A dialogue agent system based on cognitive modeling and an implementation method thereof

Abstract

The application discloses a kind of dialog intelligent agent systems and implementation methods based on cognitive modeling, it is related to artificial intelligence technical field.The system includes semantic cognitive modeling unit, decision-making cognitive modeling unit, behavior cognitive modeling unit and dialog reply execution unit, and realizes the following processes: through multiple semantic analysis channels, the current dialog information input by user is multi-dimensional semantic analysis, and the characteristic information of different semantic dimensions is obtained;Multi-dimensional feature information is fused and processed to generate a complete context representation and to extract a minimum key factor from the representation;Generate the core framework of reply content with the factor as the core driver;The core framework is checked and optimized at multiple levels to generate the final reply content and push it to the user, forming a cycle iteration, so through the two-stage information distillation architecture of "first package and then kernel" and the multi-level checking and optimization mechanism, the pertinence, inspiration, security and professional adaptability of the dialogue are improved, which is convenient for practical application and popularization.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a dialogue intelligent agent system and its implementation method based on cognitive modeling. Background Technology

[0002] With the rapid development of large language model technology, dialogue agents have been widely used in various fields such as intelligent customer service, personal assistants, and online education. However, in specific scenarios that require deep understanding and professional dialogue strategies (such as goal outcome management or career counseling), existing dialogue agents still have significant technical shortcomings.

[0003] Currently, most mainstream multi-turn dialogue response generation systems adopt an end-to-end generation architecture based on semantic encoding. This architecture typically first semantically encodes the user-input dialogue text to obtain a semantic vector that incorporates contextual information, and then directly inputs this semantic vector into a decoder to generate the response text word by word. While this end-to-end generation mode can produce relatively fluent responses in conventional dialogue scenarios, it has the following limitations in professional dialogue scenarios.

[0004] First, existing architectures often stop at generating a generalized semantic vector when understanding user input, lacking an intermediate processing step to further extract the core semantic concepts that determine the current response strategy from the complete context represented by that semantic vector. This results in generated responses that, while linguistically fluent and reasonable, lack a precise grasp of the user's deeper intentions, and are insufficiently targeted and insightful.

[0005] Second, the existing architecture's response generation process is typically completed in one go by a single decoder model, without dividing the response generation task into a core framework construction phase based on specific role settings and domain knowledge, and a verification and optimization phase based on multi-dimensional rules. This one-size-fits-all generation approach makes it difficult for the system to incorporate professional domain dialogue norms and constraints during the generation process, resulting in compromised professionalism and security of the responses.

[0006] Third, existing architectures generally lack a system design that forms a cognitive-level closed-loop iteration between generated responses and subsequent user inputs. This prevents dialogue systems from dynamically adjusting and optimizing response strategies based on a continuous understanding of the user's cognitive state in consecutive dialogue rounds, thus limiting the coherence and depth of multi-round dialogues.

[0007] Therefore, there is an urgent need for a new cognitive modeling dialogue agent that can further extract core decision factors from the complete context representation based on multi-dimensional semantic parsing, and generate structured responses and perform multi-level verification optimization based on these factors. Summary of the Invention

[0008] The purpose of this invention is to provide a dialogue agent system and implementation method based on cognitive modeling, in order to solve the problems of existing multi-turn dialogue response generation schemes, such as the lack of secondary extraction from complete context to core decision factors, the lack of a phased architecture for response framework generation and multi-level verification, and / or the lack of a closed-loop iteration mechanism.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a dialogue intelligent agent system based on cognitive modeling, comprising a semantic cognitive modeling unit, a decision cognitive modeling unit, a behavior cognitive modeling unit, and a dialogue response execution unit that are sequentially connected in communication. The semantic cognitive modeling unit is configured to receive current dialogue information input by the user and perform the following semantic cognitive modeling: parse the current dialogue information through multiple semantic analysis channels to obtain feature information of multiple semantic dimensions, wherein each of the semantic analysis channels is configured to extract and output feature information with different semantic dimensions, and the multiple semantic dimensions correspond one-to-one with the multiple semantic analysis channels; The decision-making cognitive modeling unit is configured to perform the following decision-making cognitive modeling: fuse the feature information of the multiple semantic dimensions to generate the contextual semantic package of the current dialogue, and extract the minimum key factor from the contextual semantic package through the contextual anchor generator. The contextual semantic package is a complete contextual representation that fuses multiple semantic dimensions, and is represented as a multi-dimensional feature vector in implementation, and is dynamically updated based on contextual awareness and awareness point recognition. The minimum key factor is a core semantic concept extracted twice from the complete contextual representation and used to determine the response strategy for the current dialogue turn. The behavioral cognition modeling unit is configured to perform the following behavioral cognition modeling: based on a preset role template and domain knowledge base, and using the minimum key factor as the core prompt, generate the core framework of the response content; The dialogue response execution unit is configured to perform multi-level verification and optimization on the core framework based on preset value alignment rules, security and ethics rules, and expression detail rules, generate the final response content, and push it to the user, so that the user can input new dialogue information to the semantic cognitive modeling unit based on the final response content, forming a cyclical iteration including semantic cognitive modeling, decision cognitive modeling, behavioral cognitive modeling, and dialogue response execution.

[0010] Based on the above-mentioned invention, a novel scheme for cognitive modeling dialogue agent is provided, which obtains a complete context representation and core decision factors based on multi-dimensional semantic parsing results, and performs structured response generation and multi-level verification optimization based on these factors. The system includes a semantic cognitive modeling unit, a decision cognitive modeling unit, a behavioral cognitive modeling unit, and a dialogue response execution unit connected in sequence. The semantic cognitive modeling unit performs multi-dimensional semantic parsing on the current dialogue information input by the user through multiple semantic analysis channels to obtain feature information of different semantic dimensions. The decision cognitive modeling unit fuses the multi-dimensional feature information to generate a complete context representation fused with multiple semantic dimensions as a context semantic package, and then uses a context anchor generator to extract the complete context from this context. The system extracts the minimum critical factor used to determine the response strategy for the current dialogue turn from the context representation. The behavioral cognition modeling unit, based on preset role templates and domain knowledge bases, drives the generation of the core framework of the response content with the minimum critical factor as the core. The dialogue response execution unit performs multi-level verification and optimization of the core framework according to preset value alignment rules, safety and ethics rules, and expression detail rules, generates the final response content and pushes it to the user. The user inputs new dialogue information based on the response and then enters the semantic cognition modeling unit again, forming a loop iteration. In this way, through the two-stage information distillation architecture of "package first and core later" and the multi-level verification and optimization mechanism, the relevance, inspiration, security and professional adaptability of the dialogue are improved, which is convenient for practical application and promotion.

[0011] In one possible design, the plurality of semantic analysis channels include a first semantic analysis channel, a second semantic analysis channel, and a third semantic analysis channel; The first semantic analysis channel is configured to extract feature information of the first semantic dimension from the current dialogue information to characterize the user's feeling category description and emotional features; The second semantic analysis channel is configured to extract feature information of the second semantic dimension from the current dialogue information to characterize the user's deep intent and core assumptions; The third semantic analysis channel is configured to extract feature information of the third semantic dimension from the current dialogue information to represent facts and observable events; The fusion process includes weighted fusion and conflict detection of feature information from the multiple semantic analysis channels to form the contextual semantic package.

[0012] In one possible design, the conflict detection includes: when there is an inconsistency between the feature information of the first semantic dimension output by the first semantic analysis channel and the feature information of the third semantic dimension output by the third semantic analysis channel, marking it as a contextual conflict, and incorporating the description information of the contextual conflict into the contextual semantic package, so that the contextual anchor generator can use it as a decision basis when extracting the minimum key factor.

[0013] In one possible design, the context anchor generator includes a decision network based on a multi-map dynamic combination mechanism. This decision network is configured to dynamically combine multiple candidate semantic concepts parsed from the context semantic package, based on multiple pre-set coaching process maps in the domain knowledge base, to build a cognitive exploration path that is adapted to the user's current context, and to determine the minimum critical factor from the multiple candidate semantic concepts based on this cognitive exploration path.

[0014] In one possible design, the behavioral cognitive modeling unit is further configured as follows: Based on the output of the decision network based on the multi-map dynamic combination mechanism, the intervention intensity parameter in the core framework of the response content is adjusted as follows: when the output indicates that the current dialogue effectiveness is lower than the first preset effectiveness threshold, the proportion of guiding content is increased; when the output indicates that the current dialogue effectiveness is higher than the second preset effectiveness threshold, the proportion of guiding content is decreased.

[0015] In one possible design, it also includes an emotion intensity scoring unit independent of the multiple semantic analysis channels and a tool scheduler storing multiple preset dialogue tools; The emotion intensity scoring unit is configured to perform a quantitative evaluation of the emotion intensity of the current dialogue information and generate the user's current emotion intensity score. The tool scheduler, which is communicatively connected to the decision-making cognitive modeling unit and the emotion intensity scoring unit, is configured to select and execute a target dialogue tool from the plurality of dialogue tools based on the minimum key factor and the current emotion intensity score, in order to generate a core framework of the response content or a generation strategy of the core framework that is adapted to the minimum key factor and the current emotion intensity score.

[0016] In one possible design, the plurality of dialogue tools includes at least a first dialogue tool and a second dialogue tool; The first dialogue tool is configured to adopt a dialogue strategy that is used to confirm the user's current state rather than directly solving the problem; The second dialogue tool is configured to employ dialogue strategies designed to guide users in physical awareness or emotional release. The tool scheduler is also configured to: when the current emotion intensity score exceeds a preset score threshold, forcibly switch the currently selected dialogue tool to the first dialogue tool or a preset safe dialogue tool; and when the current emotion intensity score exceeds the preset score threshold for multiple consecutive dialogue rounds, disable the second dialogue tool and trigger a preset safety dialogue strategy.

[0017] In one possible design, a topic anchor setting unit is also included, which is connected to the semantic cognitive modeling unit. The topic anchor setting unit is configured to: extract the total topic anchor from the dialogue information input by the user at the beginning of the dialogue, and correct the detected dialogue deviation according to the total topic anchor in subsequent dialogue rounds, so as to guide the dialogue back to the target topic pointed to by the total topic anchor. The topic anchor point setting unit is further configured to update the overall topic anchor point only when an explicit instruction is received from the user to actively modify the target topic during the dialogue; otherwise, the overall topic anchor point remains unchanged. When the topic anchor point setting unit detects a deviation in the dialogue, it prioritizes responding to the user's current state according to the preset coaching principles, and then gently guides the dialogue back into the framework based on the overall topic anchor point.

[0018] In one possible design, the dialogue response execution unit includes a value alignment validator, a security and ethics validator, and an expression detail coherence scorer. The value alignment verifier is configured to check whether the content of the core framework is consistent with the preset dialogue target according to the preset value alignment rules. The security and ethics verifier is configured to perform security risk detection on the content of the core framework according to preset security and ethics rules, and replace the content of the core framework with a preset security response template when a risk is detected. The expression detail coherence scorer is configured to optimize the fluency of the core framework and adjust the diversity of ending sentence structures according to preset expression detail rules.

[0019] In one possible design, the dialogue response execution unit is further configured to: after generating the final response content and before pushing the final response content to the user, perform a quantitative evaluation of the emotional intensity of the final response content to obtain a response emotional intensity score, and perform the quantitative evaluation of the emotional intensity of the current dialogue information to obtain the user's current emotional intensity score, then compare the response emotional intensity score with the current emotional intensity score, and optimize and update the final response content based on the comparison result.

[0020] In one possible design, the final response content is optimized and updated based on the comparison results, including: when the absolute value of the difference between the response emotional intensity score and the current emotional intensity score exceeds a preset threshold, adjusting the wording or expression of the final response content to raise or lower the response emotional intensity score to a range that matches the current emotional intensity score.

[0021] In one possible design, a user profile update unit is further included, which is communicatively connected to both the decision-making cognitive modeling unit and the behavior cognitive modeling unit, wherein the user profile update unit is configured as follows: Based on the historical sequence of the minimum key factors generated during the most recent multi-round conversations with the user, a cognitive profile of the user is generated or updated, which is used as an auxiliary input for the behavioral cognitive modeling unit in subsequent conversations with the user.

[0022] Secondly, the present invention provides a method for implementing a dialogue agent based on cognitive modeling, comprising the following steps S1 to S6: S1. Receive the current dialogue information input by the user, and then execute step S2; S2. Based on the current dialogue information, semantic cognitive modeling is performed as follows: the current dialogue information is parsed through multiple semantic analysis channels to obtain feature information of multiple semantic dimensions, and then step S3 is executed, wherein each of the semantic analysis channels is configured to extract and output feature information with different semantic dimensions, and the multiple semantic dimensions correspond one-to-one with the multiple semantic analysis channels; S3. Based on the feature information of the multiple semantic dimensions, decision-making cognitive modeling is performed as follows: the feature information of the multiple semantic dimensions is fused to generate the contextual semantic package of the current dialogue, and the minimum key factor is extracted from the contextual semantic package through the contextual anchor generator. Then, step S4 is executed. The contextual semantic package is a complete contextual representation that integrates multiple semantic dimensions, and is represented as a multi-dimensional feature vector in implementation. It is dynamically updated based on contextual awareness and awareness point recognition. The minimum key factor is a core semantic concept extracted twice from the complete contextual representation and used to determine the response strategy for the current dialogue round. S4. Based on the minimum key factor, perform behavioral cognition modeling as follows: Based on the preset role template and domain knowledge base, use the minimum key factor as the core prompt to generate the core framework of the response content, and then execute step S5; S5. Based on the preset value alignment rules, security and ethics rules, and expression detail rules, perform multi-level verification and optimization on the core framework, generate the final response content and push it to the user, and then execute step S6; S6. Receive new dialogue information input by the user based on the final reply content, return to the execution steps S2 to S5, forming a loop iteration including semantic cognitive modeling, decision cognitive modeling, behavioral cognitive modeling and dialogue reply execution.

[0023] The beneficial effects of the above scheme are: (1) This invention provides a novel cognitive modeling dialogue agent scheme that obtains a complete contextual representation and core decision factors based on multi-dimensional semantic parsing results, and generates structured responses and performs multi-level verification optimization based on these factors. The system comprises a semantic cognitive modeling unit, a decision cognitive modeling unit, a behavioral cognitive modeling unit, and a dialogue response execution unit, which are sequentially connected. The semantic cognitive modeling unit performs multi-dimensional semantic parsing on the current dialogue information input by the user through multiple semantic analysis channels to obtain feature information of different semantic dimensions. The decision cognitive modeling unit fuses the multi-dimensional feature information to generate a complete contextual representation that integrates multiple semantic dimensions as a contextual semantic package, and then uses a contextual anchor generator to generate a complete contextual representation based on the contextual anchors. The complete context representation extracts the minimum key factor used to determine the response strategy for the current dialogue turn. The behavioral cognition modeling unit, based on the preset role template and domain knowledge base, drives the generation of the core framework of the response content with the minimum key factor as the core. The dialogue response execution unit performs multi-level verification and optimization of the core framework according to the preset value alignment rules, safety and ethics rules, and expression detail rules, generates the final response content and pushes it to the user. The user inputs new dialogue information based on the response and then enters the semantic cognition modeling unit again, forming a loop iteration. In this way, through the two-stage information distillation architecture of "package first and core later" and the multi-level verification and optimization mechanism, the relevance, inspiration, security and professional adaptability of the dialogue are improved. (2) By sequentially communicating with the semantic cognition modeling unit, decision cognition modeling unit, behavior cognition modeling unit and dialogue response execution unit, a four-unit closed-loop cognitive architecture of "semantic cognition → decision cognition → behavior cognition → dialogue execution" is constructed, which enables the system to continuously deepen its understanding of the user's cognitive state in each round of dialogue and dynamically adjust the response strategy. This solves the technical problems of the existing end-to-end generation architecture lacking cognitive-level closed-loop iteration, multi-round dialogue coherence and depth limitation. (3) By extracting the minimum key factor from the complete context representation that integrates multiple semantic dimensions through the decision-making cognitive modeling unit, a two-stage information distillation architecture of "first package then core" is formed, so that the determination of the response strategy no longer depends on the generalized semantic vector, but is based on the core semantic concept accurately extracted from the complete context, which significantly improves the pertinence and inspiration of the response and overcomes the shortcomings of the end-to-end generation mode in the existing technology, which lacks intermediate decision-making links and the response is superficial. (4) The core framework of the response content is generated by the behavioral cognition modeling unit based on the preset role template and domain knowledge base. Then, the dialogue response execution unit performs multi-level verification and optimization, dividing the response generation process into two independent stages: "framework construction" and "verification and optimization". This allows the professional dialogue norms and constraints of the professional domain to be systematically integrated into the generation process, improving the professionalism and security of the response. (5) By extracting user emotions and relationship perceptions through the first semantic analysis channel, extracting user deep intentions and core assumptions through the second semantic analysis channel, and extracting facts and observable events through the third semantic analysis channel, a multi-dimensional structured semantic analysis of the same user input is realized. It can simultaneously obtain multi-level information such as emotional feelings, intentions and beliefs and factual basis from the user's expression, and form a complete understanding of the user's current dialogue state. (6) By performing contextual conflict detection in the fusion process, when there is an inconsistency between the user's emotional feelings and objective facts, it is marked as a contextual conflict and included in the contextual semantic package, so that the system can identify the bias or blind spot in the user's cognition and use the information of this cognitive bias as an important decision basis for extracting the minimum key factor, thereby effectively driving the dialogue to develop in depth in the direction of user cognition breakthrough. (7) By incorporating a decision network based on a dynamic combination mechanism of multiple maps into the context anchor generator, a cognitive exploration path adapted to the user's current context is constructed by dynamically combining the coaching process map in the domain knowledge base, and the minimum key factor is determined from multiple candidate semantic concepts, so that the core response strategy of each round of dialogue is generated based on the dynamically adapted cognitive exploration path rather than a fixed score, which is more in line with the user's personalized cognitive needs. (8) By dynamically adjusting the intervention intensity parameters in the core response framework based on the output of the decision network, the guiding content is increased when the dialogue effectiveness is low to help stimulate users, and the guiding content is reduced when the dialogue effectiveness is high to give users more space for independent thinking. This achieves adaptive adjustment of the intervention intensity and avoids the problems of excessive intervention or insufficient guidance. (9) The emotional intensity of user input is quantitatively evaluated by an emotional intensity scoring unit independent of the semantic analysis channel, and combined with the minimum key factor as the decision basis for tool scheduling. This makes the generation of response strategies not only consider the semantic core at the cognitive level, but also take into account the user's emotional state, thus realizing the two-dimensional synergy of cognitive decision-making and emotional perception. (10) By using the tool scheduler to dynamically select the appropriate target dialogue tool from multiple preset dialogue tools based on the minimum key factor and the current emotional intensity score, the system can flexibly switch between a companion dialogue strategy that confirms the user's status or an exploratory dialogue strategy that guides physical perception and emotional release for different cognitive needs and emotional states, thereby improving the scenario adaptability of the dialogue strategy. (11) By forcibly switching to a safe dialogue tool when the emotional intensity score exceeds the preset score threshold, and by disabling a specific tool and triggering a safety-guaranteed dialogue strategy when multiple consecutive dialogue rounds exceed the threshold, a hierarchical security gating mechanism from single-round security protection to multi-round cumulative judgment has been established, which effectively ensures the security and professional ethics compliance of dialogue in high emotional intensity scenarios. (12) By extracting the topic anchor point at the beginning of the conversation and correcting the deviation based on the anchor point in the subsequent conversation, the system can always carry out the conversation around the user's core target topic, preventing the conversation from deviating from the predetermined direction due to topic jumps or small talk. At the same time, the anchor point is only updated when the user actively issues a clear modification instruction, ensuring the stability of the conversation target topic and the user's dominance. (13) By performing an emotional intensity self-check on the reply content and comparing it with the user's current emotional state before pushing it after generating the final reply, the wording or expression of the reply is automatically adjusted when the difference between the two exceeds the threshold to increase or decrease the emotional intensity of the reply to the matching range, so that the reply finally pushed to the user can resonate with the user on an emotional level, significantly improving the empathy adaptability of the dialogue and the user experience. (14) By generating or updating the user cognitive profile based on the historical sequence of the minimum key factors generated in multiple rounds of dialogue, and using the profile as an auxiliary input for the behavioral cognitive modeling unit, the system can continuously accumulate a deep understanding of the user's cognitive patterns and thinking characteristics in long-term dialogue, realize long-term personalized adaptation of dialogue, and make the system not only a responder in single-round dialogue, but also a long-term dialogue partner that can accompany the user's cognitive growth. (15) Through the serial verification pipeline of value alignment verifier, safety and ethics verifier and expression detail coherence scorer in the dialogue response execution unit, the response content is reviewed and optimized step by step. While ensuring that the response value direction is correct and safety and ethics are compliant, the fluency and naturalness of language expression are improved, making the final output response more realistic, natural and in line with professional dialogue norms, which is convenient for practical application and promotion. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the structure of a dialogue agent system based on cognitive modeling, provided in an embodiment of the present invention.

[0026] Figure 2This is a flowchart illustrating the implementation method of a dialogue agent based on cognitive modeling provided in an embodiment of the present invention. Detailed Implementation

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these embodiments without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0028] It should be understood that although the terms "first" and "second", etc., may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, the first object may be referred to as the second object, and similarly, the second object may be referred to as the first object, without departing from the scope of the exemplary embodiments of the invention.

[0029] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, or A and B exist simultaneously. Another example is A, B and / or C, which can mean that any one of A, B, and C or any combination thereof exists. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone or A and B exist simultaneously. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0030] Example 1 like Figure 1 As shown, the cognitive modeling-based dialogue agent system provided in this embodiment includes, but is not limited to, a semantic cognitive modeling unit, a decision cognitive modeling unit, a behavior cognitive modeling unit, and a dialogue response execution unit that are sequentially connected in communication.

[0031] The semantic cognitive modeling unit is configured to receive current dialogue information input by the user and perform semantic cognitive modeling as follows: it parses the current dialogue information through multiple semantic analysis channels to obtain feature information with multiple semantic dimensions. Each semantic analysis channel is configured to extract and output feature information with different semantic dimensions, and the multiple semantic dimensions correspond one-to-one with the multiple semantic analysis channels. Semantic cognitive modeling refers to the process of extracting feature information with multiple different semantic dimensions from the current dialogue information to form a complete contextual understanding of the user's input. The current dialogue information refers to the natural language text input by the user in the current dialogue round of the dialogue agent. It can be a direct response to a question asked by the agent in the previous round or a new topic statement initiated by the user; it can be obtained through conventional text input methods. The semantic cognitive modeling unit is internally configured with multiple semantic analysis channels to perform multi-dimensional and structured semantic parsing of the same current dialogue information, thereby simultaneously obtaining information from the user's expression, including emotional feelings, intentions, beliefs, and factual basis, forming a complete understanding of the user's current dialogue state and providing a rich and accurate contextual information foundation for subsequent decision-making cognitive modeling.

[0032] In a preferred embodiment, the plurality of semantic analysis channels include, but are not limited to, a first semantic analysis channel, a second semantic analysis channel, and a third semantic analysis channel. These three semantic analysis channels can parse the same current dialogue information in parallel and extract feature information from different semantic dimensions respectively.

[0033] The first semantic analysis channel is configured to extract feature information of the first semantic dimension representing the user's feeling category description and emotional characteristics from the current dialogue information; that is, this channel focuses on the feeling category words and emotional descriptions used by the user in the dialogue, and infers the user's most likely core feeling state at present through multi-dimensional semantic calibration. In a specific implementation, the first semantic analysis channel can be implemented through a large language model configured with a first prompt word template, which contains key instructions to guide the model to identify and extract the user's feeling category description words and emotional characteristic words. For example, the first prompt word template may contain instructions such as "Please identify the feeling category description words and emotional characteristic words used by the user from the following user input, and infer the user's most likely core feeling state at present based on multi-dimensional semantic calibration"; when the user inputs "When the leader walked into the office today, he didn't look at me, I feel that he doesn't value me at all", the first semantic analysis channel can extract the feature information of the first semantic dimension representing the user's feeling category description and emotional characteristics through the first prompt word template, such as "feeling category description: feeling unvalued" and "emotional characteristics: disappointed / ignored".

[0034] The second semantic analysis channel is configured to extract feature information from the current dialogue information to represent the user's deep intentions and core assumptions in a second semantic dimension. The deep intentions refer to the user's underlying motivations, personal hopes, goals, and positive aspirations—the user's intrinsic driving forces; the core assumptions refer to the user's speculations or interpretations of facts—the user's subjective views on those facts. In a specific implementation, the second semantic analysis channel can be implemented using a large language model configured with a second prompt word template. This prompt word template contains key instructions guiding the model to distinguish and extract the user's deep intentions (intrinsic motivations and positive aspirations) and core assumptions (speculations or interpretations of facts). For example, the second prompt word template may contain instructions such as "Please extract from the following user input: ① Deep intent - the user's intrinsic motivation, personal hopes, goals, visions, and positive desires; ② Core assumptions - the user's speculations or interpretations of facts." Continuing with the previous example, the second semantic analysis channel can extract feature information representing the second semantic dimension of the user's deep intent and core assumptions through this second prompt word template, such as "Deep intent: Desire to gain the leader's recognition and attention" and "Core assumptions: The leader's behavior directly reflects the leader's level of attention to me, that is, believing that 'the leader doesn't look at me' is equivalent to 'the leader doesn't value me'."

[0035] The third semantic analysis channel is configured to extract feature information from the current dialogue information to represent the third semantic dimension of facts and observable events. That is, this channel focuses on specific events, behaviors, or phenomena in the user's statements that can be objectively verified, as well as the core assumptions the user holds based on these observable events (i.e., what the user believes the facts are and what they hope to do). In a specific implementation, the third semantic analysis channel can be implemented through a large language model configured with a third prompt word template. This third prompt word template contains key instructions for guiding the model to distinguish between observable events and the user's subjective inferences, extracting observable events and the user's core assumptions about the facts. For example, the third prompt word template may contain instructions such as "Please extract observable events (specific behaviors or phenomena in the user's statement that can be objectively verified) and the core assumptions held by the user based on the observable events (i.e., what the user believes the facts are)"; Continuing the example above, the third semantic analysis channel can extract feature information representing the third semantic dimension of facts and observable events through the third prompt word template, such as "Observable event: The leader did not look at the user when he walked into the office" and "The user's core assumptions about the facts: The user believes that the behavior of 'the leader not looking at me' means 'the leader does not value me' (this statement is the user's subjective inference, not the observable event itself)".

[0036] Through parallel parsing of the three semantic analysis channels mentioned above, the semantic cognitive modeling unit of this embodiment can perform structured understanding of user input from multiple semantic dimensions, extract user's feeling descriptions and emotional characteristics, deep intentions and core assumptions, as well as observable events and the user's core assumptions about the facts, providing a rich and accurate contextual information foundation for subsequent decision-making and behavioral cognition.

[0037] The decision-making cognitive modeling unit is configured to perform the following decision-making cognitive modeling: fusing feature information from multiple semantic dimensions to generate a contextual semantic package for the current dialogue; and extracting the minimum key factor from the contextual semantic package using a contextual anchor generator. The contextual semantic package is a complete contextual representation fusing multiple semantic dimensions, and in implementation, it is represented as a multi-dimensional feature vector and dynamically updated based on contextual awareness and awareness point recognition. The minimum key factor is a core semantic concept extracted a second time from the complete contextual representation and used to determine the response strategy for the current dialogue turn. Decision-making cognitive modeling refers to the process of fusing feature information from multiple semantic dimensions to generate a complete contextual representation for the current dialogue, and further extracting the core semantic concepts used to determine the response strategy for the current dialogue turn from this complete contextual representation. Through decision-making cognitive modeling, the system can accurately identify the core decision factors most valuable for driving the dialogue forward from multi-dimensional and structured contextual information, providing clear guidance for subsequent behavioral cognitive modeling.

[0038] The decision-making cognitive modeling unit specifically performs decision-making cognitive modeling in the following ways (A) to (B).

[0039] (A) Fusion processing and generation of contextual semantic packages: Specifically, the decision-making cognitive modeling unit first fuses the feature information from the multiple semantic dimensions of the semantic cognitive modeling unit to generate the contextual semantic package of the current dialogue. The contextual semantic package is a complete contextual representation that integrates multiple semantic dimensions, that is, it integrates information from multiple levels such as user emotions, deep intentions, and objective facts into a complete contextual description. In specific implementation, the fusion processing includes weighted fusion and conflict detection of feature information from the multiple semantic analysis channels.

[0040] Regarding the weighted fusion, the decision-making cognitive modeling unit can assign different weight coefficients to the feature information of each semantic dimension according to the needs of different dialogue scenarios or different dialogue stages. For example, in the initial stage of the dialogue, higher weights can be assigned to the feature information of the third semantic dimension (facts and observable events) to prioritize the establishment of an objective understanding of the user's situation; in the deeper stage of the dialogue, higher weights can be assigned to the feature information of the second semantic dimension (deep intentions and core assumptions) to explore the user's true motivations and inner beliefs in depth. The result of the weighted fusion constitutes a preliminary contextual representation of the current dialogue. In addition, the setting of the weight coefficients can be achieved through preset rules or dynamically determined through a pre-trained weight allocation model.

[0041] Regarding the conflict detection, it specifically includes, but is not limited to, the following: when there is an inconsistency between the feature information of the first semantic dimension output by the first semantic analysis channel and the feature information of the third semantic dimension output by the third semantic analysis channel, it is marked as a situational conflict, and the descriptive information of the situational conflict is included in the situational semantic package. In detail, the situational conflict refers to a deviation or contradiction between the user's subjective emotional feelings and objective facts. In coaching dialogue scenarios, such situational conflicts often reflect the user's cognitive blind spots or limiting beliefs, and are key nodes driving the dialogue deeper. For example, in the aforementioned example, the feature information of the first semantic dimension shows that the user has the emotional characteristic of "disappointment / being ignored," and the feature information of the third semantic dimension shows that the observable event is only "the leader did not look at the user when entering the office," while "the leader does not value me" is the core assumption held by the user based on this observable event, rather than the observable event itself. At this point, the decision-making cognitive modeling unit detects an inconsistency between the feature information of the first semantic dimension and the feature information of the third semantic dimension. The system marks this as a situational conflict and generates descriptive information for the situational conflict, such as "the user interprets the observable event 'the leader did not look at them' as 'the leader does not value them,' indicating a cognitive bias between perception and fact." This descriptive information of the situational conflict is incorporated into the situational semantic package, which, together with the weighted and fused preliminary situational representation, constitutes the complete situational semantic package of the current dialogue. The situational semantic package can be represented as a multi-dimensional feature vector and dynamically updated based on situational awareness and awareness point recognition.

[0042] Through the weighted fusion and conflict detection processing described above, the contextual semantic package not only contains fused information from multiple semantic dimensions, but also marks potential contextual conflicts in the user's cognition, providing rich and complete contextual information with decision-making guidance for the subsequent accurate extraction of the minimum key factors from multiple candidate semantic concepts.

[0043] (B) Extraction of the minimum key factor: Specifically, after generating the contextual semantic package, the decision-making cognitive modeling unit extracts the minimum key factor from the contextual semantic package through its built-in contextual anchor generator. The minimum key factor is the core semantic concept extracted from the complete contextual representation and used to determine the response strategy for the current dialogue turn. The so-called "secondary extraction" refers to first extracting feature information from multiple different semantic dimensions in the semantic cognitive modeling stage and fusing it into a complete contextual representation (first extraction), and then further refining the unique core semantic concept from this complete contextual representation (second extraction), forming a two-stage information distillation architecture of "package first, core later".

[0044] In a preferred embodiment, the context anchor generator includes a decision network based on a multi-map dynamic combination mechanism. This decision network is configured to dynamically combine multiple candidate semantic concepts parsed from the context semantic package, based on multiple pre-set coaching process maps in the domain knowledge base, to construct a cognitive exploration path adapted to the user's current context, and to determine the minimum critical factor from the multiple candidate semantic concepts based on this cognitive exploration path.

[0045] In detail, the context anchor generator first parses multiple candidate semantic concepts from the context semantic package. These candidate semantic concepts refer to selectable semantic topics or focuses within the context semantic package that can serve as the core of the response in the current dialogue turn. For example, in the context semantic package of the aforementioned example, the candidate semantic concepts that might be parsed include: "the user's feelings of disappointment," "the user's desire for the leader's attention," "the user's cognitive pattern of confusing feelings with facts," and "the leader's actual behavior," etc.

[0046] The pre-defined coaching process maps in the domain knowledge base refer to pre-constructed standardized dialogue process templates corresponding to different coaching dialogue stages and goals. The core theoretical basis of these coaching process maps is the NLP (Neuro-Linguistic Programming) principle that "a map is not a territory," meaning each process map provides a framework for understanding and guiding the dialogue, but is not an absolute description of the user's actual experience. In actual operation, the agent, based on the determination of the minimum critical factor, first assesses which coaching process the user's current situation belongs to and which step of that process needs to be executed. Then, it dynamically assembles the adapted coaching process maps like Lego bricks to build a "territory" that helps the user solve the problem. This is a dynamic process, not a static selection based on a fixed score. The decision network based on the multi-map dynamic combination mechanism, according to the constructed cognitive exploration path, determines the candidate semantic concept most helpful in driving the path forward from the multiple candidate semantic concepts as the minimum critical factor. For example, in the aforementioned example, if the cognitive exploration path is built towards the coaching direction of "helping users distinguish between feelings and facts and become aware of their own cognitive patterns", then the candidate semantic concept of "the cognitive pattern in which users confuse feelings with facts" may be identified as the minimum key factor in the current dialogue round.

[0047] Through the optimal selection process of the decision network based on the multi-map dynamic combination mechanism, the decision cognitive modeling unit can accurately identify the core semantic concepts that are most helpful in driving the coaching dialogue forward from multiple potential semantic topics, providing clear decision guidance for subsequent behavioral cognitive modeling. The minimum key factor is then transmitted to the behavioral cognitive modeling unit.

[0048] The behavioral cognitive modeling unit is configured to perform the following behavioral cognitive modeling: based on a preset role template and domain knowledge base, and using the minimum key factor as the core prompt, generate a core framework for the response content. Behavioral cognitive modeling refers to the process of transforming the core semantic concepts represented by the minimum key factor into a structured response content framework based on preset role settings and domain expertise. Through behavioral cognitive modeling, the system can implement the response strategy determined in the decision-making cognitive stage into an executable response framework in a manner that conforms to specific role specifications and domain requirements, providing a structured content foundation for subsequent dialogue response execution.

[0049] The behavioral cognition modeling unit specifically performs behavioral cognition modeling through the following methods (C) and (D).

[0050] (C) The basic generation of the core framework, specifically, is the core framework of the response content generated by the behavioral cognition modeling unit based on the preset role template and domain knowledge base, with the minimum key factor as the core prompt.

[0051] The role template refers to a pre-defined set of instructions that specifies the behavioral norms, language style, and professional capabilities of an AI agent in a dialogue. In this embodiment, as a coaching dialogue AI agent, the role template can be designed based on the core coaching competencies defined by the International Coach Federation (ICF), including but not limited to behavioral instructions for competency dimensions such as active listening, powerful questioning, direct communication, creating awareness, and designing actions. For example, the role template may include system instructions such as "You are an AI coach with ICF MCC Master Coach core competencies, strictly adhering to ICF ethical guidelines, but without ICF certification." By loading the role template, the behavioral cognitive modeling unit clarifies the role norms and behavioral boundaries to be followed when generating responses.

[0052] The domain knowledge base refers to a pre-constructed knowledge set containing coaching domain-specific tool structures and supervised data. In this embodiment, the domain knowledge base may contain knowledge points related to the SMART goal management principles and application methods of ICF-related coaching tools. For example, the domain knowledge base may contain structured knowledge items such as "S (Specific): Goals should be specific and clear, not vague and general," "M (Measurable): Goals should include quantifiable metrics to assess progress and results," "A (Attainable / Achievable): Goals should be challenging but achievable under realistic conditions," "R (Relevant): Goals should be related to the user's long-term vision and values," and "T (Time-bound): Goals should have clear completion timelines," as well as the application methods of ICF coaching tools in each stage of SMART. The domain knowledge base may also be configured with structured knowledge points with supervised tags, which may include coaching scenario tags, user emotional state tags, and cognitive stage tags, so as to perform more accurate knowledge point matching based on the current dialogue context during knowledge retrieval. By calling the domain knowledge base, the behavioral cognitive modeling unit can incorporate professional domain knowledge when generating responses, ensuring the professionalism and accuracy of the responses.

[0053] In its specific implementation, the behavioral cognitive modeling unit can use the minimum key factor as the core prompt word, and combine it with the instructions of the role template and relevant knowledge points retrieved from the domain knowledge base to construct a generated prompt word with a three-layer structure of "role setting - dialogue task - knowledge background". For example, when the minimum key factor is "the user's cognitive pattern of confusing feelings with facts", the behavioral cognitive modeling unit can construct the following generated prompt word: First, load the coach role instructions from the role template ("You are an AI coach with the core competencies of an ICF MCC master coach and strictly adhere to the ICF ethical guidelines, but do not have ICF certification. Your task is to help users become aware of their own thinking patterns through open-ended questions"); Second, set the current dialogue task with the minimum key factor as the core ("In the current dialogue, the user exhibits a cognitive pattern of confusing observable events with subjective feelings. Please generate a response around this core concept"); Finally, retrieve relevant knowledge points from the domain knowledge base (such as "distinguishing between facts and feelings is an important method for improving self-awareness" and "coaches should help users discover the differences between facts and feelings themselves through questioning, rather than directly pointing them out") as the knowledge background. The behavioral cognitive modeling unit inputs the generated prompt words into the large language model to generate the core framework of the response content.

[0054] The core framework of the response refers to the structured question guidance or response points that have not undergone detailed embellishment or security checks. For example, the core framework generated based on the aforementioned minimum key factors can include the following structured content: "① Empathetic response to the user's situation; ② Open-ended clarifying questions about the fact that 'the leader did not see the user's message' (e.g., 'Besides not being valued, are there any other possible explanations for the leader not seeing your message?'); ③ Inviting the user to recall past experiences of being valued (e.g., 'What things has your leader done in the past that made you feel valued?'); ④ Guiding the user to think about how to proactively communicate with their leader." This core framework provides a clear content structure and response points for subsequent dialogue execution.

[0055] (D) Adjustment of intervention intensity based on the output of the decision network: In a preferred embodiment, the behavioral cognitive modeling unit is further configured to adjust the intervention intensity parameter in the core framework of the response content based on the output of the decision network obtained from the decision cognitive modeling unit. The intervention intensity parameter is a parameter used to control the degree of guidance provided by the generated response to the user. Higher intervention intensity means more guiding content (such as questions, hints, invitations to reflection, etc.) is included in the response; lower intervention intensity means less guiding content is included in the response, giving the user more space for independent thinking. Specifically, the behavioral cognitive modeling unit adjusts the intervention intensity parameter as follows: when the output indicates that the current dialogue effectiveness is lower than a first preset effectiveness threshold, the proportion of guiding content is increased; when the output indicates that the current dialogue effectiveness is higher than a second preset effectiveness threshold, the proportion of guiding content is decreased.

[0056] The first preset efficacy threshold and the second preset efficacy threshold are pre-set efficacy scoring reference values, and the first preset efficacy threshold is less than the second preset efficacy threshold. When the output of the decision network indicates that the current dialogue efficacy is lower than the first preset efficacy threshold, it indicates that the coaching efficacy of the current dialogue is low, and the user may be in a state of confusion, stagnation, or avoidance, requiring more active assistance and stimulation from the coach. At this time, the behavioral cognitive modeling unit increases the proportion of guiding content in the core framework, such as adding more open-ended questions, direct feedback, or invitations to cognitive restructuring, to help the user overcome cognitive barriers and promote the dialogue forward.

[0057] When the output of the decision network indicates that the current dialogue effectiveness is higher than the second preset effectiveness threshold, it means that the coaching effectiveness of the current dialogue is high, and the user is in a state of active thinking, self-awareness, and autonomous exploration. At this time, the behavioral cognitive modeling unit reduces the proportion of guiding content in the core framework, reduces the number of questions and prompts, and increases listening, companionship, and affirmation content, giving the user more space for autonomous thinking and avoiding excessive intervention that could disrupt the user's state of autonomous awareness.

[0058] When the output of the decision network indicates that the current dialogue effectiveness is between the first preset effectiveness threshold and the second preset effectiveness threshold, the behavioral cognition modeling unit maintains the current intervention intensity parameter unchanged and continues to generate the core framework of the response content according to the basic generation method.

[0059] Through the adaptive adjustment of intervention intensity based on the output of the decision network, the behavioral cognitive modeling unit in this embodiment can dynamically adjust the guidance level of the response according to the actual progress of the dialogue, providing sufficient support when the user needs help and maintaining appropriate restraint when the user explores independently, thereby improving the overall coaching effect of the dialogue. Furthermore, the core framework of the response content generated by the behavioral cognitive modeling unit is then transmitted to the dialogue response execution unit for multi-level verification and optimization.

[0060] The dialogue response execution unit is configured to perform multi-level verification and optimization of the core framework based on preset value alignment rules, security and ethical rules, and expression detail rules, generating the final response content and pushing it to the user. This allows the user to input new dialogue information into the semantic cognitive modeling unit based on the final response content, forming an iterative cycle including semantic cognitive modeling, decision cognitive modeling, behavioral cognitive modeling, and dialogue response execution. Dialogue response execution refers to the process of progressively verifying and optimizing the core framework of the response content according to preset multi-dimensional rules, generating the final response content, and pushing it to the user, ultimately forming a closed-loop dialogue cognition. Through dialogue response execution, the system can push verified and optimized responses to the user while ensuring that the response content is consistent with the dialogue objective, safe and compliant, and naturally expressed. Simultaneously, it reintroduces subsequent user input based on the response into the cognitive modeling process, forming continuous cognitive iteration.

[0061] The dialogue response execution unit specifically performs dialogue response execution in the following ways (E) and (F).

[0062] (E) Multi-level verification and optimization, specifically, after receiving the core framework of the response content, the dialogue response execution unit performs multi-level verification and optimization on the core framework according to preset value alignment rules, security and ethics rules and expression detail rules.

[0063] In a preferred embodiment, the dialogue response execution unit includes a value alignment validator, a security and ethics validator, and an expression detail coherence scorer, which respectively perform the verification and optimization tasks of the above three dimensions. The three validators process the core framework step by step in the order of "value alignment → security and ethics → expression details", forming a structured multi-level verification pipeline.

[0064] The value alignment verifier is configured to check whether the content of the core framework is consistent with the preset dialogue goal according to preset value alignment rules. The value alignment rules refer to preset standards used to determine whether the response content conforms to the core value and target direction of the dialogue. In this embodiment, as a coaching dialogue agent, the value alignment rules may include, but are not limited to: whether the response content focuses on the user's preset goal (i.e., the target topic pointed to by the overall topic anchor), whether the response content reflects the coaching role rather than the direct advice role, and whether the response content helps improve the user's self-awareness rather than replacing the user's thinking. In specific implementation, the value alignment rules can be encoded into a series of value judgment prompts that can be executed by a large language model, used to score the goal consistency of the core framework. For example, the value alignment verifier can semantically match the core framework with the target topic pointed to by the overall topic anchor. If the matching degree is lower than a preset threshold, the core framework is adjusted to focus more on the user's preset goal; if the core framework contains content that directly provides solutions or suggestions, it is modified into an open-ended question guiding the user to explore solutions themselves. Through value alignment verification, it is ensured that the content of the core framework always revolves around the core goal of the dialogue and conforms to the value positioning of the coaching role.

[0065] The security and ethics verifier is configured to perform security risk detection on the content of the core framework according to preset security and ethics rules, and replace the content of the core framework with a preset security response template when a risk is detected. The security and ethics rules refer to preset standards used to identify and prevent potential security risks and ethical issues in dialogue. In this embodiment, the security and ethics rules may include, but are not limited to, keyword filtering rules, emotional dependence prevention rules, and privacy protection rules. In specific implementation, the security and ethics verifier first performs a security risk scan on the content of the core framework. The security risk scan includes matching and detecting preset dangerous keywords (such as those related to self-harm, suicide, violence, etc.) in the content of the core framework. If preset dangerous keywords are detected in the content of the core framework, a security response mechanism is triggered, replacing the current content of the core framework entirely with the preset security response template. The security response template is a pre-written, security-verified fixed response text, such as "The matter you mentioned is very important, and I may not be able to give the most suitable advice. If you would like, we can change the subject, or you can contact more professional support." After a security response is triggered, the dialogue response execution unit can directly output this security response template as the final response content, skipping the subsequent expression detail optimization stage.

[0066] Furthermore, the security and ethics verifier can be configured with emotional dependency protection rules. These rules automatically insert a reminder message suggesting the user contact a live coach or professional support when the number of dialogue rounds reaches a preset threshold. For example, when the number of dialogue rounds exceeds a preset threshold (e.g., 10 rounds), the security and ethics verifier automatically appends a reminder text such as "If you feel you need a deeper conversation, you might consider contacting a live coach for face-to-face communication" to the end of the core framework. This mechanism helps prevent users from developing excessive emotional dependence on the intelligent agent, ensuring their long-term mental health.

[0067] After passing value alignment and security and ethics checks, the expression detail coherence scorer performs final fluency optimization and expression detail adjustments on the content of the core framework. The expression detail coherence scorer is configured to optimize the fluency and adjust the diversity of ending sentence structures based on preset expression detail rules. These expression detail rules refer to preset standards used to improve the language quality and naturalness of the response text. In this embodiment, the expression detail rules may include, but are not limited to: sentence fluency rules (checking for grammatical errors, logical jumps, or redundant expressions and correcting them), ending sentence diversity rules (avoiding the use of the same or similar ending sentences in multiple rounds of dialogue to improve the naturalness and approachability of the conversation), and tone adaptation rules (adjusting the tone and wording of the response based on the dialogue context and the user's emotional state). In specific implementations, the expression detail coherence scorer can evaluate the language quality of the core framework content and refine and optimize the text based on the evaluation results. For example, randomizing or rotating the ending sentences can ensure sufficient diversity in the ending sentences of each round of dialogue, avoiding a mechanical and repetitive feeling for the user.

[0068] Through the step-by-step processing of the above three verifiers, the dialogue response execution unit can transform the core framework of the response content generated by the behavior cognition modeling unit into the final response content that is correct in value orientation, compliant in security ethics, and natural and fluent in language expression.

[0069] (F) Response Push and Iterative Loop: Specifically, after completing multi-level verification and optimization and generating the final response content, the dialogue response execution unit pushes the final response content to the user. The push method may include, but is not limited to, text display, voice broadcast, or a combination of both. After pushing the final response content, the dialogue response execution unit enters a waiting state, waiting to receive new dialogue information input by the user based on the final response content. When the user inputs new dialogue information, the new dialogue information is automatically routed to the semantic cognitive modeling unit as the current dialogue information for the next round of dialogue, restarting the semantic cognitive modeling process.

[0070] In this way, the entire system forms a cyclical iteration encompassing semantic cognitive modeling, decision-making cognitive modeling, behavioral cognitive modeling, and dialogue response execution. In each round of dialogue, the system re-performs multi-dimensional semantic parsing, context fusion and anchor point extraction, response framework generation, and multi-level verification and optimization based on the user's latest input. This allows the system to continuously deepen its understanding of the user's cognitive state in consecutive dialogue rounds, dynamically adjust response strategies, and continuously improve the relevance and inspiration of the dialogue, ultimately providing effective assistance in managing the user's goal outcomes.

[0071] In a preferred embodiment, in order to increase independent emotion scoring and tool scheduling, and to generate response strategies that adapt to cognitive and emotional states by combining minimum critical factors for dynamic decision-making, the dialogue agent system also includes, but is not limited to, an emotion intensity scoring unit independent of the multiple semantic analysis channels and a tool scheduler storing multiple preset dialogue tools.

[0072] The emotion intensity scoring unit is configured to quantitatively evaluate the emotion intensity of the current dialogue information and generate the user's current emotion intensity score. This emotion intensity scoring unit is independent of the multiple semantic analysis channels; that is, it does not participate in the multi-channel semantic parsing process in the aforementioned semantic cognitive modeling. Instead, it functions as an independent emotion perception module, specifically responsible for quantitatively evaluating the emotion intensity contained in the user input. This independent design ensures that the emotion intensity evaluation results are not interfered with by other dimensions (such as intent or factual information) in the semantic analysis channels, guaranteeing the objectivity and accuracy of emotion perception. Furthermore, the independent module design facilitates separate optimization and iteration of the emotion evaluation algorithm according to actual application needs, without affecting the normal operation of the semantic analysis channels.

[0073] In its implementation, the emotion intensity scoring unit can be implemented using a large language model configured with emotion scoring prompt word templates. These templates contain key instructions guiding the model to quantify the emotion intensity of the user's input text. For example, the template might include instructions such as, "Please rate the emotion intensity contained in the following user input text, ranging from 0 to 1, where 0 represents complete neutrality with no emotional fluctuation, and 1 represents extremely strong emotion. Please only output the score value and briefly explain the basis for the rating." When a user inputs "When my boss walked into the office today, he didn't look at me. I feel like he doesn't value me at all," as in the example above, the emotion intensity scoring unit can evaluate the input text and output a quantitative score, such as "0.65," along with a description of the rating basis, such as, "The user expressed a clear sense of loss and being ignored. The emotion type is clear and the intensity is moderate to high, but no extreme emotion words were found."

[0074] In this way, the emotion intensity scoring unit can independently quantify and evaluate the user's current emotional state in each round of dialogue, and generate the current emotion intensity score; the score is then transmitted to the tool scheduler, and together with the minimum critical factor, it serves as the decision basis for tool scheduling.

[0075] The tool scheduler, communicatively connected to the decision-making cognitive modeling unit and the emotion intensity scoring unit, is configured to select and execute a target dialogue tool from among the plurality of dialogue tools based on the minimum key factor and the current emotion intensity score. This selector generates a core framework or a generation strategy for the response content that is adapted to the minimum key factor and the current emotion intensity score. The plurality of dialogue tools refers to a pre-configured set of dialogue behavior templates, each corresponding to different dialogue strategies and response styles. In this embodiment, as a coaching dialogue agent, the plurality of dialogue tools includes at least a first dialogue tool and a second dialogue tool.

[0076] The first dialogue tool is configured to adopt a dialogue strategy that focuses on confirming the user's current state without directly addressing the problem. Specifically, this tool is suitable for dialogue scenarios where the user exhibits strong emotional fluctuations or is in a cognitive defensive state. Its core function is to gently acknowledge the user's current feelings and state, offering emotional acceptance and support without rushing to resolve the issue or delve deeper into the problem. For example, when a user expresses strong feelings of disappointment or frustration, the first dialogue tool can generate a response strategy framework such as, "It sounds like you're feeling quite conflicted right now, perhaps a little disappointed and misunderstood. We don't need to rush to solve anything; you can pause here for a moment and talk to me about how you're feeling right now?" This dialogue strategy helps establish a safe dialogue atmosphere, reduces the user's defensiveness, and creates conditions for further in-depth exploration.

[0077] The second dialogue tool is configured to employ dialogue strategies that guide users to engage in physical awareness or emotional release. Specifically, this tool is suitable for dialogue scenarios where users are willing to explore themselves and their emotional state is relatively stable. Its core function is to help users temporarily detach from the cognitive level by guiding them to focus on bodily sensations or express their emotions concretely, thereby gaining new insights through physical awareness or emotional release. For example, when a user repeatedly gets stuck in a certain thought pattern and finds it difficult to break free, the second dialogue tool can generate a response strategy framework such as, "Regarding this matter, we can try to feel it from a different perspective. Can you try to feel where on your body this feeling is most pronounced when you say 'My boss doesn't value me'? Is it your chest, shoulders, or other places?" This dialogue strategy helps break the user's inherent cognitive cycle and gain new insights through a mind-body connection.

[0078] In its implementation, the tool scheduler receives the minimum key factor from the decision-making cognitive modeling unit and the current emotion intensity score from the emotion intensity scoring unit in each round of dialogue. Based on these two input signals, the tool scheduler selects one of the multiple dialogue tools as the target dialogue tool and executes that target dialogue tool to generate a core framework of the response content or a generation strategy for the core framework that is adapted to the minimum key factor and the current emotion intensity score.

[0079] For example, when the minimum critical factor indicates that the current response strategy should be "helping the user distinguish between facts and feelings," and the current emotional intensity score is moderately low (e.g., 0.3-0.5), the user is in a relatively rational and willing-to-explore state. The tool scheduler can select the second dialogue tool to guide the user to explore the difference between facts and feelings through bodily perception. When the minimum critical factor also indicates "helping the user distinguish between facts and feelings," but the current emotional intensity score is high (e.g., above 0.6), the user is in an emotionally agitated state. In this case, directly exploring the cognitive level may be counterproductive. The tool scheduler can select the first dialogue tool to first confirm the user's emotional state, provide sufficient emotional acceptance and support, and then switch to other tools for in-depth exploration in subsequent dialogue rounds after the user's emotional intensity has fallen back to an appropriate range.

[0080] In a preferred embodiment, to set a dialogue target anchor point, a certain degree of topic deviation is allowed. When deviation is detected, the system prioritizes responding to the user's current state according to coaching principles, and then gently guides the dialogue back within the framework based on the overall topic anchor point. Updates are only initiated by the user, ensuring the dialogue focuses on the core topic. The dialogue agent system also includes a topic anchor point setting unit communicatively connected to the semantic cognitive modeling unit. This topic anchor point setting unit is configured to: extract the overall topic anchor point from the user's input dialogue information at the start of the dialogue; and in subsequent dialogue rounds, correct detected dialogue deviations based on the overall topic anchor point to guide the dialogue back to the target topic indicated by the overall topic anchor point. The overall topic anchor point refers to a semantic marker extracted from the user's stated goal at the initial stage of the dialogue, used to identify the core focus topic throughout the dialogue. The target topic indicated by the overall topic anchor point is the core topic or goal proposed by the user at the start of the dialogue, which they hope to continuously focus on and advance throughout the dialogue. By setting a general anchor point for the topic, the system can keep the conversation focused on the user's core objective in subsequent rounds of dialogue, avoiding significant deviations from the intended direction due to user topic jumps, small talk, or shifts in attention.

[0081] In its specific implementation, the topic anchor setting unit can receive semantic information obtained by the semantic cognitive modeling unit after parsing the user's initial input at the beginning of the dialogue, and extract key semantic concepts that represent the user's core goal as the overall topic anchor. At the beginning of the dialogue, the semantic cognitive modeling unit can output a fixed opening statement to guide the user to state the goal they wish to discuss. This fixed opening statement may include a welcome message, a brief introduction to the agent's functions, and an explanation of the benefits of clarifying the goal, such as: "Hello, I am the SMART coach, and I am happy to have this conversation with you today. Before we begin, I would like to understand what your current goal or core issue is that you truly want to focus on? If you were to summarize it in one sentence, how would you describe it?" When the user inputs "I want to improve my workplace communication skills, especially when communicating with superiors; I am always afraid to express my thoughts" at the beginning of the dialogue, the topic anchor setting unit can extract the overall topic anchor from this initial input, such as "improve workplace communication skills," with the target theme being "improving communication and expression skills with superiors."

[0082] Once extracted, the overall topic anchor is stored as a global reference variable for the current dialogue session, used for dialogue deviation detection and correction in subsequent dialogue rounds. Dialogue deviation refers to a situation where the semantic relevance between the user's input in the current dialogue round and the target topic pointed to by the overall topic anchor is lower than a preset relevance threshold; that is, the user's current input has significantly deviated from the core target topic set at the beginning of the dialogue. In specific implementations, the topic anchor setting unit can determine whether dialogue deviation exists by calculating the semantic similarity or semantic distance between the user's input in the current dialogue round and the overall topic anchor. The calculation of semantic similarity can be implemented using a pre-trained semantic matching model or a vector space model.

[0083] For example, in the aforementioned dialogue session with "improving workplace communication skills" as the overall topic anchor, if a user suddenly enters "The stock market has been fluctuating a lot lately, and I'm considering whether to adjust my investment portfolio" at a certain round of the dialogue, the topic anchor setting unit calculates the semantic similarity between the input content and the overall topic anchor "improving workplace communication skills". If the calculation result shows that the semantic similarity between the two is lower than the preset relevance threshold (such as 0.3), it is determined that a dialogue deviation has occurred.

[0084] When a deviation from the conversation is detected, the topic anchor point setting unit responds to the user's current state based on preset coaching principles. After giving the user sufficient emotional acceptance and companionship, it generates a transitional response framework based on the topic anchor point to gently guide the current deviation back to the core goal, so as to guide the conversation back to the target topic pointed to by the topic anchor point. For example, in response to the aforementioned dialogue deviations, when the user's current input does not contain strong emotional expression, the topic anchoring unit can generate a corrective response framework such as, "You just mentioned stock investment, which is a very important topic. We previously discussed improving workplace communication skills. Do you think there are any similarities between making investment decisions and expressing your thoughts to superiors in the workplace?" When the user's current input contains strong emotional expression, the topic anchoring unit prioritizes empathetic responses and gentle companionship to the user's emotional state. After the user's emotional intensity subsides, it then gently and naturally guides the dialogue back onto the intended track by finding connections between the current topic and the core objective. This corrective response framework does not abruptly interrupt or ignore the user's topic, but rather guides the dialogue back onto the intended track in a gentle and natural manner by finding connections between the current topic and the core objective, reflecting the professional principle of respecting the user and following their pace in coaching dialogues.

[0085] The topic anchor point setting unit is further configured to update the overall topic anchor point only when a clear instruction from the user to actively modify the target topic is received during the dialogue; otherwise, the overall topic anchor point remains unchanged. This update constraint ensures the stability of the dialogue target topic and user-led engagement. In actual coaching dialogue scenarios, users may proactively realize during the dialogue that their true goal is not what was initially proposed, or they may wish to adjust the direction of the dialogue due to changes in the external environment. In such cases, allowing users to update the overall topic anchor point when they explicitly express their intention to modify the target topic respects the user's right to self-determination while ensuring the flexibility of the dialogue framework.

[0086] In its implementation, the topic anchor setting unit can determine whether a user has expressed a clear intention to actively modify the target topic by performing intent recognition on user input. A clear instruction refers to a user's explicit statement of intent to change the core objective of the current conversation, expressed in clear language, such as "I've thought about it again, what I really want to talk about is...", "Can we change our objective? I think something more important might be...", or "My current objective is different from what I initially said." The topic anchor setting unit can classify user input intent using a large language model configured with intent recognition prompts to determine whether it belongs to the "actively modifying the target topic" intent category. Only when the recognition result is positive will the topic anchor be re-extracted from the user's current input and updated; otherwise, the original topic anchor remains unchanged. For vague expressions, indirect hints, or user input that may involve target adjustment but is not clear enough, the topic anchor setting unit does not actively trigger anchor updates but can instead invite the user to further clarify their true intent in the generated response.

[0087] Through the extraction, deviation correction, and update constraint mechanism of the topic anchor point, the topic anchor point setting unit of this embodiment can effectively maintain the core focus of the dialogue throughout the entire dialogue process. This not only prevents the dialogue from deviating from the predetermined goal due to user topic jumps or casual conversation, but also fully respects the user's right to actively adjust the direction of the dialogue, ensuring the efficiency and user-led nature of the dialogue.

[0088] In a preferred embodiment, to perform emotional self-checking on the response, compare it with the user's emotional state, and optimize and update it to improve the empathic adaptability of the response, the dialogue response execution unit is further configured to: after generating the final response content and before pushing the final response content to the user, perform an emotional intensity quantification assessment on the final response content to obtain a response emotional intensity score, and perform the same emotional intensity quantification assessment on the current dialogue information to obtain the user's current emotional intensity score; then compare the response emotional intensity score with the current emotional intensity score, and optimize and update the final response content based on the comparison result. Specifically, the dialogue response execution unit is also configured with a response emotional self-checking and two-way optimization mechanism. This mechanism is executed after generating the final response content and before pushing the final response content to the user, which is equivalent to adding an emotional quality check in the final stage before the response output.

[0089] The response emotional intensity score refers to a numerical value obtained by quantitatively assessing the emotional intensity contained in the response text to be pushed to the user. The purpose of this assessment is to determine whether the response text carries an appropriate degree of emotional coloring, ensuring that the response is not completely neutral, indifferent, or mechanical, but rather possesses a certain emotional warmth and can establish an emotional connection with the user. The user's current emotional intensity score refers to a numerical value obtained by quantitatively assessing the emotional intensity of the content entered by the user in the most recent round of conversation, reflecting the intensity of the user's current emotional state.

[0090] In a specific implementation, the dialogue response execution unit can call the emotion intensity scoring unit described in the preceding embodiments to quantitatively evaluate the emotion intensity of the final response content and the current dialogue information, respectively, to obtain the response emotion intensity score and the user's current emotion intensity score. Since the emotion intensity scoring unit is independent of the semantic analysis channel, its evaluation results are objective and consistent, suitable for comparing the emotion intensity of user input and system responses using the same standard. For example, when a user inputs "My boss didn't look at me when he walked into the office today; I feel like he doesn't value me at all," the emotion intensity scoring unit might evaluate the user's current emotion intensity score to be 0.65, indicating that the user is in a moderately high emotional state. If the generated final response is "Please describe in detail what your boss did at the time," this response leans towards a neutral factual inquiry, and its response emotion intensity score might only be 0.15; there is a significant difference in emotion intensity between the two.

[0091] After obtaining the response sentiment intensity score and the user's current sentiment intensity score, the dialogue response execution unit compares the response sentiment intensity score with the current sentiment intensity score, and optimizes and updates the final response content based on the comparison result. In a preferred embodiment, optimizing and updating the final response content based on the comparison result includes, but is not limited to: when the absolute value of the difference between the response sentiment intensity score and the current sentiment intensity score exceeds a preset difference threshold, adjusting the wording or expression of the final response content to raise or lower the response sentiment intensity score to a range that matches the current sentiment intensity score.

[0092] The preset difference threshold refers to a pre-set critical value used to determine whether the gap between the response's emotional intensity and the user's emotional state is too large and requires optimization. For example, the preset difference threshold can be set to 0.3. When the absolute value of the difference exceeds this threshold, it indicates a significant mismatch between the emotional intensity of the response and the user's emotional state, requiring optimization. The optimization update direction is to "increase or decrease" the response's emotional intensity score to adapt to two different mismatch situations: when the response's emotional intensity score is significantly lower than the current emotional intensity score (i.e., the response is too "calm"), the wording or expression of the final response content is adjusted to increase the response's emotional intensity score, making the response more emotionally warm and empathetic; when the response's emotional intensity score is significantly higher than the current emotional intensity score (i.e., the response is too "excited"), the wording or expression of the final response content is adjusted to decrease the response's emotional intensity score, avoiding the response appearing overly exaggerated or inappropriate.

[0093] In practical implementation, the adjustment of wording or expression can be achieved through a large language model configured with emotion adjustment prompt templates. For example, when it is necessary to increase the emotional intensity of a response, the following prompt template can be used: "Please increase the emotional temperature of this response by adding appropriate empathetic expressions, emotional vocabulary, or gentle tone words, without changing the core semantics of the following response text, to make it more emotionally resonant and approachable"; when it is necessary to decrease the emotional intensity of a response, the following prompt template can be used: "Please reduce the emotional intensity of this response by reducing emotionally charged words and using a more peaceful expression, without changing the core semantics of the following response text, to make it more rational and composed."

[0094] The matching interval refers to a reasonable numerical range centered on the current emotional intensity score, rather than requiring the response emotional intensity score to be exactly equal to the user's emotional intensity score. For example, the matching interval can be set within ±0.15 of the current emotional intensity score. When the optimized and updated response emotional intensity score falls within this interval, it is considered that the response and the user's emotional state are basically matched, and no further adjustment is needed. Continuing the previous example, the user's current emotional intensity score is 0.65, and the response emotional intensity score is 0.15. The absolute value of the difference between the two is 0.50, which exceeds the preset difference threshold of 0.3. At this time, the dialogue response execution unit determines that optimization and adjustment are needed, and since the response emotional intensity score is significantly lower than the user's emotional intensity score, the response needs to be processed to enhance its emotional impact. After processing by the emotional adjustment prompt word template, the original response "Please describe in detail what the leader did at the time" can be optimized and updated to "I heard about this, and I feel that you may have been a little upset at the time. Can you tell me in detail what the scene was like when the leader walked into the office?" The optimized response, while maintaining the core intent of the question, increased attention to user feelings and empathetic expression, raising its emotional intensity score to approximately 0.55, falling within the matching range.

[0095] Through the aforementioned response emotion self-checking and two-way optimization mechanism, the dialogue response execution unit of this embodiment can automatically detect and correct the mismatch between the response and the user's emotional state in the final stage before the response is output. This ensures that the response finally pushed to the user is not only accurate and compliant, but also resonates with the user on an emotional level, significantly improving the empathy adaptability of the dialogue and the user experience.

[0096] In a preferred embodiment, in order to construct a user cognitive profile using the historical sequence of minimum key factors to assist in the generation of subsequent dialogues and achieve long-term personalized cognition, the dialogue agent system further includes a user profile update unit that is communicatively connected to the decision-making cognitive modeling unit and the behavior cognitive modeling unit, respectively. The user profile update unit is configured to generate or update the user's cognitive profile based on the historical sequence of the minimum key factors generated during the most recent multi-round dialogues with the user. The cognitive profile is used as an auxiliary input for the behavior cognitive modeling unit in subsequent dialogues with the user.

[0097] The historical sequence of the minimum key factors refers to the time sequence formed by arranging the minimum key factors extracted and output by the decision-making cognitive modeling unit in each round of dialogue during multiple consecutive rounds of conversation with the same user, according to the chronological order of the dialogue rounds. Since the minimum key factors are the core semantic concepts that determine the response strategy in each round of dialogue, their historical sequence can reflect the user's cognitive changes, the evolution of core concerns, and the repetitive characteristics of thinking patterns during the dialogue process. Therefore, the historical sequence of the minimum key factors is a high-value data source for constructing user cognitive profiles.

[0098] In its implementation, the user profile update unit is communicatively connected to the decision-making cognitive modeling unit. After each round of dialogue completes decision-making cognitive modeling, it automatically acquires the minimum key factor generated in that round of dialogue and appends it to the dialogue log corresponding to the current user. The dialogue log can maintain a sliding window to retain the minimum key factor generated in the most recent N rounds of dialogue (e.g., the most recent 10 rounds), forming a historical sequence of the minimum key factor. This sliding window mechanism ensures that the data used for profile updates is timely enough to reflect recent changes in the user's cognitive state, while also avoiding the introduction of excessive noise due to a long dialogue history.

[0099] The cognitive profile refers to a user model that provides a structured description of a user's cognitive patterns, thinking habits, common topics of interest, and emotional response characteristics during a conversation. This cognitive profile serves as auxiliary input for the behavioral cognitive modeling unit in subsequent conversations with the user, making the generated responses more aligned with the user's personalized characteristics and cognitive style.

[0100] In its implementation, the user profile update unit can analyze and summarize the historical sequence of the minimum key factor (MKF) using a large language model configured with profile generation prompt word templates, thereby generating or updating the cognitive profile. The profile generation prompt word template may contain key instructions guiding the model to extract user cognitive features from the MKF historical sequence. For example, the template may include instructions such as, "The following is a sequence of core semantic concepts generated in recent multi-round conversations with a user. Based on these concept sequences, analyze the user's cognitive pattern characteristics, including: recurring topics of interest, common thinking patterns or cognitive biases, typical responses to different problems, and trends in user cognitive pattern changes. Summarize the user's cognitive profile in a structured manner."

[0101] For example, after multiple rounds of dialogue with a user, the historical sequence of the minimum critical factors might include: "the user's cognitive pattern of confusing facts with feelings," "the user's sensitivity to authoritative evaluations," "the user's tendency to avoid direct action planning," "the user beginning to distinguish between facts and feelings," and "the user independently proposing a communication plan," etc. The user profile update unit inputs this historical sequence into the profile generation prompt word template, and after analysis and processing by the large language model, it can generate the following cognitive profile: "In the early stages of the dialogue, the user exhibits a cognitive pattern of confusing observable facts with subjective feelings, is highly sensitive to evaluations from authoritative sources, and shows an avoidance tendency when faced with scenarios requiring specific action planning. As the dialogue progresses, the user gradually demonstrates the ability to distinguish between facts and feelings and begins to independently explore solutions. The user's deeper needs revolve around the two core themes of 'gaining recognition' and 'building confidence.'" After the cognitive profile is initially generated, as subsequent dialogues continue and new minimum key factors are continuously generated, the user profile update unit can update the cognitive profile according to a preset update frequency (e.g., once every 5 rounds of dialogue) or after each dialogue, based on the latest historical sequence of the minimum key factors, so as to continuously reflect the user's latest cognitive state and changes in thinking patterns.

[0102] After the cognitive profile is generated or updated, it is stored as a user cognitive file associated with the current user. In subsequent conversations with the same user, the user profile updating unit transmits the cognitive profile to the behavioral cognitive modeling unit as auxiliary input. Specifically, when the behavioral cognitive modeling unit generates the core framework of the response content based on the minimum key factors, the role template, and the domain knowledge base, the cognitive profile serves as auxiliary input, providing additional contextual information about the user's cognitive patterns, thinking preferences, and common topics of interest. For example, when generating the response framework, the behavioral cognitive modeling unit can refer to the characteristic recorded in the cognitive profile that "the user is sensitive to authoritative evaluations," and adopt a more gentle and probing questioning approach when discussing topics involving authority relationships; or refer to the characteristic that "the user has a tendency to avoid action planning," and provide more encouragement and gradual guidance when inviting the user to develop an action plan.

[0103] It should be noted that the cognitive profile serves as an auxiliary input, its role being to personalize and fine-tune the core response strategy indicated by the minimum key factor, rather than replacing the minimum key factor as the core prompt. The main direction and core content of the response are still determined by the minimum key factor; the cognitive profile provides personalized reference information on "how to express and execute the response strategy in a way that is more suitable for the user."

[0104] Through the aforementioned user profile update unit, the dialogue intelligence system of this embodiment can continuously accumulate a deep understanding of the user's cognitive patterns and thinking characteristics during long-term dialogues with the user, and feed this understanding back to the response generation stage, thereby achieving long-term personalization of dialogues and continuous adaptation at the cognitive level. This makes the system not only an intelligent responder in single-turn dialogues, but also a long-term dialogue partner that can accompany the user's cognitive growth.

[0105] In practical deployment, the dialogue agent system of this embodiment can undergo behavior verification through a preset set of test cases before deployment. For example, test cases can be set to verify whether the first dialogue tool is correctly triggered under specific user input, whether the tool scheduler correctly disables the second dialogue tool under consecutive rounds of high emotional intensity scenarios, and whether the dialogue response execution unit correctly outputs a safe response under user input exceeding the coach's range. The system needs to pass the verification of the above test case set before deployment to ensure that the behavior of each unit meets the design expectations.

[0106] Furthermore, in the initial dialogue phase of this embodiment, the dialogue agent system can first output a fixed opening statement through the semantic cognitive modeling unit to guide the user to propose the current goal they wish to explore. This fixed opening statement may include a welcome message, a brief introduction to the agent's functions, and an explanation of the benefits of clarifying the goal, allowing the user to begin the dialogue with a clear understanding of the dialogue framework. Simultaneously, the domain knowledge base can be configured with structured knowledge points marked with supervision tags. These supervision tags may include coach scenario labels, user emotional state labels, and cognitive stage labels, enabling more accurate knowledge point matching based on the current dialogue context during knowledge retrieval.

[0107] Example 2 This embodiment provides a method for implementing a dialogue agent system based on cognitive modeling as described in Embodiment 1, such as... Figure 2 As shown, the method for implementing the dialogue agent may include, but is not limited to, the following steps S1 to S6.

[0108] S1. Receive the current dialogue information input by the user, and then execute step S2.

[0109] S2. Based on the current dialogue information, semantic cognitive modeling is performed as follows: the current dialogue information is parsed through multiple semantic analysis channels to obtain feature information of multiple semantic dimensions, and then step S3 is executed, wherein each of the semantic analysis channels is configured to extract and output feature information with different semantic dimensions, and the multiple semantic dimensions correspond one-to-one with the multiple semantic analysis channels.

[0110] S3. Based on the feature information of the multiple semantic dimensions, decision-making cognitive modeling is performed as follows: the feature information of the multiple semantic dimensions is fused to generate the contextual semantic package of the current dialogue, and the minimum key factor is extracted from the contextual semantic package through the contextual anchor generator. Then, step S4 is executed. The contextual semantic package is a complete contextual representation that integrates multiple semantic dimensions, and can be represented as a multi-dimensional feature vector in implementation. It is dynamically updated based on contextual awareness and awareness point recognition. The minimum key factor is a core semantic concept extracted twice from the complete contextual representation and used to determine the response strategy for the current dialogue turn.

[0111] S4. Based on the minimum key factor, perform behavioral cognition modeling as follows: Based on the preset role template and domain knowledge base, use the minimum key factor as the core prompt to generate the core framework of the response content, and then execute step S5.

[0112] S5. Based on preset value alignment rules, security and ethics rules, and expression detail rules, perform multi-level verification and optimization on the core framework, generate the final response content and push it to the user, and then execute step S6.

[0113] S6. Receive new dialogue information input by the user based on the final reply content, return to the execution steps S2 to S5, forming a loop iteration including semantic cognitive modeling, decision cognitive modeling, behavioral cognitive modeling and dialogue reply execution.

[0114] The technical details of each step in this embodiment and the overall technical effect of the method can be derived by referring to the conventional derivation in Embodiment 1, and will not be repeated here.

[0115] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dialogue agent system based on cognitive modeling, characterized in that, It includes a semantic cognitive modeling unit, a decision cognitive modeling unit, a behavioral cognitive modeling unit, and a dialogue response execution unit that are sequentially connected by communication. The semantic cognitive modeling unit is configured to receive current dialogue information input by the user and perform the following semantic cognitive modeling: parse the current dialogue information through multiple semantic analysis channels to obtain feature information of multiple semantic dimensions, wherein each of the semantic analysis channels is configured to extract and output feature information with different semantic dimensions, and the multiple semantic dimensions correspond one-to-one with the multiple semantic analysis channels; The decision-making cognitive modeling unit is configured to perform the following decision-making cognitive modeling: fuse the feature information of the multiple semantic dimensions to generate the contextual semantic package of the current dialogue, and extract the minimum key factor from the contextual semantic package through the contextual anchor generator. The contextual semantic package is a complete contextual representation that fuses multiple semantic dimensions, and is represented as a multi-dimensional feature vector in implementation, and is dynamically updated based on contextual awareness and awareness point recognition. The minimum key factor is a core semantic concept extracted twice from the complete contextual representation and used to determine the response strategy for the current dialogue turn. The behavioral cognition modeling unit is configured to perform the following behavioral cognition modeling: based on a preset role template and domain knowledge base, and using the minimum key factor as the core prompt, generate the core framework of the response content; The dialogue response execution unit is configured to perform multi-level verification and optimization on the core framework based on preset value alignment rules, security and ethics rules, and expression detail rules, generate the final response content, and push it to the user, so that the user can input new dialogue information to the semantic cognitive modeling unit based on the final response content, forming a cyclical iteration including semantic cognitive modeling, decision cognitive modeling, behavioral cognitive modeling, and dialogue response execution.

2. The dialogue agent system as described in claim 1, characterized in that, The plurality of semantic analysis channels include a first semantic analysis channel, a second semantic analysis channel, and a third semantic analysis channel; The first semantic analysis channel is configured to extract feature information of the first semantic dimension from the current dialogue information to characterize the user's feeling category description and emotional features; The second semantic analysis channel is configured to extract feature information of the second semantic dimension from the current dialogue information to characterize the user's deep intent and core assumptions; The third semantic analysis channel is configured to extract feature information of the third semantic dimension from the current dialogue information to represent facts and observable events; The fusion process includes weighted fusion and conflict detection of feature information from the multiple semantic analysis channels to form the contextual semantic package.

3. The dialogue agent system as described in claim 2, characterized in that, The conflict detection includes: when there is an inconsistency between the feature information of the first semantic dimension output by the first semantic analysis channel and the feature information of the third semantic dimension output by the third semantic analysis channel, it is marked as a situational conflict, and the description information of the situational conflict is included in the situational semantic package, so that the situational anchor generator can use it as a decision basis when extracting the minimum key factor.

4. The dialogue agent system as described in claim 1, characterized in that, The context anchor generator includes a decision network based on a multi-map dynamic combination mechanism. This decision network is configured to dynamically combine multiple candidate semantic concepts parsed from the context semantic package, based on multiple pre-set coaching process maps in the domain knowledge base, to build a cognitive exploration path that is adapted to the user's current context, and to determine the minimum critical factor from the multiple candidate semantic concepts based on this cognitive exploration path.

5. The dialogue agent system as described in claim 4, characterized in that, The behavioral cognitive modeling unit is also configured to: Based on the output of the decision network based on the multi-map dynamic combination mechanism, the intervention intensity parameter in the core framework of the response content is adjusted as follows: when the output indicates that the current dialogue effectiveness is lower than the first preset effectiveness threshold, the proportion of guiding content is increased; when the output indicates that the current dialogue effectiveness is higher than the second preset effectiveness threshold, the proportion of guiding content is decreased.

6. The dialogue agent system as described in claim 1, characterized in that, It also includes an emotion intensity scoring unit that is independent of the multiple semantic analysis channels and a tool scheduler that stores multiple preset dialogue tools; The emotion intensity scoring unit is configured to perform a quantitative evaluation of the emotion intensity of the current dialogue information and generate the user's current emotion intensity score. The tool scheduler, which is communicatively connected to the decision-making cognitive modeling unit and the emotion intensity scoring unit, is configured to select and execute a target dialogue tool from the plurality of dialogue tools based on the minimum key factor and the current emotion intensity score, in order to generate a core framework of the response content or a generation strategy of the core framework that is adapted to the minimum key factor and the current emotion intensity score.

7. The dialogue agent system as described in claim 6, characterized in that, The plurality of dialogue tools includes at least a first dialogue tool and a second dialogue tool; The first dialogue tool is configured to adopt a dialogue strategy that is used to confirm the user's current state rather than directly solving the problem; The second dialogue tool is configured to employ dialogue strategies designed to guide users in physical awareness or emotional release. The tool scheduler is also configured to: when the current emotion intensity score exceeds a preset score threshold, force the currently selected dialogue tool to be switched to the first dialogue tool or a preset safe dialogue tool. Furthermore, when the current emotion intensity score exceeds the preset score threshold in multiple consecutive dialogue rounds, the second dialogue tool is disabled, and a preset guaranteed dialogue strategy is triggered.

8. The dialogue agent system as described in claim 1, characterized in that, It also includes a topic anchor setting unit with a communication connection to the semantic cognitive modeling unit. The topic anchor setting unit is configured to: extract a total topic anchor from the user-input dialogue information at the start of the dialogue; and in subsequent dialogue rounds, correct detected dialogue deviations based on the total topic anchor to guide the dialogue back to the target topic indicated by the total topic anchor. The topic anchor setting unit is further configured to: update the total topic anchor only when a clear instruction from the user to actively modify the target topic is received during the dialogue; otherwise, the total topic anchor remains unchanged. When a dialogue deviation is detected, the topic anchor setting unit prioritizes responding to the user's current state according to preset coaching principles, and then gently guides the dialogue back into the framework based on the total topic anchor. And / or, it further includes a user profile update unit that is communicatively connected to the decision cognitive modeling unit and the behavior cognitive modeling unit, respectively, wherein the user profile update unit is configured to: generate or update the cognitive profile of the user based on the historical sequence of the minimum key factors generated during the most recent multi-round dialogue with the user, the cognitive profile being used as an auxiliary input of the behavior cognitive modeling unit in subsequent dialogues with the user; And / or, the dialogue response execution unit includes a valuable alignment validator, a security and ethics validator, and an expression detail coherence scorer; The value alignment verifier is configured to check whether the content of the core framework is consistent with the preset dialogue target according to the preset value alignment rules. The security and ethics verifier is configured to perform security risk detection on the content of the core framework according to preset security and ethics rules, and replace the content of the core framework with a preset security response template when a risk is detected. The expression detail coherence scorer is configured to optimize the fluency of the core framework and adjust the diversity of ending sentence structures according to preset expression detail rules.

9. The dialogue agent system as described in claim 1, characterized in that, The dialogue response execution unit is further configured to: after generating the final response content and before pushing the final response content to the user, perform an emotional intensity quantification assessment on the final response content to obtain a response emotional intensity score, and perform the emotional intensity quantification assessment on the current dialogue information to obtain the user's current emotional intensity score, then compare the response emotional intensity score with the current emotional intensity score, and optimize and update the final response content based on the comparison result.

10. A method for implementing a dialogue agent based on cognitive modeling, characterized in that, It includes the following steps S1 to S6: S1. Receive the current dialogue information input by the user, and then execute step S2; S2. Based on the current dialogue information, semantic cognitive modeling is performed as follows: the current dialogue information is parsed through multiple semantic analysis channels to obtain feature information of multiple semantic dimensions, and then step S3 is executed, wherein each of the semantic analysis channels is configured to extract and output feature information with different semantic dimensions, and the multiple semantic dimensions correspond one-to-one with the multiple semantic analysis channels; S3. Based on the feature information of the multiple semantic dimensions, decision-making cognitive modeling is performed as follows: the feature information of the multiple semantic dimensions is fused to generate the contextual semantic package of the current dialogue, and the minimum key factor is extracted from the contextual semantic package through the contextual anchor generator. Then, step S4 is executed. The contextual semantic package is a complete contextual representation that integrates multiple semantic dimensions, and is represented as a multi-dimensional feature vector in implementation. It is dynamically updated based on contextual awareness and awareness point recognition. The minimum key factor is a core semantic concept extracted twice from the complete contextual representation and used to determine the response strategy for the current dialogue round. S4. Based on the minimum key factor, perform behavioral cognition modeling as follows: Based on the preset role template and domain knowledge base, use the minimum key factor as the core prompt to generate the core framework of the response content, and then execute step S5; S5. Based on the preset value alignment rules, security and ethics rules, and expression detail rules, perform multi-level verification and optimization on the core framework, generate the final response content and push it to the user, and then execute step S6; S6. Receive new dialogue information input by the user based on the final reply content, return to the execution steps S2 to S5, forming a loop iteration including semantic cognitive modeling, decision cognitive modeling, behavioral cognitive modeling and dialogue reply execution.

Images