration and correction unit 13e may apply the correction value only when the difference between the base value and the rule evaluation value is greater than or equal to a predetermined value, or it may calculate the corrected value by prioritizing either the base value or the rule evaluation value.

[0088] For example, if the base value of the level of abstraction is 0.74 and a correction value of +0.04 is calculated for the level of abstraction based on the fact that the abstract word dictionary match rate is above a predetermined abstract word threshold, the integration and correction unit 13e calculates 0.78 as the corrected value of the level of abstraction. Also, if the base value of the level of inference requirement is 0.77 and a correction value of +0.04 is calculated for the level of inference requirement based on the fact that the causal expression, conditional expression, or multiple semantic units are above a predetermined threshold, the integration and correction unit 13e calculates 0.81 as the corrected value of the level of inference requirement.

[0089] The feature determination unit 13f determines the final features based on the corrected values ​​calculated by the integration and correction unit 13e. In the first embodiment, the feature determination unit 13f determines the level of abstraction, the degree of inference requirement, the information density, and the metacognitive word as the final features. The feature determination unit 13f may, for example, correct the corrected values ​​to within a predetermined range if the corrected values ​​exceed a predetermined range. For example, if each feature is expressed as 0.00 or more and 1.00 or less, the feature determination unit 13f may correct values ​​less than 0.00 to 0.00 and values ​​greater than 1.00 to 1.00.

[0090] For example, the feature determination unit 13f determines the corrected value of abstraction (0.78), the corrected value of inference requirement (0.81), the information density (0.68), and the metacognitive word (0.62), calculated by the integration and correction unit 13e, as the final abstraction, inference requirement, information density, and metacognitive word. Note that the above values ​​are examples and may be changed depending on the dictionary, rules, thresholds, correction values, or correction rules.

[0091] The feature determination unit 13f outputs the determined level of abstraction, inference requirement, information density, and metacognitive words as determined features to the distribution generation unit 14. Based on the determined features obtained from the feature determination unit 13f, the distribution generation unit 14 calculates an overall score and generates a cognitive depth probability distribution.

[0092] Thus, the feature calculation unit 13 according to the first embodiment does not use the numerical evaluation results of the level of abstraction and the level of inference requirement obtained by the first large-scale language model 20 as final features, but rather uses rule evaluation values, dictionaries, rules, or correction values ​​obtained on the management server 10 side to determine the level of abstraction and the level of inference requirement. Furthermore, the feature calculation unit 13 according to the first embodiment calculates information density and metacognitive words not by numerical evaluation by the first large-scale language model 20, but by local auxiliary analysis and rule evaluation on the management server 10 side. This makes it possible to calculate features used for estimating the cognitive depth of user input by combining the flexible semantic evaluation by the first large-scale language model 20 with stable evaluation based on dictionaries or rules on the management server 10 side.

[0093] <4. Processing Sequence of Information Processing Systems> Next, the processing sequence of the information processing system according to the first embodiment will be described. Figure 4 is a diagram showing an example of the processing sequence in the information processing system according to the first embodiment. The processing sequence shown in Figure 4 is a hybrid processing example in which the level of abstraction, level of inference, information density, and metacognitive words are determined by using in combination the numerical evaluation results of the level of abstraction and level of inference requirement by the first large-scale language model 20 and the information density, metacognitive words, and correction values ​​calculated by the local auxiliary analysis on the management server 10 side.

[0094] First, user terminal 1 sends the user input entered by the user to management server 10. The input acquisition unit 11 of management server 10 acquires the user input from user terminal 1 directed to the conversational artificial intelligence system (S101). The user input is, for example, a natural language sentence entered by the user to ask the second large-scale language model 2 for a response, and we assume that the user input is "Lately, I feel like I'm not being appreciated despite working hard, and I'm starting to wonder if I have any value." This user input includes aspects of work effort, evaluation from others, and anxieties about self-worth. Furthermore, because this user input includes expressions of the user's emotional state and self-perception, it is appropriate to respond to it not merely by confirming facts, but by taking into account the background circumstances, emotions, value judgments, or self-perception.

[0095] Next, the lightweight preprocessing unit 12 of the management server 10 performs lightweight preprocessing on the acquired user input (S102). Lightweight preprocessing includes, for example, sentence segmentation, input format normalization, filler removal, stop word weight reduction, word / phrase extraction, phrase extraction, semantic unit extraction, character count calculation, or a combination thereof. The preprocessing results obtained by lightweight preprocessing are used for auxiliary analysis within the management server 10. For example, the preprocessing results are used to calculate information density, dictionary match rate, number of semantic units, or correction values. For example, the lightweight preprocessing unit 12 may divide the user input into sentences or phrases and extract semantic units such as "I'm working hard," "I feel like I'm not being appreciated," or "I don't know if I have any value."

[0096] Next, the feature calculation unit 13 generates a feature calculation prompt that includes user input. The feature calculation prompt prompts the first large-scale language model 20 to numerically evaluate the base values ​​of the features related to the user input. In the first embodiment, the feature calculation prompt includes information instructing the model to return the base values ​​of the abstraction level and inference requirement level among a plurality of features as numerical values ​​between 0.00 and 1.00, for example.

[0097] In the first embodiment, information density and metacognitive terms are not subject to numerical evaluation by the first large-scale language model 20. As described later, information density and metacognitive terms are calculated by local auxiliary analysis within the management server 10 based on the preprocessing results obtained by lightweight preprocessing and dictionaries, rules, thresholds, or correction rules stored within the management server 10.

[0098] The feature calculation unit 13 sends user input and feature calculation prompts to the first large-scale language model 20 (S103). Here, the feature calculation unit 13 sends the user input, which is the original input provided by the user, to the first large-scale language model 20. This allows the first large-scale language model 20 to numerically evaluate the base values ​​of the level of abstraction and the inference requirement while preserving the natural context of the user input.

[0099] The first large-scale language model 20 numerically evaluates the base values ​​of the level of abstraction and the inference requirement based on the user input and feature calculation prompts received from the management server 10 (S104). Specifically, the first large-scale language model 20 numerically evaluates the level of abstraction of the user input within the range of 0.00 to 1.00 based on the instructions for determining the level of abstraction included in the feature calculation prompts. The first large-scale language model 20 also numerically evaluates the inference requirement of the user input within the range of 0.00 to 1.00 based on the instructions for determining the inference requirement included in the feature calculation prompts.

[0100] When the user input is "Lately, I feel like I'm not being appreciated despite working hard, and I'm starting to doubt my own worth," the first large-scale language model 20 generates a numerical evaluation result for the level of abstraction and a numerical evaluation result for the inference requirement for that user input. In this case, the first large-scale language model 20 generates 0.74 as the numerical evaluation result for the level of abstraction and 0.77 as the numerical evaluation result for the inference requirement.

[0101] The first large-scale language model 20 returns the generated numerical evaluation results to the management server 10. The feature calculation unit 13 of the management server 10 receives numerical evaluation results from the first large-scale language model 20, including base values ​​for the level of abstraction and the inference requirement (S105). The numerical evaluation results may be returned in JSON format or other machine-readable format, for example.

[0102] Next, the feature calculation unit 13 calculates information density, metacognitive words, and correction values ​​based on the preprocessing results obtained by the lightweight preprocessing unit 12 and the dictionary, rules, thresholds, or correction rules stored in the memory unit 22 (S106). This process is a local auxiliary analysis performed within the management server 10 and is not a process that is sent to the first large-scale language model 20 or the second large-scale language model 2.

[0103] Specifically, the feature calculation unit 13 extracts words, phrases, clauses, semantic units, character counts, conditional expressions, causal expressions, sentiment expressions, abstract words, concrete words, or metacognitive expressions included in the user input based on the preprocessing results. The feature calculation unit 13 calculates the information density D and metacognitive words M based on the extraction results and the dictionary, rules, or thresholds stored in the memory unit 22.

[0104] For example, the feature calculation unit 13 calculates information density D based on the number of words, phrases, semantic units, characters, conditional expressions, named entities, or combinations thereof included in the user input. The feature calculation unit 13 also refers to the metacognitive word dictionary stored in the memory unit 22 and calculates metacognitive words M based on the frequency of appearance of expressions related to self-awareness, thought processes, judgments, confirmations, verifications, hypotheses, adjustments, or introspection included in the user input. For example, the feature calculation unit 13 compares the words included in the user input with the abstract word dictionary and concrete word dictionary stored in the memory unit 22. If the match rate between the words included in the abstract word dictionary and the user input is greater than or equal to a predetermined abstract word threshold, the feature calculation unit 13 determines that the user input tends to be highly abstract and calculates a positive correction value as the abstraction correction value Corr_A. For example, if the abstract word dictionary match rate is 0.30 or higher, the feature calculation unit 13 calculates the abstraction correction value Corr_A = +0.04.

[0105] Furthermore, if the match rate between the words included in the concrete word dictionary and the user input is greater than or equal to a predetermined concrete word threshold, the feature calculation unit 13 determines that the level of abstraction of the user input tends to be low and calculates a negative correction value as the level of abstraction correction value Corr_A. For example, if the concrete word dictionary match rate is 0.30 or higher, the feature calculation unit 13 calculates the level of abstraction correction value Corr_A = -0.04. On the other hand, if the concrete word dictionary match rate is less than a predetermined concrete word threshold, and the abstract word dictionary match rate is greater than or equal to a predetermined abstract word threshold, the feature calculation unit 13 may maintain or increase a positive correction value as the level of abstraction correction value Corr_A.

[0106] Furthermore, the feature calculation unit 13 calculates an inference requirement correction value Corr_I based on the causal expression, conditional expression, expression indicating an assumption, or expression indicating multiple points of discussion included in the user input. For example, if the user input includes words indicating causal or conditional relationships such as "why," "how," "on the other hand," "in the case," "cause," or "impact," the feature calculation unit 13 calculates a positive correction value as the inference requirement correction value Corr_I. Also, if multiple semantic units or multiple points of discussion are extracted from the user input, the feature calculation unit 13 determines that there is a high possibility that multiple stages of inference will be required and calculates a positive correction value as the inference requirement correction value Corr_I.

[0107] For example, if the number of causal or conditional expressions included in the user input is greater than or equal to a predetermined number, the feature calculation unit 13 calculates an inference requirement correction value Corr_I = +0.05. Also, if the number of semantic units extracted from the user input is greater than or equal to a predetermined number, the feature calculation unit 13 calculates an inference requirement correction value Corr_I = +0.03. On the other hand, if the user input is a single fact confirmation or a simple work instruction, and the number of causal expressions, conditional expressions, or multiple semantic units is less than a predetermined threshold, the feature calculation unit 13 may calculate a correction value of 0 or a negative value as the inference requirement correction value Corr_I.

[0108] Furthermore, the feature calculation unit 13 calculates the information density D based on the number of words, phrases, clauses, semantic units, characters, or named entities included in the user input. For example, if the number of semantic units is greater than or equal to a predetermined semantic unit threshold, the feature calculation unit 13 calculates the information density D as a relatively high value. In addition, the feature calculation unit 13 calculates the information density D by dividing the number of phrases or characters by a predetermined normalization coefficient and clipping it to a range of 0.00 to 1.00.

[0109] Furthermore, the feature calculation unit 13 compares the words included in the user input with the metacognitive word dictionary stored in the memory unit 22. If the number of occurrences of words or similar metacognitive expressions included in the metacognitive word dictionary is equal to or greater than a predetermined metacognitive word threshold, the feature calculation unit 13 calculates the metacognitive word M as a relatively high value. For example, if the user input includes expressions related to self-awareness, judgment, confirmation, or introspection such as "I don't know," "I want to organize it," "I feel like," "I want to confirm," or "I want to think about it," the feature calculation unit 13 calculates the metacognitive word M in a way that increases its value.

[0110] Next, the feature calculation unit 13 determines the final features A, I, D, and M based on the base value of abstraction level Base_A, the base value of inference requirement level Base_I, information density D, metacognitive word M, and correction value Corr (S107). Here, feature A represents the final abstraction level, feature I represents the final inference requirement level, feature D represents the information density, and feature M represents the metacognitive word.

[0111] For example, if the base value of the level of abstraction, Base_A, is 0.74, and an abstraction correction value, Corr_A = +0.04, is calculated based on the fact that the abstract word dictionary match rate is above a predetermined abstract word threshold, the feature calculation unit 13 calculates a final level of abstraction, A = 0.78, using A = Base_A + Corr_A. Also, if the base value of the inference requirement, Base_I, is 0.77, and an inference requirement correction value, Corr_I = +0.04, is calculated based on the fact that the causal expression, conditional expression, or multiple semantic units are above a predetermined threshold, the feature calculation unit 13 calculates a final inference requirement, I = 0.81, using I = Base_I + Corr_I.

[0112] Furthermore, the feature calculation unit 13 calculates the information density D=0.68 and the metacognitive word M=0.62 based on the preprocessing results. In this case, the feature calculation unit 13 determines the final features to be abstraction level A=0.78, inference requirement I=0.81, information density D=0.68, and metacognitive word M=0.62 (S107). Note that the above thresholds, correction values, and the values ​​of each feature are examples and may be changed according to the dictionary, rules, thresholds, correction values, or correction rules stored in the memory unit 22.

[0113] Next, the distribution generation unit 14 and the depth determination unit 15 determine the overall score Score, cognitive depth probability distribution P(L|U), representative depth E[L], and response depth L* based on the abstraction level A, inference requirement level I, information density D, and metacognitive word M determined by the feature calculation unit 13 (S108).

[0114] Specifically, the distribution generation unit 14 refers to the weight coefficients stored in the memory unit 22 and weights the level of abstraction A, the degree of inference requirement I, the information density D, and the metacognitive word M to calculate the overall score Score. For example, the distribution generation unit 14 calculates the overall score Score using the following formula, with the weight coefficient for the level of abstraction A set to 0.35, the weight coefficient for the degree of inference requirement I set to 0.35, the weight coefficient for the information density D set to 0.20, and the weight coefficient for the metacognitive word M set to 0.10. Score=0.35×A+0.35×I+0.20×D+0.10×M

[0115] For example, if the level of abstraction A = 0.78, the degree of inference requirement I = 0.81, the information density D = 0.68, and the metacognitive word M = 0.62, the distribution generation unit 14 calculates the overall score Score as Score = 0.35 × 0.78 + 0.35 × 0.81 + 0.20 × 0.68 + 0.10 × 0.62 = 0.755.

[0116] Next, the distribution generation unit 14 generates a cognitive depth probability distribution P(L|U) across multiple depth levels based on the calculated overall score Score. In the first embodiment, the multiple depth levels include, for example, a first depth level L1, a second depth level L2, a third depth level L3, a fourth depth level L4, and a fifth depth level L5.

[0117] The distribution generation unit 14 refers to the central value of each depth level stored in the memory unit 22 and calculates the weight of each depth level based on the difference between the overall score Score and the central value of each depth level. For example, the memory unit 22 stores 0.15, 0.35, 0.55, 0.75, and 0.92 as central values ​​corresponding to depth levels L1, L2, L3, L4, and L5, respectively. The distribution generation unit 14 calculates a larger weight for depth levels where the difference between the overall score Score and each central value is smaller.

[0118] The distribution generation unit 14 generates a cognitive depth probability distribution P(L|U) that includes the probability corresponding to each depth level by normalizing the weights of each calculated depth level. For example, if the overall score Score = 0.755 and the sensitivity coefficient α = 4.0, the distribution generation unit 14 generates a cognitive depth probability distribution P(L|U) in which depth levels L4 and L5 have relatively large probabilities. In this case, the distribution generation unit 14 generates the cognitive depth probability distribution P(L|U) as follows, for example, P(L1|U) = 0.04, P(L2|U) = 0.09, P(L3|U) = 0.20, P(L4|U) = 0.44, and P(L5|U) = 0.23.

[0119] Next, the depth determination unit 15 calculates the representative depth E[L] based on the perceived depth probability distribution P(L|U). In the first embodiment, the numerical values ​​corresponding to depth levels L1, L2, L3, L4, and L5 are set to 1, 2, 3, 4, and 5, respectively. In this case, the depth determination unit 15 calculates the representative depth E[L] by summing the products of the numerical values ​​corresponding to each depth level and the probability of that depth level.

[0120] For example, if the cognitive depth probability distribution P(L|U) is P(L1|U)=0.04, P(L2|U)=0.09, P(L3|U)=0.20, P(L4|U)=0.44, and P(L5|U)=0.23, the depth determination unit 15 calculates the representative depth E[L] as E[L]=1×0.04+2×0.09+3×0.20+4×0.44+5×0.23=3.73.

[0121] Next, the depth determination unit 15 determines the response depth L* based on the representative depth E[L]. For example, the depth determination unit 15 determines the response depth L1 when the representative depth E[L] is 1.0 or more and less than 1.5, the response depth L2 when it is 1.5 or more and less than 2.5, the response depth L3 when it is 2.5 or more and less than 3.5, the response depth L4 when it is 3.5 or more and less than 4.5, and the response depth L5 when it is 4.5 or more and 5.0 or less.

[0122] In the example above, since the representative depth E[L] is 3.73, the depth determination unit 15 determines the response depth L4 as the response depth L*. The depth determination unit 15 outputs the determined response depth L* to the control information generation unit 16. Note that the above method for determining the weight coefficient, center value, sensitivity coefficient, cognitive depth probability distribution, representative depth, and response depth is just an example and may be changed according to the set values, the number of depth levels, or the response depth determination rule stored in the memory unit 22.

[0123] Next, the control information generation unit 16 generates control information to be provided to the second large-scale language model 2 based on the response depth L* determined by the depth determination unit 15 (S109). The control information is information that causes the second large-scale language model 2 to generate a natural language response that is suitable for the user input. The control information includes, for example, a depth specification value indicating the response depth L*, a response specification corresponding to the response depth L*, a prompt for response generation, or a combination thereof.

[0124] For example, if L4 is determined as the response depth L*, the control information generation unit 16 generates the following information as control information: "L4: Responds by including explicit indication of the underlying structure, multi-stage reasoning, and attention to cognitive structure and causal relationships. It verbalizes the underlying structure, such as the connection between external evaluation and self-worth, not just the surface problem." Furthermore, the control information generation unit 16 generates a response generation prompt that includes user input, a specified depth L4, the control information mentioned above, and output conditions. The response generation prompt includes, for example, information instructing the second large-scale language model 2 to respond naturally in Japanese, without any excess or deficiency, in accordance with the specified depth.

[0125] The output unit 17 transmits the user input and the control information generated by the control information generation unit 16 to the second large-scale language model 2 (S110). In the first embodiment, the output unit 17 transmits, for example, the original input, the user input, a depth specification value indicating the response depth L*, and a response generation prompt to the second large-scale language model 2.

[0126] The second large-scale language model 2 generates a natural language response based on user input and control information received from the management server 10 (S111). The second large-scale language model 2 generates a natural language response that adjusts the descriptive granularity, level of abstraction, degree of structuring, number of inference levels, or specificity of advice of the response according to the response depth or response generation prompt included in the control information.

[0127] The second large-scale language model 2 returns the generated natural language response to the management server 10. The response acquisition unit 18 of the management server 10 receives the natural language response corresponding to the response depth from the second large-scale language model 2 (S112). The response acquisition unit 18 outputs the acquired natural language response to the response output unit 19.

[0128] The response output unit 19 transmits the natural language response acquired by the response acquisition unit 18 to the user terminal 1 (S113). This allows the user terminal 1 to present the natural language response generated by the second large-scale language model 2 to the user.

[0129] As described above, in the first embodiment, the management server 10 acquires user input and determines the level of abstraction, level of inference, information density, and metacognitive words based on the base values ​​of the level of abstraction and level of inference returned from the first large-scale language model 20, and the information density, metacognitive words, and correction values ​​calculated on the management server 10 side. The management server 10 then generates a cognitive depth probability distribution based on the determined multiple features, determines the response depth based on the cognitive depth probability distribution, and outputs control information based on the response depth to the second large-scale language model 2. This makes it possible to adapt the natural language response by the second large-scale language model 2 to the user input.

[0130] <5. Generation of cognitive depth probability distribution and determination of response depth> Next, we will explain the cognitive depth probability distribution generation and response depth determination process with reference to Figure 5. Figure 5 is a flowchart that shows the process of S108 shown in Figure 4 in more detail. Note that explanations that overlap with the content explained in Figure 4 will be omitted as appropriate.

[0131] As shown in Figure 5, the distribution generation unit 14 obtains the level of abstraction A, the inference requirement I, the information density D, and the metacognitive word M determined by the feature calculation unit 13. Here, the level of abstraction A and the inference requirement I are features determined based on the base values ​​of the level of abstraction and the inference requirement returned from the first large-scale language model 20 and the correction values ​​calculated by local auxiliary analysis within the management server 10. On the other hand, the information density D and the metacognitive word M are features calculated by local auxiliary analysis within the management server 10, and are not subject to numerical evaluation by the first large-scale language model 20.

[0132] Next, the distribution generation unit 14 retrieves the weight coefficients, the central values ​​of each depth level, the sensitivity coefficients, the generation formula for the cognitive depth probability distribution, and the normalization conditions stored in the memory unit 22. Depending on the method for generating the cognitive depth probability distribution, the distribution generation unit 14 may retrieve only a portion of this information.

[0133] The distribution generation unit 14 then calculates an overall score Score by weighting the acquired features A, I, D, and M. For example, the distribution generation unit 14 calculates the overall score Score using the formula Score = 0.35A + 0.35I + 0.20D + 0.10M, where the weight coefficient for abstraction level A is 0.35, the weight coefficient for inference requirement level I is 0.35, the weight coefficient for information density D is 0.20, and the weight coefficient for metacognitive word M is 0.10.

[0134] Next, the distribution generation unit 14 calculates the weight of each depth level based on the difference between the overall score Score and the central value of each depth level L1 to L5. For example, if the central value corresponding to each depth level Li is denoted as ci and the sensitivity coefficient as α, the distribution generation unit 14 calculates the weight wi of each depth level Li using the following formula. wi = exp(-α × |Score - ci|)

[0135] Here, the central value ci is, for example, 0.15 for L1, 0.35 for L2, 0.55 for L3, 0.75 for L4, and 0.92 for L5. The sensitivity coefficient α is, for example, 4.0. In this case, the closer the overall score Score is to the central value ci for each depth level, the larger the weight wi corresponding to that depth level becomes.

[0136] The distribution generation unit 14 generates a cognitive depth probability distribution P(L|U) by normalizing each calculated weight wi. For example, the distribution generation unit 14 calculates the probability P(Li|U) corresponding to each depth level Li using the following formula. P(Li|U)=wi / Σwj

[0137] This generates a cognitive depth probability distribution P(L|U) where the sum of the probabilities corresponding to each depth level L1 to L5 is 1. The cognitive depth probability distribution P(L|U) represents the probability that the user input U corresponds to each of the multiple depth levels L1 to L5.

[0138] The weight calculation formula, central value, and sensitivity coefficient described above are examples only and are not limited thereto. The distribution generation unit 14 may calculate the weights for each depth level based on an arbitrary function that assigns larger weights to each depth level as the difference between the overall score and the central value of each depth level decreases. For example, the distribution generation unit 14 may generate the cognitive depth probability distribution P(L|U) based on an exponential function, a Gaussian function, an inverse distance function, a linear function, a piecewise linear function, a sigmoid function, a softmax function, or a combination thereof.

[0139] Next, the depth determination unit 15 calculates a representative depth E[L] based on the cognitive depth probability distribution P(L|U), and determines the response depth L* based on the representative depth E[L]. The depth determination unit 15 outputs the determined response depth L* to the control information generation unit 16.

[0140] <6. Correspondence between response depth and response specifications> Next, the correspondence between response depth and response specification will be explained. Figure 6 is a table showing an example of the correspondence between response depth and response specification according to the first embodiment. Based on the response depth determined by the depth determination unit 15, the control information generation unit 16 refers to the correspondence shown in Figure 6 and generates a response specification or a response generation prompt as control information to be provided to the second large-scale language model 2.

[0141] As shown in Figure 6, the multiple response depths are divided into five stages, for example, from the first response depth L1 to the fifth response depth L5. Response depth L1 corresponds to the shallowest response depth, and response depth L5 corresponds to the deepest response depth. However, the number of response depth stages is not limited to five; any number of stages, two or more, may be used.

[0142] The response depth L1 is, for example, the response depth corresponding to a specific factual confirmation or a simple work instruction. The response specification corresponding to the response depth L1 may include presenting the conclusion concisely and keeping the explanation of reasons or background to the minimum necessary. For example, when the control information generation unit 16 determines the response depth L1, it may generate a response generation prompt that includes response specifications such as "answer concisely with the conclusion," "keep the explanation short," or "add only the minimum necessary supplementary information."

[0143] Response depth L2 is, for example, a response to a specific question that includes a basic summary of the situation or a summary of key points. The response specification corresponding to response depth L2 may include stating the conclusion and then briefly explaining the main points or basic reasons. For example, when the control information generation unit 16 determines the response depth L2, it may generate a response generation prompt that includes response specifications such as "stating the conclusion first and summarizing the main points," "briefly explaining the necessary reasons," or "summarizing the key points in bullet points."

[0144] Response depth L3 is the response depth corresponding to an input that requires an explanation of reasons, background circumstances, or causal relationships. The response specification corresponding to response depth L3 may include organizing the circumstances included in the user input and explaining the reasons, background, or causal relationships. For example, when response depth L3 is determined, the control information generation unit 16 may generate a response generation prompt that includes a response specification such as "organize and explain the background circumstances," "explain including reasons and causal relationships," or "basically compare from multiple perspectives."

[0145] Response depth L4 is a response depth that corresponds to an input that includes, for example, multiple points of discussion, implicit assumptions, emotional states, value judgments, or multiple stages of reasoning. The response specification corresponding to response depth L4 may include structuring the user input, reasoning step by step from multiple perspectives, and presenting solutions or decision-making information. For example, when response depth L4 is determined, the control information generation unit 16 may generate a response generation prompt that includes a response specification such as "structure and organize the situation," "explain the background circumstances, causal relationships, and options step by step," or "consider from multiple perspectives and present feasible solutions."

[0146] Response depth L5 is a response depth that corresponds to inputs such as highly abstract questions, consultations involving self-awareness, issues involving value judgments, or inputs requiring metacognitive organization. The response specification corresponding to response depth L5 may include organizing the structure, value judgments, self-awareness, or abstract framework behind the user input and supporting introspection or policy consideration as needed. For example, when response depth L5 is determined, the control information generation unit 16 may generate a response generation prompt that includes response specifications such as "organize the underlying value judgments or cognitive structure," "present perspectives to deepen self-awareness," or "present both an abstract framework and concrete action plans."

[0147] The control information generation unit 16 may output the response specification corresponding to the response depth directly to the second large-scale language model 2, or it may generate a response generation prompt including the response specification and output it to the second large-scale language model 2. For example, when the response depth L4 is determined, the control information generation unit 16 may output a response generation prompt to the second large-scale language model 2 that includes instructions such as "organize the situation, explain the background circumstances and causal relationships, and present step-by-step countermeasures," in addition to the user input.

[0148] Furthermore, the control information generation unit 16 may adjust the response specifications not only according to the response depth, but also according to the type of user input, user attributes, purpose of use, dialogue history, type of second large-scale language model 2, or service settings. For example, even with the same response depth L4, in a business support scenario, a response specification that emphasizes issue clarification, risk assessment, and execution procedures may be selected. On the other hand, in a consultation support scenario, a response specification that emphasizes emotional clarification, background information analysis, and step-by-step advice may be selected.

[0149] The correspondence shown in Figure 6 is just one example, and the correspondence between response depth and response specification is not limited thereto. For example, a response specification for each response depth may be defined as any combination including at least one of the following: level of explanation, level of abstraction, number of inference stages, degree of structuring, level of expertise, order of presentation of conclusions, presence or absence of explanation of reasons, presence or absence of examples, presence or absence of confirmation questions, or specificity of advice. The control information generation unit 16 may generate response specifications by referring to a depth correspondence table stored in the storage unit 22, or it may dynamically generate response specifications based on the response depth.

[0150] As described above, in the first embodiment, the control information generation unit 16 generates a response specification or a response generation prompt based on the response depth determined by the depth determination unit 15. This allows specifying the explanation granularity, degree of structuring, or number of inference stages to the second large-scale language model 2 according to the cognitive depth of the user input. As a result, the natural language response by the second large-scale language model 2 can be better adapted to the user input.

[0151] <7. Effects> As described above, in the information processing system according to the first embodiment, the management server 10 acquires user input and determines a plurality of features, including level of abstraction, degree of inference request, information density, and metacognitive words, based on the user input. The management server 10 then generates a cognitive depth probability distribution based on the determined plurality of features, determines the response depth based on the cognitive depth probability distribution, and outputs control information based on the response depth to the second large-scale language model 2. Therefore, the natural language response by the second large-scale language model 2 can be controlled according to the cognitive depth of the user input. As a result, the natural language response by the second large-scale language model 2 can be adapted to the user input.

[0152] Furthermore, in the information processing system according to the first embodiment, the abstraction level and the degree of inference required of the user input are included as features, so that the determination of the response depth can reflect whether the user input is closer to concrete fact-finding or closer to an abstract consultation, and how much inference is required to generate the response. Therefore, it is possible to generate a natural language response that corresponds to the depth of explanation or degree of inference required by the user input, rather than a response based solely on the superficial words or topics of the user input.

[0153] Furthermore, the information processing system according to the first embodiment includes the information density of user input and metacognitive words as features, so that the amount of information contained in the user input, the number of semantic units, and the degree of expression related to self-awareness, introspection, or thought processes can be reflected in the determination of response depth. Therefore, it becomes easier to distinguish between cases where the user input is merely a short question and cases where it is a consultation involving multiple circumstances, emotions, backgrounds, or self-awareness, and the natural language response can be better adapted to the user input.

[0154] Furthermore, in the information processing system according to the first embodiment, the feature calculation unit 13 determines the final features based on the base values ​​of the level of abstraction and the inference requirement returned from the first large-scale language model 20, and the information density, metacognitive words, rule evaluation values, or correction values ​​calculated on the management server 10 side. Therefore, flexible semantic evaluation by the first large-scale language model 20 and stable evaluation based on dictionaries, rules, or thresholds can be used in combination. As a result, the features of user input can be calculated more stably compared to the case where the system relies solely on the numerical evaluation results of the first large-scale language model 20.

[0155] Furthermore, in the information processing system according to the first embodiment, the level of abstraction and the degree of inference requirement are determined using numerical evaluation results from the first large-scale language model 20, while the information density and metacognitive words are calculated by local auxiliary analysis on the management server 10. Therefore, while the flexible evaluation capabilities of the first large-scale language model 20 are utilized for features requiring semantic judgment, features that can be easily calculated stably based on the number of words, the number of semantic units, or the dictionary match rate can be calculated on the management server 10. As a result, multiple features can be calculated efficiently and stably while minimizing dependence on external models.

[0156] Furthermore, in the information processing system according to the first embodiment, the feature calculation unit 13 can correct the base values ​​of the level of abstraction and the level of inference requirement returned from the first large-scale language model 20 using an abstract word dictionary, a concrete word dictionary, rules, thresholds, correction values, or correction rules. In addition, the feature calculation unit 13 can calculate information density and metacognitive words based on a metacognitive word dictionary, the number of words, the number of phrases, the number of semantic units, the number of characters, or the frequency of occurrence of metacognitive expressions. Therefore, words, symbols, phrases, semantic units, the number of characters, causal expressions, conditional expressions, or metacognitive expressions included in user input can be reflected in the calculation of features. As a result, contextual or semantic features of user input can be more easily reflected in the estimation of cognitive depth.

[0157] Furthermore, in the information processing system according to the first embodiment, the distribution generation unit 14 calculates an overall score by weighting the determined level of abstraction, inference requirement, information density, and metacognitive words, and generates a cognitive depth probability distribution based on the overall score. Therefore, it is possible to estimate the cognitive depth of user input not only by handling multiple features individually, but also based on an integrated index that reflects the contribution of each feature. As a result, the cognitive depth of user input can be handled in a multifaceted and quantitative manner.

[0158] Furthermore, in the information processing system according to the first embodiment, the distribution generation unit 14 generates a cognitive depth probability distribution spanning multiple depth levels. Therefore, instead of immediately classifying user input into a single depth level, it can be represented as a probabilistic spread across multiple depth levels. As a result, cognitive depth can be estimated more flexibly, for example, even when user input has intermediate or boundary-like characteristics.

[0159] Furthermore, in the information processing system according to the first embodiment, the depth determination unit 15 calculates a representative depth based on the cognitive depth probability distribution and determines the response depth based on the representative depth. Therefore, the cognitive depth probability distribution across multiple depth levels can be converted into a representative value that is easy to use for response generation control. As a result, the probabilistically estimated cognitive depth can be smoothly reflected in the response depth provided to the second large-scale language model 2.

[0160] Furthermore, in the information processing system according to the first embodiment, the representative depth may be an expected value calculated by summing the products of the numerical value corresponding to each depth level and the probability of each depth level. Therefore, not only the single depth level with the highest probability, but the entire probabilistic distribution across multiple depth levels can be reflected in the determination of the response depth. As a result, the cognitive depth of user input can be converted to the response depth more smoothly.

[0161] Furthermore, in the information processing system according to the first embodiment, the control information generation unit 16 generates control information to be provided to the second large-scale language model 2 based on the determined response depth. Therefore, it is possible to specify the granularity of the response explanation, the level of abstraction, the degree of structuring, the number of inference stages, the order in which conclusions are presented, or the specificity of the advice to the second large-scale language model 2. As a result, the natural language response of the second large-scale language model 2 can be controlled according to the cognitive depth of the user input.

[0162] Furthermore, in the information processing system according to the first embodiment, the control information generation unit 16 can generate a response specification or a response generation prompt by referring to the correspondence between the response depth and the response specification. Therefore, for example, a simple factual response can be generated for a shallow response depth, and a response including a summary of the situation, an explanation of the background circumstances, a summary of causal relationships, or step-by-step advice can be generated for a deep response depth. As a result, it becomes easier to select a response mode that matches the nature of the user input.

[0163] Furthermore, in the information processing system according to the first embodiment, the management server 10 intervenes between the user terminal 1 and the second large-scale language model 2, outputting user input and control information to the second large-scale language model 2. Therefore, it is possible to control response generation according to the cognitive depth of user input for an existing conversational artificial intelligence system without constructing the second large-scale language model 2 itself. As a result, natural language responses can be adapted to user input while utilizing an existing conversational artificial intelligence system.

[0164] Furthermore, in the information processing system according to the first embodiment, the response acquisition unit 18 acquires the natural language response generated by the second large-scale language model 2, and the response output unit 19 outputs the natural language response to the user terminal 1. Therefore, the management server 10 can perform a series of processes from acquiring user input to determining the response depth, generating control information, acquiring the natural language response, and outputting it to the user terminal 1. As a result, the user terminal 1 can be smoothly presented with a natural language response corresponding to the response depth.

[0165] As described above, according to the first embodiment, a cognitive depth probability distribution can be generated based on multiple features including the level of abstraction, the degree of inference requirement, the information density, and metacognitive words of the user input. Based on this cognitive depth probability distribution, the response depth can be determined, and control information based on the response depth can be output to the second large-scale language model 2. Therefore, the natural language response by the second large-scale language model 2 can be adjusted according to the cognitive depth of the user input, and a natural language response that is suitable for the user input can be generated.

[0166] In the above embodiment, an example was described in which the management server 10 intervenes between the user terminal 1 and the second large-scale language model 2, but the embodiment is not limited to this. All or part of the information processing system according to this embodiment may be configured integrally with an interactive artificial intelligence system or a service that provides said interactive artificial intelligence system.

[0167] For example, at least a portion of the input acquisition unit 11, feature calculation unit 13, distribution generation unit 14, depth determination unit 15, control information generation unit 16, response acquisition unit 18, and response output unit 19 may be implemented as internal functions of the interactive artificial intelligence system. In this case, the interactive artificial intelligence system may acquire user input, calculate multiple features based on the user input, determine a cognitive depth probability distribution and response depth, and generate a natural language response based on control information or response generation conditions corresponding to the response depth.

[0168] Furthermore, the functions of the management server 10 may be incorporated into the servers, applications, inference infrastructure, orchestration layer, prompt control unit, or response generation control unit that constitute the interactive artificial intelligence system. In this case, the configuration described as the management server 10 does not need to be a physically or logically independent server, but may be implemented as a single functional unit or processing module within the interactive artificial intelligence system.

[0169] Furthermore, while the above embodiment describes an example in which the management server 10 acquires the natural language response generated by the second large-scale language model 2 and outputs the natural language response to the user terminal 1, the embodiment is not limited to this. For example, the conversational artificial intelligence system or the service providing the conversational artificial intelligence system may directly output the generated natural language response to the user terminal 1. In this case, the management server 10 or a corresponding functional unit may output user input and control information to the conversational artificial intelligence system and not relay the natural language response itself.

[0170] Furthermore, the format of the control information is not limited to response generation prompts. The control information may consist of, for example, response specifications, response generation prompts, system messages, developer messages, parameters, tags, metadata, response templates, response policies, output constraints, generation conditions, or a combination thereof.

[0171] Furthermore, the control information may be directly input to the second large-scale language model 2, or it may be input to a prompt control unit, orchestration unit, response generation control unit, or inference control unit included in the interactive artificial intelligence system. In this case, the control information may be the final prompt itself given to the second large-scale language model 2, or it may be intermediate information for generating the final prompt.

[0172] Furthermore, control information may be input to the interactive artificial intelligence system via an application programming interface, communication interface, browser extension, plugin, middleware, external integration services, or a combination thereof. In this case, the control information may be sent as a single request together with user input, or it may be sent as separate parameters, header information, metadata, or configuration information from the user input.

[0173] Furthermore, in this specification, local auxiliary analysis within the management server 10 refers to analysis processing performed on the management server 10 side or within the same management environment, separate from the process of outputting user input to the first large-scale language model 20 and obtaining numerical evaluation results. Local auxiliary analysis may include, for example, processing to extract words, symbols, phrases, clauses, semantic units, character counts, conditional expressions, causal expressions, named entities, sentiment expressions, or metacognitive expressions contained in the user input, matching with dictionaries or rules, thresholding processing, normalization processing, weighting processing, correction value calculation processing, statistical processing, or a combination thereof.

[0174] Furthermore, the local auxiliary analysis may include analysis using a lightweight language model, classification model, embedding model, natural language processing model, or local large-scale language model, which is performed within the management server 10 or within the same management environment as the management server 10. In this case, the feature calculation unit 13 may calculate all or part of the level of abstraction, inference requirement, information density, and metacognitive words based on the evaluation results obtained by the local model, without using the numerical evaluation results returned from the first large-scale language model 20, or in combination with said numerical evaluation results.

[0175] Furthermore, the combination of features to be numerically evaluated by the first large-scale language model 20 and features calculated by the local auxiliary analysis is not limited to the examples above. For example, one, two, or three features from abstraction level, inference requirement level, information density, and metacognitive terms may be numerically evaluated by the first large-scale language model 20, and the remaining features may be calculated by the local auxiliary analysis. In this case, the features to be numerically evaluated by the first large-scale language model 20 are not limited to abstraction level and inference requirement level, but may also include information density or metacognitive terms. Alternatively, all of abstraction level, inference requirement level, information density, and metacognitive terms may be calculated by the local auxiliary analysis.

[0176] Furthermore, the combination of features used to generate the cognitive depth probability distribution is not limited to the examples above. The feature calculation unit 13 may calculate at least one of information density and metacognitive words as additional features, in addition to the two features including abstraction level and inference requirement. That is, the distribution generation unit 14 may generate the cognitive depth probability distribution based on three features: abstraction level, inference requirement, and information density, or it may generate the cognitive depth probability distribution based on three features: abstraction level, inference requirement, and metacognitive words. Alternatively, the distribution generation unit 14 may generate the cognitive depth probability distribution based on four features: abstraction level, inference requirement, information density, and metacognitive words.

[0177] In this case, the distribution generation unit 14 may change the weight coefficients depending on the type of features used to generate the cognitive depth probability distribution. For example, when using abstraction level A, inference requirement level I, and information density D, the distribution generation unit 14 may calculate the overall score using Score = wA × A + wI × I + wD × D. Alternatively, when using abstraction level A, inference requirement level I, and metacognitive word M, the distribution generation unit 14 may calculate the overall score using Score = wA × A + wI × I + wM × M. Here, wA, wI, wD, and wM are the weight coefficients of the corresponding features, and may be set or normalized depending on the combination of features used.

[0178] Furthermore, the feature calculation unit 13 may calculate at least some of multiple features based on symbols included in the user input. For example, the feature calculation unit 13 may supplementarily calculate the information density, inference requirement, strength of sentiment expression, or degree of metacognitive expression of the user input based on the number, position, or combination of occurrences of question marks, exclamation marks, parentheses, bullet points, quotation marks, separators, or emphasis symbols.

[0179] Furthermore, the feature calculation unit 13, the distribution generation unit 14, or the depth determination unit 15 may perform a predetermined fallback process if the user input is an exceptional input or a boundary input. Exceptional inputs or boundary inputs may include, for example, extremely short inputs, inputs with few semantic units, ambiguous inputs where multiple interpretation axes compete, noisy inputs including misspellings, omissions, colloquial fragments, or speech recognition errors, inputs with inconsistencies between multiple features, inputs where the cognitive depth probability distribution is widely distributed across multiple depth levels, or inputs where the representative depth is located near the boundary value of the response depth.

[0180] For example, if the user input is extremely short or the number of semantic units is less than a predetermined number, the feature calculation unit 13 may calculate a relatively low information density, and the depth determination unit 15 may prioritize an intermediate or conservative response depth. Also, if the user input is ambiguous and multiple interpretation axes conflict, the management server 10 may determine a confirmation intent to verify the interpretation axis and generate confirmation questions based on that confirmation intent. Furthermore, if the user input is noisy, the lightweight preprocessing unit 12 may perform feature calculation after correcting inconsistencies in notation, removing fillers, or correcting misspellings, and if it is difficult to restore the meaning, it may generate control information to request re-input.

[0181] Furthermore, if there is inconsistency between features, for example, if the level of abstraction is high while the information density is extremely low, or if the inference requirement is high while the metacognitive terminology is low, the feature calculation unit 13 or the depth determination unit 15 may prioritize the level of abstraction and the inference requirement as core features and treat the information density and metacognitive terminology as auxiliary features. Also, if the cognitive depth probability distribution is not sufficiently biased towards a particular depth level, or if the representative depth is near the boundary value of the response depth, the depth determination unit 15 may refer to the probability difference of adjacent depth levels in the cognitive depth probability distribution, perform reestimation as necessary, and adopt an intermediate or conservative response depth.

[0182] (Second embodiment) Next, an information processing system according to the second embodiment will be described. The information processing system according to the second embodiment differs from the first embodiment in that it determines the need for re-estimation of the natural language response generated by the second large-scale language model 2, readjusts the response depth as necessary, and regenerates control information. In addition, for the configuration of the information processing system according to the second embodiment, components similar to those in the first embodiment are denoted by the same reference numerals, and redundant explanations are omitted. In the following section, to explain the re-estimation process in the second embodiment, we will use an example based on user input different from the specific example described in the first embodiment.

[0183] Figure 7 is a diagram showing the processing sequence of the information processing system according to the second embodiment. In Figure 7, the processing from S101 to S109 described in the first embodiment is omitted, and the process from the point onward in which the management server 10 transmits the user input U and control information to the second large-scale language model 2 is shown. That is, in the second embodiment, similar to the first embodiment, feature quantities A, I, D, and M are determined based on the user input U, the overall score Score, the cognitive depth probability distribution P(L|U), the representative depth E[L], and the response depth L* are determined, and control information based on the response depth L* is generated.

[0184] In Figure 7, the processes of sending user input U and control information to the second large-scale language model 2, the process by which the second large-scale language model 2 generates a natural language response, and the process by which the management server 10 receives the natural language response are shown as S110, S111, and S112, respectively, as processes common to the first embodiment. On the other hand, in the second embodiment, after receiving the natural language response in S112, the processes from S201 onward are performed to determine the need for re-estimation, re-determine the response depth, and regenerate the natural language response.

[0185] In the following, in order to explain the re-estimation process in the second embodiment, we will use an example based on a different user input U than the specific example described in the first embodiment.

[0186] For confirmation, I will also explain S110 to S112. The output unit 17 of the management server 10 transmits user input U and control information to the second large-scale language model 2 (S110). The control information includes, for example, a depth specification value indicating the response depth L*, a response specification corresponding to the response depth L*, a response generation prompt, or a combination thereof, as described in the first embodiment. The second large-scale language model 2 generates a natural language response based on the user input U and control information received from the management server 10 (S111).

[0187] The response acquisition unit 18 of the management server 10 receives a natural language response corresponding to the response depth L* from the second large-scale language model 2 (S112). Hereinafter, the natural language response received from the second large-scale language model 2 may be referred to as the post-generation response R.

[0188] Next, the management server 10 determines the need for re-estimation of the post-generated response R (S201). The need for re-estimation is determined, for example, based on the degree of fit between the response depth range expected based on the user input U and the depth range of the actually generated post-generated response R. Specifically, the management server 10 generates a cognitive depth probability distribution PR(L|R) corresponding to the post-generated response R by applying the same feature calculation, overall score calculation, and cognitive depth probability distribution generation procedures used for the user input U to the post-generated response R.

[0189] The management server 10 calculates a consistency score based on the overlap between the cognitive depth probability distribution PU(L|U) generated based on the user input U and the cognitive depth probability distribution PR(L|R) generated based on the post-generated response R. For example, the management server 10 calculates the consistency score using the following formula. Consistency=Σi min(PU(Li|U),PR(Li|R))

[0190] Here, Li represents one of several depth levels, L1 to L5. The consistency score is a value in the range of 0.00 to 1.00, and a higher value indicates that the depth range of the response expected based on the user input U is consistent with the depth range of the actually generated post-generated response R.

[0191] For example, if the consistency score is less than 0.60, the management server 10 determines that there is a significant discrepancy between the expected response depth range for the user input U and the depth range of the generated response R, and determines that re-estimation is necessary. The management server 10 also determines that re-estimation is necessary if the difference between the representative depth EU[L] corresponding to the user input U and the representative depth ER[L] corresponding to the generated response R is greater than or equal to a predetermined threshold. For example, if |EU[L]-ER[L]| is 1.0 or greater, the management server 10 determines that the depth range assumed for the input and the actual response depth range are off by one level or more, and determines that re-estimation is necessary.

[0192] On the other hand, the management server 10 determines that re-estimation is unnecessary if the consistency score is 0.60 or higher and the difference in representative depth is less than a predetermined threshold. In this case, the management server 10 sends the post-generation response R received in S112 to the user terminal 1 as the final natural language response.

[0193] If S201 determines that reestimation is necessary, the management server 10 generates a reestimation prompt. The reestimation prompt is a prompt to check for any inconsistency between the generated response R and the expected response depth range based on the user input U. The reestimation prompt may include information instructing, for example, a re-evaluation from the abstract side, a re-evaluation from the concrete side, checking for any excess or deficiency in response depth, or checking whether the generated response R satisfies the response specification corresponding to the response depth L*. For example, if the user input U subject to the reestimation process includes an expression relating to the link between external evaluation and self-value, the reestimation prompt may include instructions to check whether the generated response R adequately handles that link.

[0194] At this time, the verification axes to be included in the re-estimation prompt are not determined autonomously by the first large-scale language model 20, but are determined by the management server 10 based on the consistency score, representative depth difference, inconsistency with the response specification, or the content of the post-generated response R. The first large-scale language model 20 performs a re-evaluation based on the verification axes determined by the management server 10.

[0195] The output unit 17 of the management server 10 sends the user input U, the post-generation response R, and the re-estimation prompt to the first large-scale language model 20 (S202). The re-estimation prompt includes, for example, the user input U, the initial response depth L*, the post-generation response R, and the perspective to be re-evaluated.

[0196] The first large-scale language model 20 re-evaluates the base values ​​of the level of abstraction and the level of inference requirement based on the user input U, the post-generation response R, and the re-estimation prompt received from the management server 10 (S203). Specifically, the first large-scale language model 20 refers to the user input U and the post-generation response R according to the confirmation axis included in the re-estimation prompt and numerically evaluates the re-evaluated base value Base_A for the level of abstraction and the re-evaluated base value Base_I for the level of inference requirement.

[0197] For example, if the reestimation prompt instructs "check whether the post-generated response R adequately handles the underlying structure of the user input U, and re-evaluate the level of abstraction and the degree of inference required," the first large-scale language model 20 generates re-evaluated Base_A and Base_I based on the extent to which the post-generated response R includes connections between external evaluation and self-value, background information, causal relationships, or multi-stage inference.

[0198] The first large-scale language model 20 returns the numerical evaluation results, including the re-evaluated Base_A and Base_I, to the management server 10. The feature calculation unit 13 of the management server 10 receives the numerical evaluation results of the re-evaluated Base_A and Base_I from the first large-scale language model 20 (S204). These numerical evaluation results are returned in, for example, JSON format or other machine-readable format.

[0199] Next, the management server 10 re-determines the overall score Score, cognitive depth probability distribution P(L|U), representative depth E[L], and response depth L* based on the re-evaluation results, information density D, metacognitive word M, and correction value Corr, and determines consistency (S205). Specifically, the feature calculation unit 13 calculates the re-evaluated abstraction level A by applying the abstraction level correction value Corr_A to the re-evaluated Base_A. The feature calculation unit 13 also calculates the re-evaluated inference requirement level I by applying the inference requirement level correction value Corr_I to the re-evaluated Base_I.

[0200] In this case, the information density D and metacognitive words M are the values ​​calculated by local auxiliary analysis within the management server 10, as in the first embodiment. That is, the management server 10 uses the information density D and metacognitive words M calculated based on the lightweight preprocessing results, dictionary, rules, thresholds, or correction rules. However, in another example, the management server 10 may recalculate the information density D and metacognitive words M based on the post-generation response R.

[0201] The distribution generation unit 14 calculates the overall score after re-evaluation by weighting the re-evaluated abstraction level A, re-evaluated inference requirement level I, information density D, and metacognitive word M. For example, the distribution generation unit 14 calculates the overall score after re-evaluation using the following formula. Score=0.35×A+0.35×I+0.20×D+0.10×M

[0202] The distribution generation unit 14 calculates the weight for each depth level based on the difference between the re-evaluated overall score Score and the central value of each depth level L1 to L5, and generates the re-evaluated cognitive depth probability distribution P(L|U) by normalizing these weights. The depth determination unit 15 calculates the re-evaluated representative depth E[L] based on the re-evaluated cognitive depth probability distribution P(L|U), and re-determines the response depth L* based on this representative depth E[L].

[0203] For example, if the re-evaluated features have an abstraction level A of 0.75, an inference requirement I of 0.79, an information density D of 0.66, and metacognitive words M of 0.45, the distribution generation unit 14 calculates the overall score after re-evaluation as Score = 0.35 × 0.75 + 0.35 × 0.79 + 0.20 × 0.66 + 0.10 × 0.45 = 0.716.

[0204] In this case, the distribution generation unit 14 generates, for example, P(L1|U)=0.05, P(L2|U)=0.12, P(L3|U)=0.23, P(L4|U)=0.37, and P(L5|U)=0.23 as probability distributions of cognitive depth after re-evaluation. The depth determination unit 15 calculates the representative depth E[L] after re-evaluation, using E[L]=1×0.05+2×0.12+3×0.23+4×0.37+5×0.23=3.61.

[0205] The depth determination unit 15 determines the consistency of the re-evaluation results based on the difference between the representative depth in the initial estimation and the representative depth after re-evaluation. For example, if the representative depth in the initial estimation is 3.68 and the representative depth after re-evaluation is 3.61, the difference is |3.68-3.61|=0.07. In this case, since the difference is within 0.15, the depth determination unit 15 determines that the initial estimation and the re-evaluation results are consistent.

[0206] If the re-evaluation results are determined to be consistent, the depth determination unit 15 adopts either the response depth based on the initial estimation or the response depth based on the representative depth after re-evaluation as the final response depth. For example, if both the response depth in the initial estimation and the response depth after re-evaluation are L4, the depth determination unit 15 adopts the response depth L4 as the final response depth L*.

[0207] On the other hand, if the difference between the representative depth in the initial estimation and the representative depth after re-evaluation exceeds 0.15, the depth determination unit 15 determines that the initial estimation and the re-evaluation result are inconsistent. In this case, the management server 10 performs one of the following actions: averaging the initial estimation and the re-evaluation result, obtaining additional information by presenting a verification axis, or adopting a conservative depth. For example, the management server 10 may determine the response depth L* based on the average value of the representative depth in the initial estimation and the representative depth after re-evaluation. Furthermore, if the management server 10 determines that the response is excessively abstract or excessively specific, it may adopt a conservative response depth to reduce the cognitive load on the user.

[0208] Next, the control information generation unit 16 generates regeneration control information based on the re-determined response depth L* (S206). The regeneration control information is control information for causing the second large-scale language model 2 to regenerate a natural language response, and includes, for example, a depth specification value indicating the re-determined response depth L*, a response specification corresponding to the response depth L*, a regeneration prompt, or a combination thereof.

[0209] For example, if the re-determined response depth L* is L4, the control information generation unit 16 generates regeneration control information that specifies the response to include explicit representation of the underlying structure, multi-stage reasoning, cognitive structure, or focus on causal relationships, as the response specification corresponding to response depth L4. The control information generation unit 16 may also include auxiliary instruction information that instructs the second large-scale language model 2 to supplement any missing perspectives detected in the post-generation response R.

[0210] The output unit 17 transmits the user input U and regeneration control information to the second large-scale language model 2 (S207). The second large-scale language model 2 regenerates a natural language response based on the user input U and regeneration control information received from the management server 10 (S208).

[0211] The response acquisition unit 18 receives the regenerated natural language response from the second large-scale language model 2 (S209). The response output unit 19 transmits the received regenerated natural language response to the user terminal 1 (S210). As a result, even if the initial response does not sufficiently match the response depth corresponding to the user input U, the information processing system according to the second embodiment can provide the user with a natural language response that better fits the user input U through reestimation and regeneration.

[0212] Furthermore, an upper limit may be set on the number of re-estimations. For example, the management server 10 may limit the number of re-estimations to two or fewer. If the consistency does not meet a predetermined standard even after the upper limit of re-estimations has been reached, the management server 10 will either adopt a conservative depth, average the initial estimation and the re-evaluation result, or generate an acknowledgment prompting the user to input additional information. This prevents the re-estimation process from being excessively repeated while improving the stability of the response depth.

[0213] In the second embodiment, the determination of the need for reestimation, the generation of prompts for reestimation, the analysis of re-evaluation results, the recalculation of the overall score, the regeneration of the cognitive depth probability distribution, the re-determination of the response depth, the consistency determination, and the generation of control information for regeneration are all performed on the management server 10 side. The first large-scale language model 20 is responsible for returning numerical evaluation results based on the verification axis defined by the management server 10, and the second large-scale language model 2 is responsible for generating or regenerating natural language responses based on the control information generated by the management server 10. Therefore, in the second embodiment as well, the estimation of cognitive depth, the determination of response depth, and the generation of control information are performed as processing on the system side, centered on the management server 10.

[0214] As described above, according to the second embodiment, the post-generation response R generated by the second large-scale language model 2 is checked for consistency with the expected response depth range based on the user input U, the response depth L* is re-determined as necessary, and control information is regenerated. Therefore, even if the post-generation response R is excessively shallow, excessively deep, or generated based on a different interpretation axis relative to the user input U, a natural language response that better fits the user input U can be regenerated.

[0215] Furthermore, according to the second embodiment, the management server 10 calculates a consistency score based on the overlap between the cognitive depth probability distribution PU(L|U) based on the user input U and the cognitive depth probability distribution PR(L|R) based on the post-generation response R. This allows for quantitative determination of the discrepancy between the expected response depth range for the user input U and the actual depth range of the generated post-generation response R. As a result, the management server 10 can control whether to output the generated post-generation response R directly to the user terminal 1 or to re-determine the response depth L* and regenerate a natural language response.

[0216] Furthermore, according to the second embodiment, averaging, acquisition of additional information, adoption of a conservative response depth, or regeneration can be performed based on the difference between the response depth after reestimation and the response depth before reestimation, the consistency score, or the explanatory granularity, level of abstraction, number of inference stages, or degree of structuring included in the generated response R. This stabilizes the determination of the response depth even when the user input U is ambiguous, includes multiple interpretation axes, or there is an inconsistency between the generated response R and the expected response depth range.

[0217] Furthermore, according to the second embodiment, when the management server 10 performs the acquisition of additional information by presenting confirmation axes, the management server 10 determines the confirmation intent or confirmation axis and causes the second large-scale language model 2 to translate the confirmation intent or confirmation axis into natural language. As a result, compared to the case where the second large-scale language model 2 arbitrarily generates confirmation questions, it is possible to generate confirmation questions that correspond to the location of ambiguity, inconsistency, or depth differences in the user input U, and the acquired additional information can be reflected in subsequent feature calculation, generation of cognitive depth probability distributions, or determination of response depth.

[0218] (Third embodiment) Next, with reference to Figure 8, the information processing system according to the third embodiment will be described. The information processing system according to the third embodiment differs from the first embodiment in that it determines the response depth L* using the level of abstraction A and the degree of inference requirement I of the user input. That is, in the third embodiment, the information density D and metacognitive word M from the first embodiment are not used, or the weights of information density D and metacognitive word M are set to 0, and the overall score Score, cognitive depth probability distribution P(L|U), representative depth E[L], and response depth L* are determined based on the level of abstraction A and the degree of inference requirement I.

[0219] In the third embodiment, the feature calculation unit 13 determines features A and I based on the base value of abstraction, Base_A, and the base value of inference request, Base_I, returned from the first large-scale language model 20. The feature calculation unit 13 may, if necessary, apply correction values ​​Corr_A and Corr_I calculated by local auxiliary analysis in the management server 10 to Base_A and Base_I.

[0220] For example, if the abstract word dictionary match rate is equal to or greater than a predetermined abstract word threshold, the feature calculation unit 13 calculates an abstraction level correction value Corr_A = +0.04 and calculates the abstraction level A using A = Base_A + Corr_A. Also, if the number of causal expressions or conditional expressions is equal to or greater than a predetermined number, the feature calculation unit 13 calculates an inference requirement correction value Corr_I = +0.05 and calculates the inference requirement I using I = Base_I + Corr_I.

[0221] The distribution generation unit 14 calculates an overall score Score based on the level of abstraction A and the inference requirement I. For example, the distribution generation unit 14 calculates the overall score Score using the following formula, with the weight coefficient for the level of abstraction A set to 0.50 and the weight coefficient for the inference requirement I set to 0.50. Score = 0.50 × A + 0.50 × I

[0222] For example, if the level of abstraction A = 0.78 and the inference requirement I = 0.81, the distribution generation unit 14 calculates the overall score Score as Score = 0.50 × 0.78 + 0.50 × 0.81 = 0.795.

[0223] The distribution generation unit 14 generates a cognitive depth probability distribution P(L|U) spanning multiple depth levels L1 to L5, similar to the first embodiment, based on the calculated overall score Score. The depth determination unit 15 calculates a representative depth E[L] based on the cognitive depth probability distribution P(L|U) and determines the response depth L* based on the representative depth E[L].

[0224] According to the third embodiment, compared to the first embodiment, the number of features to be calculated can be reduced, thus simplifying the feature calculation process, rule evaluation process, or correction process. Furthermore, since the level of abstraction A and the degree of inference requirement I are core features in estimating cognitive depth, even without using information density D and metacognitive words M, it is possible to determine the overall score Score, cognitive depth probability distribution P(L|U), representative depth E[L], and response depth L* according to user input. As a result, the management server 10 can generate control information to be given to the second large-scale language model 2 based on a relatively small number of features, and can generate natural language responses that are suitable for user input while suppressing processing load, dictionary or rule setting burden, and dependence on external models or local auxiliary analysis. [Explanation of symbols]

[0225] 1 user terminal 2. Second Large-Scale Language Model 10 Management Server 11 Input acquisition unit 12 Lightweight pre-processing 13 Feature Calculation Unit 13a Feature calculation prompt generation unit 13b Input Analysis Unit 13c Rule Evaluation Department 13d Numerical evaluation result acquisition unit 13e Integration and Correction Unit 13f Feature Determination Unit 14 Distribution generator 15 Depth determining section 16 Control Information Generation Unit 17 Output section 18 Response Acquisition Unit 19 Response Output Section 20. The First Large-Scale Language Model 22 Memory section A Level of abstraction I Reasoning requirements D Information density M Metacognitive words< / url:>

Claims

1. An input acquisition unit that acquires user input directed at an interactive artificial intelligence system, A feature calculation unit calculates a plurality of features, including the level of abstraction of the user input and the degree of inference required for the user input, based on the user input obtained. A distribution generation unit generates a cognitive depth probability distribution across multiple depth levels based on the multiple features calculated above, A depth determination unit that determines the response depth corresponding to the user input based on the generated cognitive depth probability distribution, A control information generation unit generates control information, which is information for causing the conversational artificial intelligence system to generate a natural language response that is suitable for the user input, based on the response depth determined above. An output unit that outputs the user input and the control information to the interactive artificial intelligence system, An information processing system characterized by comprising the following features.

2. The information processing system according to claim 1, characterized in that the plurality of features include at least one of the information density of the user input and the degree of metacognitive expression contained in the user input.

3. The information processing system according to claim 2, characterized in that the distribution generation unit weights the calculated feature quantities to calculate an overall score and generates the cognitive depth probability distribution based on the calculated overall score.

4. The information processing system according to claim 3, characterized in that the depth determination unit calculates a representative depth based on the cognitive depth probability distribution.

5. The information processing system according to claim 4, characterized in that the depth determination unit determines the response depth based on the calculated representative depth.

6. The information processing system according to claim 5, characterized in that the representative depth is an expected value calculated by summing the products of numerical values ​​corresponding to the plurality of depth levels and the probabilities of each depth level in the cognitive depth probability distribution.

7. The information processing system according to claim 6, characterized in that the feature calculation unit outputs a feature calculation prompt including the user input to a large-scale language model, and calculates at least a portion of the plurality of features based on the numerical evaluation results returned from the large-scale language model.

8. The information processing system according to claim 6, characterized in that the feature calculation unit calculates at least a portion of the plurality of features based on the words, symbols, phrases, semantic units included in the user input, or the number of characters in the user input.

9. The information processing system according to claim 6, characterized in that the feature calculation unit outputs a feature calculation prompt including the user input to a large-scale language model, and calculates at least a portion of the plurality of features based on the numerical evaluation result returned from the large-scale language model and the words, symbols, phrases, semantic units, or the number of characters in the user input.

10. On the computer, Input acquisition processing to obtain user input directed at an interactive artificial intelligence system, A feature calculation process that calculates multiple features, including the level of abstraction of the user input and the degree of inference required for the user input, based on the user input obtained, A distribution generation process that generates a cognitive depth probability distribution across multiple depth levels based on the multiple features calculated above, A depth determination process that determines the response depth corresponding to the user input based on the generated cognitive depth probability distribution, A control information generation process generates control information, which is information for causing the conversational artificial intelligence system to generate a natural language response that is suitable for the user input, based on the response depth determined above. A program characterized by causing the interactive artificial intelligence system to perform an output process that outputs the user input and the control information.

11. Computers An input acquisition step that obtains user input directed at an interactive artificial intelligence system, A feature calculation step that calculates multiple features including the level of abstraction of the user input and the degree of inference required for the user input based on the acquired user input, A distribution generation step that generates a cognitive depth probability distribution across multiple depth levels based on the multiple features calculated above, A depth determination step in which the response depth corresponding to the user input is determined based on the cognitive depth probability distribution generated above, A control information generation step, which generates control information that causes the conversational artificial intelligence system to generate a natural language response that is suitable for the user input, based on the response depth determined above, An information processing method characterized by performing an output step of outputting the user input and the control information to the interactive artificial intelligence system.

Images