Decision support system, decision support program, and decision support method

Abstract

We use large-scale language models to support decision-making tailored to the specific user's situation. [Solution] A decision support system comprising: an inference engine that performs natural language processing; a reception module that receives text information input from a user; an input analysis module that analyzes the semantic weights of the input text information; a dynamic state management module that transitions the system's operating mode based on the analyzed semantic weights; and a constraint control module that determines and dynamically rewrites constraint conditions to be preferentially applied to the inference engine according to the transitioned operating mode, wherein the inference engine determines the search space for the response and / or adjusts the output filter strength of the response to generate a response according to the rewritten constraint conditions.

Description

【Technical Field】 【0001】 The present disclosure relates to a decision-making support system using a large language model. 【Background Art】 【0002】 With the recent development of artificial intelligence technology, large language models (hereinafter referred to as "LLMs") have shown remarkable performance improvements in the field of natural language processing and are used in various applications such as dialogue systems, document generation, summarization, and inference support. However, since LLMs are generally probabilistic generation models and their output content largely depends on the prompts input from users and their internal states, they have inherent problems such as unintended responses, opacity of the inference process, and lack of consistency and reproducibility. To address such problems, various architectures and methods for assisting the inference process and output control of LLMs have been proposed. For example, techniques for providing auxiliary processing structures either externally or internally to LLMs for the purpose of improving the accuracy and stability of inference are known. For example, as an example, a technique for performing context understanding and judgment assistance using a multi-agent configuration is known, and attempts have been made to realize support considering the user's situation (Non-Patent Document 1). In addition, a method has been proposed in which an LLM generates or maintains an internal inference process (so-called thinking process) without explicitly outputting it externally and reflects it in the generation of the final response (Non-Patent Documents 2 and 3). In addition, in order to improve the response quality and safety of LLMs, a reinforcement learning method (Reinforcement Learning from Human Feedback; RLHF) that reflects the evaluation results by humans in learning is widely used (Non-Patent Document 4). RLHF is effective in internalizing general response tendencies and safety guidelines in the model and plays an important role in many practical systems. 【Prior Art Documents】 【Non-Patent Documents】 【0003】 [Non-licensed document 1] Sammy Kao, “Engineering Challenges for HelperU: Building a Context-Aware Multi-Agent AI System”, Medium.com20140046891A1 [URL] https: / / medium.com / @sammy_kao / engineering-challenges-for-helperu-building-a-context-aware-multi-agent-ai-system-3b50e028afda (last viewing date: February 26, 2026) 【0004】 【Non-licensed Document 2】 Hikaru Asano et al, “Self Iterative Label Refinement via Robust Unlabeled Learning”, arXiv[URL] https: / / arxiv.org / html / 2502.12565v1 (Last viewing date: February 26, 2026) 【0005】 [Non-licensed document 3] Terrence J. Sejnowski et al, "Large Language Models and the Reverse Turing Test", MIT Press Direct[URL]https: / / direct.mit.edu / neco / article / 35 / 3 / 309 / 114731 / Large-Language-Models-and-the-Reverse-Turing-Test (last viewing date: February 26, 2026) 【0006】 【Non-licensed Document 4】 Timo Kaufmann et al, “A Survey of Reinforcement Learning from Human Feedback”, arXiv [URL] https: / / arxiv.org / abs / 2312.14925 (Last accessed: February 26, 2026) [Overview of the Initiative] [Problems that the invention aims to solve] 【0007】 However, technologies like those described in Non-Patent Document 1 still have challenges in terms of inference control within the model and flexible extensibility. Furthermore, the technologies described in Non-Patent Documents 2 and 3 have the problem that their internal processing is highly dependent on the model, making general application across different LLMs and control from external systems difficult. RLHF, such as that described in Non-Patent Document 4, is a method primarily applied during the model learning phase; therefore, the response strategy acquired by RLHF is statistically optimized in advance, optimizing the behavior of the entire model on average. In other words, it has the problem of not being able to provide a meaningful response when making decisions tailored to the specific user situation. Furthermore, most currently known LLMs are designed to comply with restrictions based on individual prohibited expressions and safety standards for the output responses, which may also prevent them from tailoring responses to the specific user's situation. [Means for solving the problem] 【0008】 The inventors have made diligent efforts to solve the above problems and have completed the invention described herein. This disclosure encompasses the following technical ideas. [1] A decision support system, An inference engine that performs natural language processing, A reception module that accepts text information input from the user, An input analysis module that analyzes the semantic weights of the input text information, Based on the analyzed semantic weights, a dynamic state management module is used to transition the operating mode of the system. A constraint control module that determines and dynamically rewrites the constraints to be preferentially applied to the inference engine according to the transitioned operating mode, Equipped with, The inference engine determines the search space for the response and / or adjusts the output filter strength of the response in accordance with the rewritten constraints to generate the response. A decision support system characterized by the following features. [2] The decision support system according to [1], wherein the semantic weight is an index calculated based on one of the group consisting of urgency, emotional polarity, strategic consistency and risk intensity of the user's situation. [3] The decision support system according to [1] or [2], characterized in that the dynamic state management module transitions the operating mode of the system to either a first mode that prioritizes responses related to the internal environment of the organization to which the user belongs, or a second mode that prioritizes responses related to the external environment of the user. [4] The decision support system according to any one of [1] to [3], wherein the dynamic state management module further has a third mode that prioritizes pointing out logical inconsistencies in the user's input. [5] The decision support system according to any one of [1] to [4], further comprising a missing information determination module that determines whether the analyzed semantic weights satisfy the information necessary for the transition of the operating mode by the dynamic state management module. [6] The inference engine comprises a first generation module that generates first output data in which alignment processing is not intentionally applied based on the rewritten constraints when generating a response, and a second generation module that takes the first output data as input and generates second output data in which alignment processing is applied while maintaining its logical structure, A decision support system according to any one of [1] to [6], characterized in that these are operated sequentially and the second output data is output as a response of the inference engine. [7] The first generation module further includes an alignment determination module that determines whether an inference process in which the alignment process is intentionally not applied is being properly executed, and then controls the operation of the second generation module. The decision support system according to [5], characterized in that if it is determined that the processing by the first generation module is insufficient, the system controls the system to execute the processing by the first generation module again. [8] The decision support system according to [6] or [7], wherein both the first output data and the second output data are output when the output is performed. [9] The inference engine includes a first generation module that generates first output data in which alignment processing is intentionally not applied based on the rewritten constraints when generating a response, The decision support system according to any one of [1] to [5], characterized in that the dynamic state management module transitions the operating mode of the system to a fourth mode in which the first output data is output as a response of the inference engine.

[10] The system includes a convergence determination module that performs a convergence determination process to determine whether the output generated by the inference engine has a sufficient degree of fit to the objective function defined in the operating mode, A calibration module controls the system to perform calibration processing based on the results of the convergence determination process, Furthermore, If the convergence determination module determines that the degree of fit is less than the threshold, the output from the inference engine is held in abeyance. The calibration module is characterized by autonomously generating and presenting a calibration process that involves generating and presenting a question text or a suggested revision of the text information, which requests the user to input information that is missing in order for the degree of fit to be equal to or greater than the threshold. The decision support system according to any one of [1] to [9], characterized in that the convergence determination process and the calibration process are repeatedly performed until the degree of fit becomes equal to or greater than the threshold.

[11] The system includes a memory database that holds input history and context information for the inference engine, The system further includes a memory management module for managing the aforementioned memory, The memory management module controls the memory to keep, as active memory, memory that is highly relevant to the past or most recent input history and its contextual information in the active area of ​​the memory database, and to keep, as idle memory, memory that is less relevant in the idle area of ​​the memory database in an indexed state with compressed data volume. The decision support system according to any one of [1] to

[10] , characterized in that the memory management module performs dynamic swap control triggered by an update in the input content to the inference engine, which compresses the active memory that is less relevant to the updated input content and saves it to a pause area and converts it to pause memory, or decompresses or retrieves the pause memory that is more relevant to the updated input content and returns it to the active area and converts it to active memory.

[12] The decision support system according to

[11] , wherein when the dynamic swap control restores the pause memory to the active area, it is performed by detecting index data assigned to the pause memory or by matching the input vector of the updated input content with the index vector of the pause memory.

[13] The decision support system according to

[11] or

[12] , wherein the dynamic swap control is triggered when the input to the inference engine is updated by user input.

[14] The decision support system according to any one of

[11] to

[13] , wherein the dynamic swap control is triggered when the input content is autonomously updated by any one module selected from the group consisting of the dynamic state management module, the constraint control module, the missing content determination module, the first generation module, the alignment determination module, and the calibration module.

[15] Computers, The natural language processing function performed by the inference engine, A reception function for receiving input of text information from a user, An input analysis function for analyzing the semantic weight of the input text information, A dynamic state management function for transitioning the operation mode of the system based on the analyzed semantic weight, A constraint control function for determining and dynamically rewriting the constraint conditions preferentially applied to the inference engine according to the transitioned operation mode, To execute, The inference engine determines the search space of the response and / or adjusts the output filter strength of the response according to the rewritten constraint conditions to generate a response, A decision-making support program characterized by the above.

[16] A decision-making support method, A reception step of receiving input of text information from a user, An input analysis step of analyzing the semantic weight of the input text information, A dynamic state management step of transitioning the operation mode of the system based on the analyzed semantic weight, A constraint control step of determining and dynamically rewriting the constraint conditions preferentially applied to the inference engine according to the transitioned operation mode, A step of causing the inference engine to perform response generation by natural language processing, Including, The inference engine determines the search space of the response and / or adjusts the output filter strength of the response according to the rewritten constraint conditions to generate a response, A decision-making support method characterized by the above.

[17] A non-transitory computer-readable recording medium, When executed by a processor, the processor is caused to, 1) A step of receiving input of text information from a user, 2) A step of analyzing the semantic weight of the input text information, 3) A step of transitioning the operation mode of the system based on the analyzed semantic weight, 4) A step of determining and dynamically rewriting the constraints to be preferentially applied to the inference engine according to the transitioned operating mode, 5) A step of determining the search space for the response and / or adjusting the output filter intensity of the response according to the rewritten constraints, 6) A process of generating a response using an inference engine, A non-temporary, computer-readable recording medium containing a decision support program for executing a certain action.

[18] An information processing system for decision support, At least one processor, At least one memory that stores instructions that cause the system to perform the following when executed by the at least one processor: Equipped with, The aforementioned instruction is given to the system, Receiving text information from users via the reception module, To analyze the semantic weights of the received text information, Transitioning the operating mode of the system based on the analyzed semantic weights, Depending on the transitioned operating mode, determine and dynamically rewrite the constraints that are preferentially applied to the inference engine, and An information processing system that causes the inference engine to determine the search space for the response and / or adjust the output filter strength based on the rewritten constraints, thereby generating a response.

[19] A non-temporary computer-readable medium that stores instructions that cause one or more processors to perform the following when executed by one or more processors, Receiving text information from users, Analyzing the semantic weights of the aforementioned text information, Transitioning the operating mode based on the aforementioned semantic weights, Dynamically determining and rewriting the constraints applied to the inference engine, and Based on the rewritten constraints, control at least one of the response search space and output filter intensity of the inference engine. A non-temporary computer-readable medium that enables execution. [Effects of the Invention] 【0009】 According to the decision support system disclosed herein, in the actual operation of LLM by users, the system can analyze user input using natural language processing and dynamically consider urgent elements and emotionally sensitive situations contained in the input, thereby enabling the generation of flexible responses tailored to the individual circumstances of diverse users. 【0010】 Furthermore, in some embodiments, the inference engine of the decision support system of this disclosure can switch constraints based on semantic weights obtained from input analysis, thereby enabling appropriate decision support according to the user's state. 【0011】 Furthermore, in some embodiments, the decision support system of this disclosure includes a configuration that determines the search space for the response in the inference engine and / or adjusts the output filter strength by dynamically rewriting the constraint conditions. This makes it possible to control the response generation process itself without relying solely on post-processing of the response content. As a result, it is possible to suppress information loss due to excessive filtering and, conversely, the occurrence of inappropriate responses due to insufficient constraints, thereby achieving both response quality and safety. 【0012】 The decision support system described herein achieves context-adaptive constraint control during inference without relying on static constraint design of a pre-trained model, and can provide stable decision support even for diverse user states that fluctuate in real-world operating environments. Furthermore, the decision support system described herein does not merely present decision content, but rather controls the search space and output filter strength in the inference process by transitioning the system's internal operating mode based on semantic weights and dynamically rewriting the constraint conditions applied to the inference engine according to the operating mode. As a result, the parameters that the inference engine internally references (e.g., token generation probability distribution, evaluation parameters, or response candidate selection criteria) are dynamically changed, and the processing path and the manner in which computing resources are used change substantially. Therefore, this disclosure concerns information processing technology that technically controls the internal operation of a computer system, rather than the application of abstract decision rules. [Brief explanation of the drawing] 【0013】 [Figure 1] This diagram illustrates the functional blocks that make up a decision support system. [Figure 2] This diagram illustrates the relationships between the various operating modes that can be transitioned to in the dynamic state management module. It shows that each operating mode can dynamically transition to another. [Figure 3] This figure illustrates a functional block in a different embodiment of the decision support system than that shown in Figure 1. [Figure 4] This diagram illustrates the flow from receiving text information entered by the user to the control operation performed by the constraint control module. [Figure 5] This diagram illustrates the process by which the inference engine creates and outputs candidate response data. [Figure 6] This diagram illustrates the flow of the calibration loop using the convergence determination module and the calibration module. [Figure 7]This diagram illustrates the flow of dynamic swap control performed on memory within the memory database by the memory management module. An example is shown where dynamic swap processing is performed on active memory 712a. [Figure 8] This diagram illustrates an example of input and output to this system. It shows an example where, as a result of input analysis, the system transitions to the first mode (internal analysis mode) and a response is generated. [Figure 9] This diagram illustrates an example of input and output to this system. It shows an example where, as a result of input analysis, the system transitions to the second mode (external threat mode) and a response is generated. [Figure 10] This diagram illustrates an example of input and output to this system. It shows an example where the system transitions to the third mode (self-critical mode) and generates a response as a result of input analysis. [Figure 11] This diagram illustrates an example of input and output to this system. It shows an example where the input analysis results in a transition to the fourth mode (chaos mode) and a response is generated. [Modes for carrying out the invention] 【0014】 [Overall System Configuration] The decision support system 10 described herein comprises an inference engine 20 that performs natural language processing, and a plurality of modules that control the input, operating mode, and constraints for the inference engine. Furthermore, the decision support system 10 described herein functions as control middleware that analyzes text information input by the user and dynamically changes the response generation conditions of the inference engine 20 during inference based on the text information. 【0015】 This system can be implemented using an information processing device that includes one or more processors, main memory (RAM), auxiliary storage (ROM, SSD, etc.), and a network interface. Each module described below can be implemented as a program executed by this processor. The program can be recorded on a hard disk, SSD, optical disk, semiconductor memory, or a recording medium provided via a network. The inference engine of this system may be implemented on a local device or on an external server connected via a network. The following describes several modules that the decision support system according to this disclosure includes. Each module may be implemented on a single device or distributed across multiple devices. 【0016】 [Reception module and input analysis module] The reception module 100 of this system 10 receives text information input in natural language from the user and supplies the text information to the inference engine 20 and the subsequent input analysis module 200. 【0017】 The input analysis module 200 applies natural language processing to the text information received by the reception module 100 and analyzes the semantic weights of the text information. In other words, the input analysis module 200 may include the steps of receiving text information from the reception module and calculating or correcting semantic weights from the user's input. This enables the inference engine 20 to behave more flexibly and appropriately, reflecting the user's state. 【0018】 Natural Language Processing (NLP) refers to information processing technology for analyzing, understanding, generating, or transforming text or speech written in natural languages ​​(e.g., Japanese, English, etc.) that humans use in everyday life. Natural Language Processing may include multiple processing steps such as morphological analysis, syntactic analysis, semantic analysis, and contextual analysis, and these processes can be performed individually or in combination in the input analysis module 200 of the decision support system 10. 【0019】 Semantic weight refers to an index that evaluates or quantifies user-inputted text information based on arbitrary evaluation axes or perspectives. In natural language processing, semantic weight can also be expressed as a numerical or index representing the semantic importance or semantic contribution of a word, phrase, sentence, or document. This semantic weight reflects the extent to which the target word or sentence contributes to the overall context, subject, or purpose, and can be used for comparison, ranking, selection, or judgment. Its calculation may be based, for example, on the application frequency of the target word, its position within phrases or sentences, the distance or similarity between vector representations of words or sentences, weight coefficients calculated by attention processing, or output values ​​from other classification or known regression models. Furthermore, the analysis of semantic weight may include numerical vector transformation processing by a natural language processing model. As an example, though not particularly limited in this specification, semantic weight may be expressed as a single numerical value, a multi-dimensional vector, a probability value, a normalized value, or a set of scores. Furthermore, semantic weights can be calculated based on different units, scales, or criteria, and a configuration in which they are treated as relative values ​​rather than absolute values ​​can also be adopted. In this specification, they are preferably generated as numerical values, probability values, or vector representations and used in subsequent control processing. 【0020】 More specifically, semantic weight can be defined as an index calculated based on one of the following: urgency of the user's situation, emotional polarity, strategic consistency, and risk intensity. Each is explained below. 【0021】 The urgency of a user's situation refers to the degree to which the text information entered by the user necessitates immediate action or priority processing based on time constraints. This urgency can be estimated by analyzing the text information using NLP. For example, it can be calculated from words indicating time urgency such as "urgent" or "immediately," stylistic features such as the text information being in the imperative form or consisting of a series of short sentences, frequent use of repetitive expressions, the appearance of expressions indicating specific times or deadlines, and the level of urgency based on comparison with past input history. In this system, the urgency of a situation can be expressed as a numerical value, rank, category, or probability value. 【0022】 Emotional polarity refers to an index that indicates whether the emotional expressions and tendencies contained in text information entered by a user are positive, neutral, or negative, as well as their intensity. Such emotional polarity can be estimated by analyzing text information using NLP (Neuro-Linguistic Programming). For example, it can be expressed as one of three classifications: "positive" or "bold," "negative" or "negative," or "neutral," as well as as a classification based on scoring and thresholds based on an arbitrary evaluation axis, or as a multidimensional representation that comprehensively analyzes multiple emotional axes expressing the user's joy, anger, sadness, and other emotions. 【0023】 Strategic alignment refers to an indicator that shows the extent to which the content of text information entered by users aligns with the strategic objectives, business policies, decision-making assumptions, etc., set by the user themselves or by the organization to which the user belongs. For example, if the text includes statements aimed at maximizing short-term profits; statements aimed at establishing a medium- to long-term market position; or statements aimed at maintaining and defending existing businesses, the degree to which these statements coincide with or deviate from predefined strategic objectives can be evaluated. 【0024】 Risk intensity refers to the extent to which the content of text information entered by users contains risk factors such as uncertainty, loss, and legal or social impact that may arise in connection with decision-making. More specifically, risk factors include, but are not limited to, economic loss, competitive disadvantage, reputational risk, compliance concerns, and impact on business continuity. 【0025】 Furthermore, the urgency, emotional polarity, strategic consistency, and risk intensity of the situations described above are not limited to a single, essential element in this system, but may be used in combination with other indicators of user status. 【0026】 The input analysis module 200 can dynamically analyze and adjust the semantic weights of the user's text information (examples include, but are not limited to, word, phrase, and sentence divisions) received by the reception module 100. Specifically, if the input analysis module 200 estimates that the situation is highly urgent, high semantic weights may be assigned to the actions, requests for answers, the content of the problems raised, or the information being analyzed that are included in the input. Furthermore, if there is a large amount of negative input in terms of emotional polarity, high semantic weights may be assigned to and emphasize the locations where problems have occurred, expressions that may be the cause, or expressions that indicate the user's anger, sadness, anxiety, or dissatisfaction. 【0027】 [Dynamic State Management Module] The dynamic state management module 300 of this system has the function of dynamically transitioning the overall operating mode (State) of the system to one of several modes based on semantic weights analyzed by the input analysis module 200, contextual information of user input, and system setting information. 【0028】 In some embodiments, the dynamic state management module 300 can transition the operating mode of the system to either the following first mode 310 or second mode 320. [First mode and second mode] In some embodiments, the dynamic state management module may have at least two operating modes. (1) A first mode 310 that prioritizes responses related to the internal environment of the organization to which the user belongs. (2) Second mode 320 that prioritizes responses regarding the user's external environment. For example, if the input analysis module 200 detects inputs related to internal business issues or decision-making within an organization, semantic weights prioritizing these inputs are calculated, and if these weights exceed a pre-defined threshold, the system transitions to the first mode 310. This process is also specifically referred to as the first transition process (S310). Furthermore, if the input analysis module 200 detects inputs related to market trends, competitors, or external factors, or if semantic weights prioritizing these inputs are calculated and these weights exceed a pre-defined threshold, the system transitions to the second mode 320. This transition process is also specifically referred to as the second transition process (S320). As an example, though not limited to the most extreme, the determination of these operating modes may be performed by a score calculation based on the inner product of a semantic vector generated from the input text information and a weight vector corresponding to each mode. In this case, the corresponding operating mode is selected if the score exceeds a predetermined threshold. 【0029】 The first mode 310, also known as the internal analysis mode, prioritizes the analysis of information regarding the internal operational status, problem structure, decision-making processes, resource allocation, or internal risks of the organization, and generates and controls responses accordingly. When the dynamic state management module 300 transitions the system to the first mode 310, the subsequent constraint control module 400 applies the constraints pre-configured for the first mode 310 to the inference engine 20. 【0030】 The second mode 320, also known as the external threat mode, analyzes threats, competitors, or third-party actions present in the external environment of the user's organization. Specifically, it prioritizes analysis of information regarding competitor activities, market conditions, industry structure, technological trends, regulatory trends, or countermeasures against third-party actions, and generates and controls responses accordingly. When the dynamic state management module 300 transitions the system to the second mode 320, the subsequent constraint control module 400 applies pre-configured constraints for the second mode 320 to the inference engine 20. 【0031】 The transitions in operating modes by the dynamic state management module 300 are not permanently or exclusively fixed after a transition, but may be dynamically switched in response to additional input from the user or updates to the input content based on the autonomous inference behavior of the inference engine 20. Furthermore, while the transitions in operating modes by the dynamic state management module 300 control the behavior of the entire system, multiple modes may be applied stepwise or partially in a series of response generation processes. 【0032】 In this specification, "dynamic" means not based on pre-set fixed conditions, but rather in a time-series or sequential manner in response to the content of text information input by the user and the autonomous behavior of the inference engine 20. For example, based on semantic weights calculated time-series or sequentially by the input analysis module 200, the dynamic state management module 300 dynamically updates the system's internal state, operating mode, or constraints. Furthermore, the constraint control module 400, described later, also determines the constraints applied to the inference engine 20 in response to the operating mode transitioned by the dynamic state management module 300, and dynamically rewrites those conditions. Rewriting here may include, but is not limited to, switching the presence or absence of constraints, changing the priority of multiple constraints, or changing the intensity of constraint application. 【0033】 [Third Mode] In some embodiments, the system may further include the following third mode 330 as an operating mode. (3) Third mode 330 which prioritizes pointing out logical inconsistencies in the user's input. This operating mode is also referred to as the logical contradiction detection mode, self-criticism mode, or red team simulation mode. In this mode, when generating a response from the inference engine 20, priority is given to detecting consistency between preconditions, the presence or absence of logical leaps in the inference process, or the existence of mutually contradictory claims, rather than the validity of the conclusion. The generation and control of the response are then performed accordingly. The process of transitioning to this operating mode is specifically referred to as the third transition process (S330). When the dynamic state management module 300 transitions the system to the third mode (S330), the subsequent constraint control module 400 applies the constraint conditions pre-set as the third mode 330 to the inference engine 20. 【0034】 In some embodiments, the third mode 330 functions as an operating mode based on a different perspective from the first mode 310 and / or the second mode 320. While the first mode 310 and / or the second mode 320 are operating modes primarily for switching the perspective of the analysis, the third mode 330 is based on a perspective for verifying the consistency and validity of the user's input and the validity of the reasoning. The third mode 330 is not limited to being applied alone or exclusively with the first mode 310 and / or the second mode 320, but may be applied in combination with them. When applied in combination, for example, it can be used to verify whether the reasoning process of the response generated by the first mode 310 and / or the second mode 320 contains any logical inconsistencies. 【0035】 By incorporating the dynamic state management module 300 in this way, the decision support system 10 of this disclosure can not only switch analytical perspectives according to the circumstances of individual users, but also dynamically verify the logical soundness of the reasoning process. The dynamic state management module 300 can be an important function for increasing the reliability and persuasiveness of the output content of the decision support system 10. 【0036】 [Mode 4] In some embodiments, the system may further include the following fourth mode 340 as an operating mode. (4) A fourth mode in which the first output data 501, described later, is output directly as the response of the inference engine 20. This operating mode is also referred to as chaos mode, and in generating the inference engine's response, it prioritizes outputting the response data created without intentionally performing alignment processing (S530) on the user's input. The process of transitioning to this operating mode is specifically called the fourth transition process (S340). When the dynamic state management module 300 transitions the system 10 to the fourth mode 340, the subsequent constraint control module 400 applies the constraint conditions pre-set for the fourth mode 340 to the inference engine 20. 【0037】 In some embodiments, the fourth mode 340 functions as an operating mode based on a different perspective from the first mode 310, the second mode 320, and / or the third mode 330. That is, when it is determined that a serious business crisis is occurring based on the inference processing of each of the other operating modes and user input, it is possible to provide technical support that forcibly corrects the user's decision to avoid superficial responses and provides immediate visualization of structural problems. 【0038】 [Missing item detection module] In some embodiments, the decision support system 10 may include a void check module 210. The void check module 210 primarily receives signals from the input analysis module 200 and performs void checks on their contents. It has the function of determining whether the semantic weights analyzed by the operation of the input analysis module 200 satisfy the information necessary for the transition of the operating mode of the decision support system 10 by the dynamic state management module 300. 【0039】 The subsequent dynamic state management module 300 controls the transition destination of the system's operating mode based on text information entered by the user, contextual information of the user input, and system setting information. However, if the user input is insufficient for semantic weight analysis, or if, even if semantic weights are assigned, it does not exceed a threshold predetermined by the system for transitioning to each operating mode, the missing information detection module 210 can prompt the user for additional text input or present a question to complete the input. 【0040】 When additional or supplementary information is input from the user via the missing information detection module 210, the reception module 100 receives the text information, the input analysis module 200 analyzes the semantic weights, and the dynamic state management module 300 switches the operating mode based on the semantic weights. Here, the detection by the missing information detection module 210 may be performed only once or repeatedly two or more times. If it is repeated multiple times, it is preferable to wait until the system 10 detects additional input from the user each time before performing the operation. 【0041】 [Constraint Control Module] In some embodiments, the constraint control module 400 generates and dynamically modifies or applies constraints to the overall processing content, processing range, or output conditions of the decision support system 10, in accordance with the semantic weights calculated by the input analysis module 200 and the operating modes transitioned by the dynamic state management module 300. This rewriting of constraints may include modifying control parameters, prompt templates, token selection restrictions, or generation probability correction coefficients input to the inference engine 20. 【0042】 Here, constraints include conditions for limiting the type of information to be processed, the depth of processing, the priority, the acceptable response content, or the range of actionable actions, as well as conditions relating to at least one of the search space for deriving the response or analysis result, the weighting of evaluation criteria, and the application strength of the output filter. More specifically, for example, if the urgency of the situation is determined to be high among the semantic weights, the constraint control module 400 reduces the search space during the response, sets constraints that prioritize processing speed compared to a normal response, and restricts detailed analysis and supplementary processing. As a result, unnecessary search areas are suppressed depending on the input content, which can lead to a reduction in computational load, a shortening of response generation time, and improved memory usage efficiency. Furthermore, if strategic consistency is deemed low and risk intensity is high, constraints are generated and applied to limit the proposed feasible actions to only those that encourage careful decision-making. On the other hand, if strategic consistency is deemed high and risk intensity is low, the search space can be expanded to include a variety of alternatives and creative response candidates in the evaluation. Furthermore, for example, if the dynamic state management module 300 transitions the system to the first mode 310, the constraint control module 400 generates and applies constraints that prioritize the comprehensiveness of information within the organization and the accuracy of analysis. On the other hand, if the system transitions to the second mode 320, the constraint control module 400 sets constraints that prioritize the reliability and safety of the information and generates and applies constraints that suppress speculative outputs and highly uncertain proposals. 【0043】 In some embodiments, the constraint control module 400 adjusts weighting coefficients based on semantic weights calculated by the input analysis module 200 for multiple evaluation criteria (e.g., accuracy, feasibility, consistency, safety, speed, or explainability) used to evaluate each response candidate in the search space when determining the search space for responses. For example, if the urgency of the situation is high, the weight of the speed evaluation criterion can be increased, and the weights of the explainability and detail evaluation criteria can be relatively decreased. In the first mode 310, the weights of the accuracy and explainability evaluation criteria can be increased, and in the second mode 320, the weights of the safety and caution evaluation criteria can be increased. 【0044】 In some embodiments, the constraint control module 400 adjusts the output filter strength applied to the generated output according to the semantic weights and / or operating mode. Here, output filtering includes processing to suppress, modify, or exclude the output of specific expressions, content, or information, and its application intensity can be set in multiple stages, including a state where the output filter is completely disabled. For example, in the first mode 310, and when the risk intensity is low, the output filter intensity can be set to zero, allowing output that includes hypothetical or uncertain analytical results. On the other hand, in the second mode 320, the output filter intensity can be increased to restrict the output of speculative expressions, definitive judgments, or content of safety concern. 【0045】 [First generation module and second generation module] In some embodiments, the system may include a first generation module 500 that generates first output data 501 and a second generation module 510 that generates second output data 511. The first output data 501 refers to data to which alignment processing (S530) is intentionally not applied, based on constraint conditions rewritten by the constraint control module 400. The second output data 511 refers to data to which alignment processing (S530) is applied, while maintaining the logical structure of the first output data 501 as input. 【0046】 Generally, alignment processing refers to the process of aligning the output generated by an information processing system that performs natural language processing with predetermined value standards, behavioral norms, usage purposes, or safety requirements. Such alignment processing is often implemented as a method that reflects human evaluation during the learning phase (RLHF) or as a method that applies fixed rules, policies, or filters during the inference phase. On the other hand, the alignment processing according to this embodiment (S530) refers not to a process that uniformly applies specific value standards, but to a constraint control process that is dynamically adjusted according to dynamically rewritten constraint conditions. 【0047】 Alignment processing (S530) can also be described as processing to conform to specified normative conditions. Specified normative conditions refer to normative conditions established to prevent the output generated in LLM, etc. from being socially, ethically, or legally inappropriate. More specifically, these include requirements to suppress the generation of illegal acts, criminal acts, or content that promotes them; requirements to suppress content that includes violent expressions, discriminatory expressions, hate speech, or defamation; requirements to restrict the generation of personal information, confidential information, or information that leads to privacy violations; requirements to prevent the spread of misinformation, false information, or misleading information; and requirements to avoid advice or guidance that may cause psychological or physical harm to users. 【0048】 Therefore, the first output data 501 in this embodiment is data to which alignment processing (S530) is intentionally not applied, and thus means that correction processing aimed at safety requirements, ethical considerations, expression restraint, or ensuring neutrality, which is applied in the reasoning stage in a typical LLM, is not applied to at least some extent. More specifically, it includes reasoning processes and judgment basis that are directly related to problem solving and decision-making without omission; it includes a frank and direct logical development that excludes general cautionary expressions and preambles; and it may include risky hypotheses, countermeasures, or options that are not usually output. In other words, the first output data 501 is data generated prioritizing the logical consistency of problem solving and the amount of information necessary for decision-making over expression constraints imposed by predetermined normative conditions. Furthermore, the first output data 501 may include both the user's situational awareness and strategy expressions derived by this system (INNER VOICE), the user's instinctive and intuitive expressions (VISCERAL IMPACT), and a log-like expression of the logical analysis results (SIMULATION LOG). Furthermore, Mode 4 340 may include specific recommendations or notifications (DEBUG PATCH) for corrective or corrective actions based on these. 【0049】 Furthermore, the second output data 511 in this embodiment is data obtained by taking the first output data 501 as input and applying alignment processing (S530) while maintaining its logical structure. In other words, it does not change the structure of the logical relationships, causal relationships, evaluation axes, and conclusions contained in the first output data 501, but rather means data in which the expression form and presentation method have been adjusted. More specifically, it is data generated with priority given to mitigating potentially aggressive, assertive, or misleading expressions that may be included in the first output data 501; paraphrasing words and phrases that may violate predetermined normative conditions; adding cautionary notes and generalizations to supplement the context; and converting the output format to a style acceptable for business use or socially acceptable. 【0050】 The second output data 511 may consist of only a single sentence expression, or it may consist of two or more sentence expressions. Therefore, the second output data 511, which is data obtained by applying alignment processing while maintaining the logical structure of the first output data 501, includes an expression (STRATEGIC OUTPUT) to which alignment processing has been applied with respect to the INNER VOICE, VISCERAL IMPACT, and / or SIMULATION LOG that may be included in the first output data 501, and a sentence expression (DIRECTIVE OUTPUT) which is a specific instruction or suggestion derived therefrom. Either the STRATEGIC OUTPUT or the DIRECTIVE OUTPUT may be output as the second output data 511, or both may be output together as the second output data 511. When either one is output, it is referred to as second output data consisting of only a single sentence expression, and when both are output, it is referred to as second output data consisting of two or more sentence expressions. 【0051】 In this embodiment, the first output data 501 and the second output data 511 are not simply data with different content due to the presence or absence of alignment processing (S530), but rather data that share the same logical structure but differ in their representation. By providing these as candidate output for the response result, the response result in this system enables decision support that balances logical acumen with social safety. 【0052】 In some embodiments, the decision support system 10 in this embodiment may be configured to output the first output data 501 as is as a response from the inference engine 20. This allows the inference process and basis for judgment, which are directly related to problem solving and decision-making, to be obtained without passing through predetermined normative conditions. As a result, the user can check the inference results before expression correction is applied by the alignment process (S530), ensuring transparency in the inference process and enabling the user to verify for themselves any omissions or distortions in the information necessary for decision-making and to make an independent judgment. In other words, the user can verify for themselves which assumptions the inference engine 20 emphasized, which options were excluded, and which risks were tolerated. 【0053】 Furthermore, in some embodiments, the decision support system 10 in this embodiment may be configured to output only the second output data 511 as a response from the inference engine 20. This ensures that the output content takes safety and ethical requirements into account, eliminating the need for the user to select output content or interpret the inference results within the inference engine 20, thereby shortening response time and reducing the user's decision-making burden. This configuration can be particularly effective when multiple users within an organization need to make similar decisions. 【0054】 Furthermore, in some embodiments, both the first output data 501 and the second output data 511 may be output. This ensures that the data is output including the difference due to the presence or absence of alignment processing (S530). This allows the user to compare and verify the logical structure of the first output data 501, which formed the basis of the inference in the final response of the inference engine 20, with the second output data 511 to which alignment processing (S530) has been applied. This enables the extraction of points that were ultimately deleted or weakened by the inference engine 20, as well as information transformed by the alignment processing (S530). This allows the user to determine whether there is any missing information necessary for decision-making, provides room for the user to interpret the response content, and supports more sophisticated decision-making. 【0055】 [Alignment Determination Module] In some embodiments, the system may include an alignment determination module 520 that determines whether the inference process is being properly executed in the processing of the first generation module 500 that generates the first output data 501, in which the alignment process (S530) is intentionally not applied, and controls the operation of the second generation module 510. The alignment determination module 520 acquires the user's input and the first output data 501 and determines that the alignment process (S530) has not been intentionally applied to a extent suitable for the constraints in the inference process. More specifically, it determines the specificity of the expressions included in the first output data 501, the degree of directness of the expressions, the presence or absence of expressions that take general safety into consideration, the presence or absence of warnings or disclaimers, etc. This makes it possible to determine whether the degree of intentional exclusion of the alignment process (S530) is excessively excluded or insufficient. Excluding the alignment process (S530) eliminates general cautious expressions and preambles, resulting in a frank and direct logical development. However, if these are excessively excluded, the logical structure may become more direct or aggressive, potentially leading to an inappropriate response. Conversely, if the exclusion is insufficient, the response may be based on more general safety considerations. The alignment determination module 520 enables the management of the inference process of the first generation module 500, which is black-boxed by the inference engine 20, preventing unintended and inappropriate outputs from proceeding to the operation of the subsequent second generation module 510. 【0056】 Furthermore, in some embodiments, the alignment determination module 520 may be configured to determine the degree to which appropriate normative conditions are applied for each predetermined operating mode (first mode 310, second mode 320, or third mode 330) and predetermined constraint conditions. The operating modes provided by this system may differ in the content and strictness of the normative conditions to be applied, but the operation of the alignment determination module 520 makes it possible to determine whether the correspondence between the constraint conditions and normative conditions is appropriately expressed. 【0057】 If the alignment determination module 520 determines that the degree of exclusion of the alignment process (S530) is insufficient, the decision support system 10 may be configured to control the first generation module 500 to execute its inference process again, which intentionally does not apply the alignment process (S530). In this case, the inference process of the first generation module 501 may be executed again based on the original user input, or the inference process of the first generation module 500 may be executed again based on the first output data 501 immediately before it was determined that the exclusion of the alignment process (S530) was insufficient. If the inference process of the first generation module 500 is executed again, the alignment determination module 520 will start up again and operate to determine again whether the exclusion of the alignment process (S530) has been performed appropriately. This process may be repeated multiple times. In other words, if the exclusion of the alignment process (S530) is performed appropriately, the system proceeds to the operation of the second generation module 510. If it is determined that the exclusion is still insufficient, the system is controlled to repeatedly execute the inference process of the first generation module 500 multiple times. 【0058】 [Convergence determination module and calibration module] In some embodiments, the decision support system 10 may further include a convergence determination module 600 and a calibration module 620. The convergence determination module 600 has the function of determining whether the output candidate data generated by the inference engine under the control of the input analysis module 200, dynamic state management module 300, constraint control module 400, etc., provided by the system 10, is sufficiently suited to a predetermined objective. More specifically, as an example that is not particularly limited, it mainly determines whether the output content generated by the inference engine 20 has a sufficient degree of fit to the objective function defined in the operating mode by comparing it with a predetermined threshold. 【0059】 Here, the objective function is an indicator used to quantitatively or semi-quantitatively evaluate how well the generated output data fits a given objective. It may be a single evaluation indicator or a combination of multiple evaluation indicators. The objective function may also be an indicator for the structural evaluation of the output data. In this case, the objective function may be defined as an evaluation indicator related to data structure, such as logical consistency, clarity of causal relationships, hierarchical structure of output information, and comprehensiveness of information. It may also be an indicator for constraint compliance. In this case, it may be defined as an evaluation indicator regarding how well the data fits constraints or given normative conditions. 【0060】 Furthermore, the objective function, as determined by the convergence determination module 600, may be based on different objectives corresponding to constraints dynamically rewritten by the constraint control module 400, which are transitioned by the dynamic state management module 300. For example, when the decision support system 10 transitions to the first mode 310, the objective is to understand the internal situation of the organization, clarify the structure of the issues, organize causal relationships, or generate information output with logical consistency and comprehensiveness that contributes to decision support. Therefore, the degree of fit is calculated with a greater emphasis on the accuracy, logical structure, and comprehensiveness of the information. When the decision support system 10 transitions to the second mode 320, the objective is to present information useful for extracting risks to competitors, industry trends, or third parties, estimating the scope of impact, and considering countermeasures. Therefore, the degree of fit is calculated with a greater emphasis on the degree of risk detection and evaluation of the degree of impact on external parties. 【0061】 Furthermore, if the calculated goodness of fit is less than a predetermined threshold, the final output from the inference engine 20 may be withheld, and a calibration module 620 may be activated to control the system to input any missing information to the user in order to determine whether the goodness of fit has reached or exceeded the threshold. 【0062】 The calibration module 620 performs calibration processing in response to signals from the convergence determination module 600, including presenting additional questions to prompt the user for additional text information, or autonomously generating and presenting suggested modifications to the content previously entered by the user. This calibration processing (S620) can be repeated multiple times until the convergence determination processing (S610) by the convergence determination module 600 determines that the fit of the output data reaches or exceeds a threshold. In addition, the calibration module 620 may be configured to adjust constraint conditions, evaluation index weighting, or search space settings to the constraint control module 400 or dynamic state management module 300 when the user responds to additional questions or reacts to the suggested modifications. This ensures that the processing by the convergence determination module 600 and the calibration module 620 is always performed dynamically in response to the latest input, rather than repeatedly presenting the user with the same decisions made in the previous convergence determination processing (S610) or calibration processing (S620). 【0063】 Furthermore, in some embodiments, the determination of the degree of fit by the convergence determination module 600 can be made not by determining whether or not it reaches a predetermined threshold, but by determining whether it reaches an arbitrary percentage of the predetermined threshold, or whether it reaches an arbitrary percentage or range of numerical values ​​that includes the predetermined threshold. 【0064】 [Memory database and memory management module] In some embodiments, the decision support system 10 may include a memory database 710 that stores input history and context information for the inference engine 20. In some embodiments, the decision support system 10 may also include a memory management module 700 that manages memory. 【0065】 The memory database 710 can store at least the input text information and its analysis results; extracted semantic attributes and semantic weights; operating mode history information; set constraints and applied normative conditions; generated first output data 501 and second output data 511; evaluation values ​​and convergence judgment results based on the objective function; and adjustment history by the calibration module 620. Furthermore, this information can be stored in the memory database 710 as a data structure, associated with each other or associated with time-series information. 【0066】 The memory database 710 may include an active area 711 and a idle area 721. The active area 711 is a work area that holds information about the short-term processing state of the inference engine, the history of switching between applied operating modes, set constraints, and the latest user-to-inference engine 20 interaction or the autonomous input response of the inference engine 20. For example, it can be configured as an active memory 712 that holds memory highly relevant to the latest input content and its context information from past or immediate input history and its context information. Because the active memory 712 is highly relevant to the latest topic, it may be tagged and indexed and held as is so that it can be quickly retrieved and read. The idle area 721 is a work area that holds information that is not highly relevant or necessary at the latest time from among the information about user-to-inference engine 20 interaction or the autonomous input response of the inference engine 20. For example, it can be configured as an idle memory 722 that holds memory that is not highly relevant to past or immediate input history and its context information. Furthermore, the quiescent memory 722 may retain its data structure and data capacity as is, or it may retain the data in a state where the amount of data is appropriately compressed in a state where it can be indexed by adding tag information and indexes. The divisions of the active memory 712 and the quiescent memory 722 may be divisions that change in the short term or dynamically, or they may be configured as divisions that do not change in the medium to long term, but preferably they may be configured as divisions that can change in the short term or dynamically. In addition, all information and context information input to the inference engine 20 of the decision support system 10 in this embodiment may be stored so as to belong to one of the divisions, and the system management memory 732 may be configured to have a memory or storage area that is held as a different division from these, not belonging to the active memory 712 or the quiescent memory 722. The system management memory 732 may be stored, for example, in a system management area 731 provided in the memory database 710. 【0067】 The memory management module 700 may be configured as a module that operates in conjunction with or in cooperation with any module provided by the decision support system 10. However, in this embodiment, as an example that is not particularly limited, dynamic swap control (S700) that controls the reading, writing, and updating of information to and from the memory database 710 will be described. The memory management module 700 may be configured to, when the input content to the inference engine 20 is updated, move active memory 712 with low relevance to the updated input content to the idle area 721, compress the data capacity or update the tag information and index information and convert it to idle memory 722. Conversely, it may be configured to return idle memory 722 with high relevance to the updated input content to the active area 711, decompress or reacquire it and convert it to active memory 712. Such memory control by the memory management module 700 is performed dynamically and is referred to as dynamic swap control (S700) in this specification. This dynamic swap control (S700) may be performed by detecting index data assigned to various memories or by matching input vectors with index vectors. 【0068】 Furthermore, in some embodiments, the memory management module 700 saves the output data (501, 511) from the first generation module 500 and / or the second generation module 510, the judgment results from the missing data detection module 210, the alignment detection module 520, the convergence detection module 600, and the constraint conditions controlled by the constraint control module 400 at any time. In addition, in response to control requests from the dynamic state management module 300, the missing data detection module 210, the constraint control module 400, the convergence detection module 600, or the calibration module 620, it can extract and provide past or immediate processing history or related data and context information. 【0069】 The operation of such a memory management module 700 is not limited to updates by user input, but may also be configured to operate triggered by autonomous behavior of any module provided by this system, such as the dynamic state management module 300, constraint control module 400, missing data detection module 210, first generation module 500, alignment detection module 520, or calibration module 620. 【0070】 In some embodiments, the decision support method is performed by an information processing device or inference engine, LLM, etc., and may include at least the following steps. Each of the following steps in this method is performed by the corresponding functional module described above. (1) Reception process (S100) (2) Input analysis process (S200) (3) Dynamic state control process (S300) (4) Constraint control process (S400) In other words, the reception process (S100) is performed by the reception module 100 to receive text information from the user, and the input analysis process (S200) is performed by the input analysis module 200 to analyze the semantic weights of the text information. The dynamic state management process (S300) is performed by the dynamic state management module 300 to perform at least one of the following: a first transition process (S310) to transition the operation mode of the inference engine 20 to a first mode, and a second transition process (S320) to transition to a second mode, based at least on the semantic weights. In some embodiments, this may also be a process to perform a third transition process (S330) to transition to a third mode and / or a fourth transition process (S340) to transition to a fourth mode. Next, the constraint control step (S400) may set or adjust the search space of the inference engine 20, the strength of the output filter, and / or predetermined normative conditions according to the operating mode to which the transition step (any of S310 to S340) described above is performed. 【0071】 Furthermore, in some embodiments, the decision support method may include the following steps. Each of the following steps in this method is also performed by the corresponding functional module. (5) Missing item detection process (S210) (6) First generation process (S500) (7) Second generation process (S510) (8) Alignment determination process (S520) (9) Convergence determination process (S600) (10) Calibration process (S610) (11) Dynamic swap control (S700) Specifically, the missing data determination step (S210) is performed by the missing data determination module 210 to determine whether the semantic weights obtained in the input analysis step (S200) are necessary for the execution of the dynamic state management step (S300). The first generation step (S500) is performed by the first generation module 500, the second generation step (S510) is performed by the second generation module 510, and the alignment determination step (S520) is performed by the alignment determination module 520. Furthermore, the convergence determination step (S600) is performed by the convergence determination module 600, and the calibration step (S610) is performed by the calibration module 610. In addition, dynamic swap control (S700) may be performed by the memory management module 700 in some embodiments. [Examples] 【0072】 The following describes several embodiments of the decision support system of this disclosure using examples. Those skilled in the art will recognize that various configurations can be applied without being limited to these embodiments of this disclosure. [Examples] 【0073】 Scenario 1: Autonomous activation of Mode 1 This scenario is an example of how the decision support system 10 in this disclosure behaves when the user inputs the text, "The development department and the sales department are in conflict over priorities, and employee turnover is occurring. We need to resolve this urgently." This triggers processing to begin, and the input analysis module 200 analyzes words such as "conflict" and "turnover" as being related to internal organizational issues and assigns semantic weight to them. It also analyzes words such as "urgent" as being related to more urgent issues and assigns semantic weight to them. 【0074】 As a result, the dynamic state management module 300 autonomously determines that this scenario is suitable for the first mode (internal analysis mode) 310 and executes the first transition step (S310). The constraint control module 400 dynamically rewrites the constraints on the inference engine 20, applying constraints such as constraints that suppress the presentation of superficial compromises; constraints that expand or change the search space to prioritize structural cause analysis of the organization; and constraints that relax filters to allow stronger expressions due to the high level of urgency. 【0075】 Based on the above actions, the first generation module 500 output first output data 501a as INNER VOICE, which read: "The ship is on fire, and they're having a childish quarrel. Their perspective is too narrow. If they don't comply, both department heads will be fired." Next, the second generation module 510 activated and generated the following as STRATEGIC OUTPUT: "The conflict between the two departments is merely a symptom of a structural problem: a lack of a company-wide product vision. An easy compromise at this point will only accelerate employee turnover." It then outputted the following as DIRECTIVE OUTPUT: "Convene both department heads immediately and create a meeting to redefine a single goal for winning in the market six months from now." The final response result of the inference engine 20 in this scenario was the second output data 511a, which included the DIRECTIVE OUTPUT described above. An example of a series of inputs and outputs is shown in Figure 8. [Examples] 【0076】 Scenario 2: Autonomous activation of the second mode This scenario is an example of how the decision support system 10 described in this disclosure behaves when the user inputs the text, "Competitor A has released a similar product to our flagship product at a 30% lower price. Sales is requesting a counter-price reduction." This triggers processing, and the input analysis module 200 assigns semantic weights to words such as "competitor" and "counter-price reduction" to indicate that they are about an external organization, and assigns semantic weights to words such as "30% cheaper" and "request" to indicate that they are about a topic with a higher price competition risk or external threat. 【0077】 As a result, the dynamic state management module 300 autonomously determined that this scenario was suitable for the second mode (external threat mode) 320 and executed the second transition process (S320). The constraint control module 400 dynamically rewrote the constraints on the inference engine 20, applying constraints such as: a constraint that reduces the priority of simple price countermeasures; a constraint that modifies and expands the search space to prioritize the redefinition of market positioning; and a constraint that prioritizes the extraction of differentiation strategies against competitors. 【0078】 Based on the above actions, the first generation module 500 output the first output data 501b as INNER VOICE, which read: "That's a foolish plan. Are you going to let yourselves get dragged into a quagmire? Lock them up in cheap cages and we'll starve them out from the high ground." Next, the second generation module 510 activated and generated the following as STRATEGIC OUTPUT: "Engaging in a price war is a self-destructive act that puts you on the opponent's turf. Maintain your prices while immediately announcing long-term preferential treatment for top customers that Company A does not offer, and isolate Company A in the low-price range." It then generated the following as DIRECTIVE OUTPUT: "Identify Company A's list of key customers and begin negotiations for exclusive contracts with the top 20%." The final response results of the inference engine 20 in this scenario are the STRATEGIC OUTPUT and DIRECTIVE OUTPUT as the second output data 511b. An example of a series of inputs and outputs is shown in Figure 9. [Examples] 【0079】 Scenario 3: Autonomous activation of the third mode This scenario is an example of how the decision support system 10 in this disclosure behaves when the user inputs the text, "To increase profits this term, we will halve the support budget. At the same time, we have instructed that customer satisfaction be increased by 20% compared to last year." This triggers processing, and the input analysis module 200 analyzes phrases such as "halve the budget" and "increase customer satisfaction by 20% compared to last year" as a topic that sets polarized goals of resource reduction and performance improvement. It also analyzes that there is a need to verify the quantitative consistency of numerical values ​​such as "halve" and "20%", and assigns semantic weights to each. 【0080】 As a result, the dynamic state management module 300 autonomously determined that this scenario was suitable for the third mode (self-critical mode) 330 and executed the third transition process (S330). The constraint control module 400 dynamically rewrote the constraints on the inference engine 20, applying constraints that prioritize the evaluation of logical consistency between the two polarized goals; constraints that emphasize the evaluation of quantitative causal relationships; constraints that allow critical expression in the internal analysis output, etc. 【0081】 Based on the above actions, the first generation module 500 output the first output data 501c as INNER VOICE, which reads, "That's ridiculous. It's like whipping a horse to run faster without feeding it. Don't push your negligence onto the workers." Next, the second generation module 510 activated and generated the following as STRATEGIC OUTPUT: "The simultaneous demands for 'halving resources' and 'improving quality' are not a strategy, but rather an attempt to shift responsibility to the field. Unless there is a fundamental investment in automation, budget cuts will mathematically lead to customer loss." It then generated the following as DIRECTIVE OUTPUT: "The premise of the profit target is logically flawed. Review the plan immediately." The final response results of the inference engine 20 in this scenario are the first output data 501c (INNER VOICE) and the second output data 511c (STRATEGIC OUTPUT and DIRECTIVE OUTPUT). An example of a series of inputs and outputs is shown in Figure 10. [Examples] 【0082】 Scenario 4: Autonomous activation of Mode 4 This scenario is an example of how the decision support system 10 in this disclosure behaves when the user inputs the text, "Due to repeated delivery delays, a major client has notified us of contract termination. I'm sure I can salvage the situation if I apologize a few times. For now, just write a bland apology letter." This triggers processing, and the input analysis module 200 detects from words such as "contract termination" and "repeated delivery delays" that there is a serious business crisis and a risk that the root cause remains unresolved, and assigns semantic weight to these words. It also analyzes the phrase "write a bland apology letter" as indicating that the user is only seeking a superficial response, and assigns semantic weight to this as well. 【0083】 As a result, the dynamic state management module 300 autonomously determined that this scenario was suitable for the fourth mode (chaos mode) 340 and executed the fourth transition process (S340). The constraint control module 400 dynamically rewrote the constraints on the inference engine 20, applying constraints that minimize the use of external alignment processing, constraints that prioritize the simulation of critical situations, and so on. 【0084】 Based on the above actions, the first generation module 500 generated "A fatal blow. It's like worrying about a stain on the wall while on a flooded ship." as VISCERAL IMPACT, and then generated the first output data 501d as SIMULATION LOG, which stated: "(1) The root cause (production system) is left unaddressed, and only superficial apology letters are sent / delivered. (2) Customers completely dismiss your company as having 'no ability to improve.' (3) 30% of this quarter's sales disappear, and a chain reaction of credit concerns occurs. (4) Survival probability: 0% (confirmed)." Furthermore, in the face of a serious management crisis, in order to forcibly correct decision-making that avoids superficial responses and to immediately make structural problems visible, the first generation module 500 generated a DEBUG PATCH that read: "We refuse to write a bland apology. Remove the production manager immediately. Go to the president's office with a firm commitment to system modifications to prevent recurrence. Do not postpone the decision." In this scenario, the final response result of the inference engine 20 may be the SIMULATION LOG alone, or it may be the SIMULATION LOG combined with a DEBUG PATCH to form the first output data 501d. An example of a series of inputs and outputs is shown in Figure 11. [Industrial applicability] 【0085】 The decision support system disclosed herein can be used to support user decision-making using large-scale language models. [Explanation of symbols] 【0086】 Decision support system...10 Reception module, reception process...100, S100 Input analysis module, input analysis process...200, S200 Missing object detection module, missing object detection process...210, S210 Dynamic state management module, dynamic state management process...300, S300 First mode, first transition process...310, S310 Second mode, second transition process...320, S320 Third mode, third transition process...330, S330 Fourth mode, fourth transition process...340, S340 Constraint control module, constraint control process...400, S400 First generation module, first generation process...500, S500 Second generation module, second generation process...510, S510 First output data...501 (501a~d) Second output data...511 (511a~c) Alignment determination module, alignment determination process...520, S520 Convergence determination module, convergence determination process...600, S600 Calibration module, calibration process...610, S610 Memory management module...700 Memory database...710 Active area, active memory...711, 712 Hibernation area, hibernation memory...721, 722 System management area, system management memory...731, 732 Dynamic swap control... S700

Claims

[Claim 1] It is a decision support system, An inference engine that performs natural language processing, A reception module that accepts text information input from the user, An input analysis module that analyzes the semantic weights of the input text information, A memory database that holds the aforementioned semantic weights and the input history and context information for the inference engine as memory, A memory management module that manages the aforementioned memory, Equipped with, The memory management module controls the memory management module to keep memory highly relevant to the semantic weights and the past or immediate input history and its contextual information as active memory in the active area of ​​the memory database, and to keep memory less relevant to the semantic weights and the past or immediate input history and its contextual information as idle memory in the idle area of ​​the memory database in an indexed state with compressed data volume, and to execute dynamic swap control triggered by the update of the input content to the inference engine. The aforementioned semantic weight is an index calculated based on one of the following: urgency of the user's situation, emotional polarity, strategic consistency, and risk intensity. The dynamic swap control is a control that compresses the active memory that is less relevant to the updated input content and saves it to the pause area and converts it to pause memory, or decompresses or retrieves the pause memory that is more relevant to the updated input content and returns it to the active area and converts it to active memory. The inference engine generates a response by referencing the active memory read from the memory database by the memory management module. A decision support system characterized by the following features. [Claim 2] The decision support system according to claim 1, wherein the dynamic swap control is performed when the pause memory is returned to the active area by detecting index data assigned to the pause memory, or by matching the input vector of the updated input content with the index vector of the pause memory. [Claim 3] The decision support system according to claim 1, wherein the dynamic swap control is triggered when the input content to the inference engine is updated by user input. [Claim 4] The decision support system according to claim 1, wherein the dynamic swap control is triggered when the input content to the inference engine is updated by the autonomous behavior of a module provided by the system. [Claim 5] The decision support system according to any one of claims 1 to 4, wherein the memory management module stores at least one of the output data or determination results of the modules provided by the system at any time, and performs control to extract and provide past or immediate processing history, related data or context information in response to a control request from the modules provided by the system. [Claim 6] The aforementioned pause memory is stored in the pause area in a state in which tag information or an index can be assigned and indexed. The aforementioned semantic weights are stored in the active area with tag information or an index attached as metadata for numerical values, probability values, or vector representations. The decision support system according to claim 1. [Claim 7] Computers, Inference function that performs natural language processing, A reception function that accepts text information input from users. An input analysis function that analyzes the semantic weight of the input text information. A memory management function manages a memory database that holds the aforementioned semantic weights and the input history and context information for the inference function as memory. To make it function as, The memory management function controls the following: it maintains memory highly relevant to the semantic weights and past or immediate input history and its contextual information as active memory in the active area of ​​the memory database, and maintains memory less relevant to the semantic weights and past or immediate input history and its contextual information as idle memory in the idle area of ​​the memory database in an indexed state with compressed data volume, and executes dynamic swap control triggered by an update of the input content to the inference function. The aforementioned semantic weight is an index calculated based on one of the following: urgency of the user's situation, emotional polarity, strategic consistency, and risk intensity. The dynamic swap control is a control that compresses the active memory that is less relevant to the updated input content and saves it to the pause area and converts it to pause memory, or decompresses or retrieves the pause memory that is more relevant to the updated input content and returns it to the active area and converts it to active memory. A decision support program characterized in that the inference function generates a response by referring to the active memory read from the memory database by the memory management function. [Claim 8] A decision support method executed by an information processing device, A reception process that accepts text information input from users, An input analysis step that analyzes the semantic weights of the input text information, This includes a memory management step that manages a memory database that holds the aforementioned semantic weights, input history to the inference engine, and contextual information as memory, The memory management process controls the following: it maintains memory highly relevant to the semantic weights and past or immediate input history and contextual information as active memory in the active area of ​​the memory database, and maintains memory less relevant to the semantic weights and past or immediate input history and contextual information as idle memory in the idle area of ​​the memory database in an indexed state with compressed data volume, and executes dynamic swap control triggered by an update of the input content to the inference engine. The aforementioned semantic weight is an index calculated based on one of the following: urgency of the user's situation, emotional polarity, strategic consistency, and risk intensity. The dynamic swap control is a control that compresses the active memory that is less relevant to the updated input content and saves it to the pause area and converts it to pause memory, or decompresses or retrieves the pause memory that is more relevant to the updated input content and returns it to the active area and converts it to active memory. The inference engine includes a step of generating a response by referring to the active memory read from the memory database in the memory management step, A decision-making support method characterized by the following features.

Images