A multi-agent emotional intervention system and method based on emotion focusing

Abstract

This application discloses a multi-agent emotion intervention system and method based on emotion focus, relating to the field of artificial intelligence application technology. The system includes an active emotion perception module, a user profile database, and a multi-agent intervention module. The multi-agent intervention module comprises multiple agent units. The active emotion perception module generates structured emotion quadruples based on multimodal data published by the user and stores these structured emotion quadruples in the user profile database. The multi-agent intervention module loads the target user's structured emotion quadruples from the user profile database and, according to a preset emotion-focused therapy process logic, controls multiple cooperating agent units to engage in dialogue with the target user. Each agent unit is responsible for a specific stage of the dialogue. This application can improve the reliability of emotion intervention.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence application technology, and in particular to a multi-agent emotion intervention system and method based on emotion focus. Background Technology

[0002] With the widespread use of the internet and social media, people's mental health is increasingly affected by the online environment. Frequent exposure to negative information on social media can easily trigger psychological stress, anxiety, and other problems, leading to a surge in demand for online mental health services.

[0003] In existing technologies, artificial intelligence (AI) emotion support tools based on large language models (LLMs) have shown great application potential. However, these existing technical solutions mainly suffer from the following technical problems: (1) Passive intervention mode: Most existing AI emotional support systems adopt a responsive paradigm, that is, passively waiting for users to actively send out clear help signals. This mode cannot cover potential user groups who are unwilling or unable to actively seek help due to a vague understanding of their own psychological state or due to the "stigma of illness", resulting in limited service coverage.

[0004] (2) Insufficient depth of intervention: When providing emotional support, systems based on generalized large-scale language models often respond with generalized and patterned comforting words. Due to the lack of professional psychotherapy theories as structured guidance, the system is unable to conduct in-depth personalized interventions based on the user's personal experiences and emotional patterns, and cannot reach and address the user's deep-seated core emotional problems, resulting in intervention effects that remain superficial.

[0005] (3) System architecture limitations lead to logical inconsistencies: When a system using a single, black-box large-scale language model as its core performs complex long-term, multi-stage psychological intervention tasks, it is difficult to maintain logical consistency in dialogue, professionalism of roles, and long-term memory of user states. This results in the system's inability to effectively reproduce the dynamic and structured process in real psychotherapy, affecting the reliability and professionalism of the intervention. Summary of the Invention

[0006] The purpose of this application is to provide a multi-agent emotion intervention system and method based on emotion focus, which can improve the reliability of emotion intervention.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a multi-agent emotion intervention system based on emotion focus, including an active emotion perception module, a user profile database, and a multi-agent intervention module; the multi-agent intervention module includes multiple agent units. The proactive emotion perception module is used to generate structured emotion quadruples based on the multimodal data published by the user, and store the structured emotion quadruples in the user profile database. The multi-agent intervention module is used to load the target user's structured emotional quadruple from the user profile database, and control multiple cooperative agent units to conduct dialogue with the target user according to the preset emotion-focused therapy process logic; each agent unit is responsible for a specific stage of the dialogue.

[0008] Optionally, the multi-agent intervention module includes five agent units, namely a coordinator, an empathy connector, an emotion explorer, an emotion converter, and a narrative integrator. The coordinator is used to load the target user's structured emotional quadruple from the user profile database and activate the empathy connector; The empathic connector is used to engage in dialogue with the target user after being activated, and to determine whether the healing alliance has been successfully established after each round of empathic dialogue. If the healing alliance is successfully established, a first completion signal is sent to the coordinating commander; otherwise, the next round of dialogue with the target user continues. The coordinating commander is also used to activate the emotion explorer and send the structured emotion quadruple to the emotion explorer upon receiving the first completion signal. The emotion explorer, after being activated, uses the structured emotion quadruple and a built-in emotion-focused therapy questioning strategy to engage in dialogue with the target user, extract the target user's emotion markers, and when it is determined that the emotion markers are recognized by the target user, records the emotion markers in the emotion exploration record sheet, and sends a second completion signal and the emotion exploration record sheet to the coordinator. The coordinating commander is also used to activate the emotion converter and send the emotion exploration record form to the emotion converter upon receiving the second completion signal. The emotion conversion therapist, after being activated, engages in dialogue with the target user and performs an intervention task of emotion-focused therapy based on the emotional markers in the emotion exploration record table. During the execution of the intervention task, the extracted emotion conversion markers of the target user are recorded in the emotion conversion record table, and the emotional conversion of the target user is detected based on the emotion conversion markers. If the emotional conversion is achieved, a third completion signal is sent to the coordinator and the emotion conversion record table is transmitted; otherwise, the intervention task of emotion-focused therapy continues to be executed. The coordinating commander is also used to activate the narrative integrator and send the emotion transformation record form to the narrative integrator upon receiving the third completion signal. The narrative integrator, once activated, engages in dialogue with the target user based on the emotion transformation record form. When it is determined that the target user has completed the construction of a new personal narrative, the new personal narrative is sent as a final report to the coordinating commander.

[0009] Optionally, the emotional markers include four categories: conflict and division markers, unresolved complex markers, problem reaction points, and vulnerability markers; Based on the emotion markers described in the emotion exploration record form, perform the intervention task of emotion-focused therapy, specifically including: When the emotion marker is a conflict / split marker, the intervention task is to conduct a dialogue with the target user using two-chair dialogue technology; When the emotion marker is an incomplete complex marker, the intervention task is to conduct a dialogue with the target user using empty chair dialogue technology; When the emotional marker is a problem response point, the intervention task is to use system arousal unfolding technology to have a dialogue with the target user; When the emotional marker is a vulnerability marker, the intervention task is to engage in dialogue with the target user using empathic affirmation techniques.

[0010] Optionally, the empathic connector is obtained by fine-tuning the parameters of a large language model using a fine-tuning dataset, wherein the fine-tuning dataset is a public psychological dialogue dataset. The emotion converter is a hybrid architecture that combines a target large language model with retrieval-enhanced generation; the target large language model is a fine-tuned large language model.

[0011] Optionally, the multimodal data includes at least one of text, images, and videos.

[0012] Optionally, the active emotion perception module includes a data preprocessing unit and a data analysis unit, wherein the data analysis unit has multiple large models built-in; The data preprocessing unit is used to perform word segmentation and stop word removal on the text to obtain preprocessed text, and to extract key frame images from the video according to a preset time interval or scene change detection. The data analysis unit is used for: The semantic consistency score between the preprocessed text and the input image is calculated. If the semantic consistency score is greater than a set value, the various models are controlled to select the multimodal enhancement instruction. If the semantic consistency score is less than or equal to the set value, the various models are controlled to select the text-dominant instruction. The input image includes images from the multimodal data and keyframe images extracted from the video. The various models output one or more initial structured sentiment quadruples through the multimodal enhancement instruction or the text-dominant instruction. The structured sentiment quadruple is obtained by weighted integration of multiple initial structured sentiment quadruples output by multiple large models.

[0013] Optionally, the structured sentiment quadruple is obtained by weighted integration of multiple initial structured sentiment quadruples output by the large model, specifically including: The weighted Boda counting method is used to fuse multiple initial structured sentiment quadruples output by multiple large models to obtain the structured sentiment quadruples.

[0014] Optionally, the large model includes Qwen-VL and Deepseek-VL.

[0015] Optionally, the structured emotion quadruple is represented as {aspect, category, emotion, opinion}, where the aspect is the entity or event discussed by the user, the category is the mental health dimension corresponding to the aspect, the emotion is the emotional polarity corresponding to the aspect, and the opinion is the vocabulary used by the user to express the emotion; the mental health dimension includes occupational stress and self-awareness.

[0016] Secondly, this application provides a multi-agent emotion intervention method based on emotion focus, which is applied to the aforementioned multi-agent emotion intervention system based on emotion focus. The multi-agent emotion intervention method based on emotion focus includes: The proactive emotion perception module generates structured emotion quadruples based on the multimodal data posted by the user and stores the structured emotion quadruples in the user profile database. The multi-agent intervention module loads the target user's structured emotional quadruple from the user profile database and controls multiple cooperating intelligent agent units to engage in dialogue with the target user according to the preset emotion-focused therapy process logic.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a multi-agent emotion intervention system and method based on emotion-focused therapy. The system generates structured emotion quadruples through the proactive emotion perception module and stores them in a user profile database. The multi-agent intervention module controls multiple cooperating agent units to engage in dialogue with the target user according to a preset emotion-focused therapy (EFT) process logic. Each agent unit is responsible for a specific stage of the task, making the emotion intervention more targeted and improving its reliability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the structure of a multi-agent emotion intervention system based on emotion focusing, provided as an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] In one exemplary embodiment, this application provides a multi-agent emotion intervention system based on emotion focus, such as... Figure 1 As shown, the emotion-focused multi-agent emotion intervention system includes an active emotion perception module, a user profile database, and a multi-agent intervention module; the multi-agent intervention module includes multiple agent units.

[0023] The proactive emotion perception module is used to generate structured emotion quadruples based on the multimodal data published by the user, and store the structured emotion quadruples in the user profile database.

[0024] The multi-agent intervention module is used to load the target user's structured emotional quadruple from the user profile database, and control multiple cooperative agent units to conduct dialogue with the target user according to the preset emotion-focused therapy process logic; each agent unit is responsible for a specific stage of the dialogue.

[0025] The target users are those whose structured emotion quadruple sets meet the preset rules. Among them, the target users can be those whose emotion polarity is negative in the structured emotion quadruple sets.

[0026] The active emotion perception module and the multi-agent intervention module of this application are connected through a dynamic user profile database. The active emotion perception module actively identifies emotional signals from the user's publicly released multimodal data and generates a structured user profile. The multi-agent intervention module receives the user profile and executes a structured multi-stage emotional intervention process, which includes controlling multiple cooperative intelligent agent units to engage in dialogue with the target user according to a preset emotion-focused therapy process logic. Each intelligent agent unit is responsible for a specific stage of the task, making the emotional intervention at each stage more targeted and improving the reliability of the emotional intervention.

[0027] This application constructs a system capable of proactively sensing users' emotional states and providing structured, in-depth intervention based on professional emotion-focused therapy. By computationally and modularizing the complex psychotherapy process, breaking it down into tasks collaboratively executed by multiple clearly defined intelligent agents, it addresses the shortcomings of single models in terms of professionalism, consistency, and intervention depth, aiming to provide a proactive, professional solution with deep emotional support capabilities.

[0028] This application implements a closed-loop system from proactive emotion perception to structured intervention. The system first proactively analyzes the user's multimodal data through a front-end module to construct a user profile. Then, it initiates a back-end EFT multi-agent intervention module to conduct targeted, phased, in-depth intervention. The multi-agent intervention module reads user profile information from a user profile database as the initial context for intervention and engages in dialogue with the user through a human-computer interaction interface.

[0029] In one exemplary embodiment, the active emotion perception module is responsible for proactively analyzing the multimodal content posted by users on platforms such as social media to accurately perceive the user's emotional state in different aspects and to build a structured user profile.

[0030] The multimodal data includes at least one of text, images, and videos.

[0031] The multimodal data may also include data from the user's wearable devices (such as heart rate and sleep patterns), conversations from personal diary applications (requiring user authorization), or instant messaging tools, to build a more comprehensive and dynamic user profile.

[0032] The structured sentiment quadruple is represented as {Aspect, Category, Sentiment, Opinion}. The aspect is the entity or event discussed by the user, the category is the corresponding mental health dimension, the sentiment is the corresponding sentiment polarity, and the opinion is the vocabulary used by the user to express the sentiment. The mental health dimension includes occupational stress and self-awareness. Sentiment polarity includes positive, negative, and neutral. For example, for the user text "I've been under a lot of work pressure lately, and I feel like I'm on the verge of a breakdown every day," the structured sentiment quadruple extraction result would be: ("Work pressure", "Occupational stress", "Negative", "On the verge of a breakdown").

[0033] The proactive emotion perception module employs a multimodal aspect-level sentiment analysis (MABSA) workflow based on weighted ensemble of large models. First, multiple open-source large models undergo parameter efficient fine-tuning (PEFT) in the field of mental health. Then, the analysis results from multiple models are fused through weighted ensemble (e.g., using Borda counting) to improve the accuracy and robustness of the analysis. Finally, the module outputs structured sentiment quadruples containing {aspect, aspect category, sentiment polarity, opinion}, which are stored in a dynamic user profile database.

[0034] The large models include Qwen-VL and Deepseek-VL.

[0035] The active emotion perception module includes a data preprocessing unit and a data analysis unit, and the data analysis unit has multiple large models built in.

[0036] The data preprocessing unit is used to perform word segmentation and stop word removal on the text to obtain preprocessed text, and to extract key frame images from the video according to a preset time interval or scene change detection.

[0037] The data analysis unit is used to analyze the preprocessed data, specifically including: The semantic consistency score between the preprocessed text and the input image is calculated. If the semantic consistency score is greater than a set value, the various models are controlled to select a multimodal enhancement instruction; if the semantic consistency score is less than or equal to the set value, the various models are controlled to select a text-dominant instruction. The input image includes images from the multimodal data and keyframe images extracted from the video. Each model outputs one or more initial structured sentiment quadruples through the multimodal enhancement instruction or the text-dominant instruction. When the large model encounters an unknown entity, it obtains external knowledge as supplementary context through a search tool (ST). The various models are domain-fine-tuned large models.

[0038] The structured sentiment quadruple is obtained by weighted integration of multiple initial structured sentiment quadruples output by multiple large models. Specifically, the predicted probability distributions of multiple large models are integrated to output one or more structured sentiment quadruples.

[0039] The formula for weighted Porta count is: Scoring function: ,in, To score points, It is the total number of emotional polarity categories. It is emotional polarity In the The large model predicts the ranking in the sorted list.

[0040] Total score calculation: ,in, For emotional polarity Total score It is the total number of large models. It is the first The weights of each large model.

[0041] Final prediction: Choose the one with the highest total score. The polarity is taken as the final result.

[0042] In an exemplary embodiment, the structured sentiment quadruple is obtained by weighted integration of multiple initial structured sentiment quadruples output by multiple large models. Specifically, this includes using a weighted Boda count method to fuse multiple initial structured sentiment quadruples output by multiple large models to obtain the structured sentiment quadruple.

[0043] In one exemplary embodiment, the multi-agent intervention module receives a user profile and executes a structured, multi-stage intervention process that follows the four core processes of EFT theory (emotional acceptance, exploration, transformation, and meaning construction) to provide deep emotional support.

[0044] The multi-agent intervention module includes five cooperative agent units, namely, a coordinator, an empathy connector, an emotion explorer, an emotion converter, and a narrative integrator.

[0045] ① The coordinator, acting as the central control unit of the system, is responsible for the orchestration, status management, and dynamic scheduling of the entire intervention process. It maintains control and data interaction connections with the other four intelligent agents, responsible for sending control commands (activation / stop) and receiving status reports and data. The empathy connector, emotion explorer, emotion converter, and narrative integrator are activated sequentially under the coordinator's direction and interact with the user. More specifically, the coordinator activates and deactivates other expert agents (intelligent agent units) according to preset EFT process logic, manages conversation history and user profiles, and integrates security monitoring tools to identify high-risk discourse.

[0046] The functions of the coordinating commander are implemented as follows: 1) State Management and Planning: Internally, a state machine is maintained to record the current intervention stage. Upon receiving the completion signal and data payload from the expert agent, the system transitions to the next state according to the preset EFT process rules and decides to activate the corresponding agent unit.

[0047] 2) Memory module: manages two types of information, namely short-term conversation memory (the complete dialogue history of the current conversation, the activation state of each agent and key deliverables) and long-term user profile (emotional knowledge read from the user profile database and dynamically updated during intervention).

[0048] 3) Security Monitoring Tool: Integrates a tool that monitors conversation content in real time through keyword matching and semantic analysis to identify high-risk statements with malicious intent. Once an alert threshold is triggered, the normal process is immediately halted, and preset security protocols are executed (such as sending warnings or providing emergency assistance resources).

[0049] ② Empathy Connector Function: To establish and maintain a safe and trusting therapeutic alliance with the user in the early stages of intervention, laying the foundation for subsequent in-depth intervention. It also has a built-in dialogue filtering mechanism to ensure the empathetic quality of all system outputs.

[0050] Implementation: A high-efficiency parameter fine-tuning (PEFT) process is performed on a basic large-scale language model to deeply internalize empathy capabilities. The fine-tuning dataset is based on publicly available psychological dialogue datasets such as Soulful-Chat and PsyQA, and is refined and supplemented by psychology professionals according to the empathy principles of PEFT, ensuring that the model deeply internalizes professional empathy capabilities rather than simply following instructions.

[0051] ③ Emotional Explorer Function: Based on the EFT exploratory questioning strategy, guide users to identify, name and explore their core emotions in depth, and locate precise emotional markers and related unmet needs.

[0052] Implementation: Achieved through sophisticated role design and prompt engineering. Its system prompts systematically encode EFT's emotion recognition theory and exploratory questioning techniques (such as open-ended questions, clarifying questions, and empathic inferences), enabling it to generate professional and insightful guiding questions based on the dialogue context.

[0053] The Emotion Explorer uses a large language model based on the built-in role setting system prompts.

[0054] The prompt systematically encodes Emotion-Focused Therapy (EFT) theory and questioning techniques, such as open-ended questions, clarifying follow-up questions, empathic inference, emotion differentiation, and having the user retell their emotional experience. An example of a prompt might be: "You are a professional EFT therapist. Please help the user identify their current emotion in detail, encouraging them to use specific words and explore the causes, changes, and effects of the emotion from multiple perspectives. Based on the user's expression below, please analyze the polarity of their emotion and the underlying unmet needs step by step, asking 1-2 guiding questions in each round." After the user answers the guiding questions, the emotion explorer organizes and records the user's emotional markers and confirms with the user through questions whether such an emotional issue exists.

[0055] In each round, the system dynamically generates a customized prompt for the emotion explorer based on the dialogue context and the structured four-tuple content, until the system recognizes that the user "confirms" the emotional markers found by the emotion explorer.

[0056] ④ Emotional Transformation Specialist Function: To implement core intervention techniques of EFT (such as the "empty chair" dialogue), guide users to have a deep emotional experience and transformation, and promote the occurrence of corrective emotional experience.

[0057] Implementation: A hybrid technical architecture combining model fine-tuning and Retrieval Enhancement Generation (RAG) is employed. Fine-tuning enables the model to master the language patterns and rhythm of the intervention; RAG retrieves relevant knowledge from external professional knowledge bases (including EFT manuals, clinical operation guidelines, case studies, etc.) to enhance the professionalism and accuracy of the intervention and reduce the risk of "illusion".

[0058] In the dialogue iteration loop, there is a conditional trigger that determines whether the thinking results of the large model in the hybrid technology architecture include the need to use core intervention techniques, such as empty chairs or double chairs. If they are needed, the link search answer of RAG retrieval is triggered and returned to the large model in the hybrid technology architecture to integrate the content of the next dialogue.

[0059] ⑤ Narrative Integrator: Purpose: In the later stages of intervention, guide users to reflect on the entire intervention process, extract new insights, and construct a coherent and empowering new personal narrative to consolidate the treatment effect.

[0060] Implementation: Achieved through highly structured prompt engineering. The system instructions are coded in detail with specific steps and language paradigms to guide users in reflecting, refining understanding, and constructing narratives, ensuring stable and consistent execution of integrated functions.

[0061] The Narrative Integrator, also known as the Large Language Model following the built-in character setting system, is specified in the Prompt as having the following main tasks: 1. Guide users to recall key emotional turning points and experiences; 2. Help users summarize and understand in their own words; 3. Assist users in reconstructing their personal narratives, such as interpreting the meaning of life events.

[0062] Template Example: "Please remind the user to reflect on their emotional transformation process and review in detail which changes touched them the most, and try to summarize it using phrases like 'My new understanding' or 'My current hope is...'. ... Gradually guide the user until they can fluently and completely express their new life narrative and give a positive evaluation of their own changes." The prompt content embeds detailed structure and step-by-step prompts to ensure stable steps and consistent output.

[0063] In one exemplary embodiment, the multi-agent intervention module executes a structured, phased deep emotional intervention process. This process is centrally coordinated by a coordinator and can be broken down into the following four sub-steps (phases): 1. Emotional Acceptance and Alliance Building Phase The coordinator is used to load the target user's structured emotional quadruple from the user profile database and activate the empathy connector.

[0064] The empathic connector is used to engage in dialogue with the target user after being activated, and to determine whether the therapeutic alliance has been successfully established after each round of empathic dialogue. If the therapeutic alliance is successfully established, it sends a first completion signal to the coordinating commander and transmits the initially collected emotional cues (structured emotional quadruples); otherwise, it continues to engage in the next round of dialogue with the target user.

[0065] In this phase, the empathic connector establishes a trusting relationship (therapeutic alliance) with the user through multiple rounds of empathic dialogue. Establishing the therapeutic alliance involves analyzing the user's responses to determine if the user exhibits signs of trust, openness, or emotional calm. When the trust level reaches a threshold, the therapeutic alliance is considered successfully established. The trust level is determined based on scoring the user's responses.

[0066] 2. Emotional Exploration Stage The coordinating commander is also used to activate the emotion explorer upon receiving the first completion signal and send the structured emotion quadruple to the emotion explorer.

[0067] The emotion explorer, upon activation, engages in dialogue with the target user based on the structured emotion quadruple and employs a built-in emotion-focused therapy questioning strategy. The explorer extracts the target user's emotional markers and, upon determining that the target user acknowledges these markers, records them in an emotion exploration record sheet. A second completion signal and the record sheet are then sent to the coordinator. Using natural language understanding technology, the emotion explorer extracts and confirms core emotional vocabulary and associated unmet needs from the user's statements. Once the user acknowledges these identifications, a clear emotional marker is identified.

[0068] 3. Emotional Transition Stage The coordinating commander is also used to activate the emotion converter and send the emotion exploration record form to the emotion converter upon receiving the second completion signal.

[0069] The emotion conversion therapist, upon activation, engages in dialogue with the target user and executes an emotion-focused therapy intervention task based on the emotional markers in the emotion exploration record table. During the intervention, the extracted emotion conversion markers of the target user are recorded in an emotion conversion record table. The therapist then checks whether the target user's emotion conversion has been achieved based on these markers. If so, a third completion signal is sent to the coordinator, and the emotion conversion record table is transmitted. Otherwise, the emotion-focused therapy intervention task continues. During the execution of the emotion-focused therapy intervention task, a professional solution is retrieved from an external knowledge base using the RAG mechanism to guide the generation of interventions, ensuring the professionalism of the intervention.

[0070] Emotional transformation markers refer to the cognitive shifts and deeper expressions that users exhibit during emotional intervention. These markers are specifically determined through a large-scale model used by the emotional transformation specialist. The specialist analyzes the user's responses after the intervention dialogue to detect changes in the positivity of their emotional expression, the emergence of cognitive restructuring, or an increase in self-efficacy. When the changes in these indicators (emotional transformation markers) exceed preset thresholds, the emotional transformation goal is considered achieved. These changes are extracted in real-time from the large-scale model and the JSON data of the user's dialogue.

[0071] The RAG mechanism refers to the use of contextual information in a large model to generate dialogue content whenever a psychological intervention is performed on a user. This information is obtained by retrieving relevant cases, related theoretical terms, core intervention methods, and other materials from a database.

[0072] 4. Meaning Construction and Integration Stage The coordinating commander is also used to activate the narrative integrator and send the emotion transformation record form to the narrative integrator upon receiving the third completion signal.

[0073] The narrative integrator, once activated, engages in dialogue with the target user based on the emotion transformation record form. When it is determined that the target user has completed the construction of a new personal narrative (meaning construction), the new personal narrative is sent as a final report to the coordinating commander.

[0074] Among them, the narrative integrator guides users to reflect on their experience during the transformation process, extract new insights, and help users build an empowering new personal narrative.

[0075] The narrative integrator determines whether meaning construction is complete. Meaning construction is considered complete when the user can clearly and coherently retell their new personal narrative and express positive self-evaluation. Upon receiving the final report, the coordinator comprehensively assesses the user's overall condition and decides whether to end the intervention session or identify new therapeutic focuses and return to the emotion exploration phase, initiating a new intervention cycle.

[0076] Determining that the new personal narrative of the target user has been constructed includes: obtaining the structured emotional quadruple of the new personal narrative; comparing the structured emotional quadruple of the target user loaded from the user profile database with the structured emotional quadruple of the new personal narrative; if the emotional polarity changes to positive, then it is determined that the new personal narrative of the target user has been constructed.

[0077] In one exemplary embodiment, the emotion markers include four categories: conflict / split markers, unresolved complex markers, problem reaction points, and vulnerability markers.

[0078] When a visitor (user) is identified to have self-criticism or compulsive conflict, the emotional marker is identified as a conflict-split marker.

[0079] When a visitor is identified to have unexpressed, persistent emotions toward significant others, the emotional marker is identified as an unresolved complex marker.

[0080] When a visitor is identified as confused about their own emotional responses, emotional markers are identified as problem response points.

[0081] When a visitor reveals core feelings of shame or helplessness, the emotional marker is identified as a vulnerability marker.

[0082] Based on the emotion markers described in the emotion exploration record form, perform the intervention task of emotion-focused therapy, specifically including: When the emotion marker is a conflict / split marker, the intervention task is to conduct a dialogue with the target user using two-chair dialogue technology; When the emotion marker is an incomplete complex marker, the intervention task is to conduct a dialogue with the target user using empty chair dialogue technology; When the emotional marker is a problem response point, the intervention task is to use system arousal unfolding technology to have a dialogue with the target user; When the emotional marker is a vulnerability marker, the intervention task is to engage in dialogue with the target user using empathic affirmation techniques.

[0083] In one exemplary embodiment, the empathic connector is obtained by fine-tuning the parameters of a large language model using a fine-tuning dataset, wherein the fine-tuning dataset is a public psychological dialogue dataset. The emotion converter described here employs a hybrid architecture combining a target large language model with retrieval-enhanced generation; the target large language model is a fine-tuned version of the large language model. This fine-tuned large language model is trained using publicly available dialogue datasets in the field of mental health, such as Soulful-Chat and PsyQA, which contain a large amount of authentic question-and-answer and emotion intervention texts from clients and therapists.

[0084] In one exemplary embodiment, emotion-focused therapy can also be replaced by cognitive behavioral therapy (CBT) or acceptance and commitment therapy (ACT). In this alternative, the roles and tasks of the agents need to be adapted to the new theory (e.g., CBT may include agents such as "cognitive reconstructors" and "behavioral experimenters"), but the multi-agent collaboration model of "decomposition-coordination-execution" remains unchanged.

[0085] The technical effects of this application are as follows: 1. This invention realizes a paradigm shift from "passive response" to "proactive care", expanding the service coverage.

[0086] Reason / Technical Point: This advantage stems from the proactive emotion perception module in the technical solution. Most existing technologies passively wait for users to seek help, while this invention proactively analyzes users' publicly available multimodal data through this module, identifying users with potential psychological distress who have not actively sought help, and initiating timely intervention. This solves the problems of passive and narrow-coverage service models in existing technologies.

[0087] 2. This invention significantly enhances the professionalism and depth of intervention, enabling it to address deep-seated core emotions.

[0088] Reasons / Technical Points: This advantage stems from a multi-agent collaborative framework based on EFT theory. Existing technologies rely on a single, generalized, large model, resulting in generalized and superficial responses. This invention decomposes the complex EFT process into tasks for five specialized agents (empathy, exploration, transformation, integration, and coordination), with each agent responsible for a specific stage of the specialized task. In particular, the RAG combined with fine-tuning technique used by the emotion transformation specialist ensures that when performing core techniques such as the "empty chair" exercise, it can operate based on a professional external knowledge base, thereby guaranteeing the theoretical fidelity and clinical effectiveness of the intervention and solving the problem of insufficient intervention depth in existing technologies.

[0089] 3. This invention ensures the logical consistency and high reliability of long-term, multi-stage intervention processes.

[0090] Reasons / Technical Points: This advantage stems from a multi-agent architecture centered on a coordinator. A single LLM (Leveled Model) is prone to "amnesia" or logical drift in long dialogues. This invention, through deterministic state management and process scheduling by the coordinator, ensures that intervention is strictly carried out in an orderly manner according to the four stages of EFT (Extra-Frequency Exploration). Each agent has a clear division of labor, and information is transmitted through structured record tables (such as the "Emotional Exploration Record Table"), avoiding information loss and logical confusion, and solving the problem of poor consistency in complex tasks caused by a single-model architecture.

[0091] 4. This invention improves the efficiency and accuracy of intervention.

[0092] Reason / Technical Point: This advantage stems from the accurate user profiles provided by the proactive emotion perception module. Through refined MABSA analysis, this module provides a structured initial context containing specific "aspects" and "emotions" for subsequent interventions. This allows intelligent agents such as emotion explorers to ask more targeted questions and provide guidance from the outset, reducing aimless exploration in the early stages, shortening the number of dialogue rounds to pinpoint the core issues, and thus improving overall intervention efficiency.

[0093] Based on the same inventive concept, this application also provides a method for implementing the aforementioned emotion-focused multi-agent emotional intervention method. The solution provided by this method is similar to the implementation described above. Therefore, the specific limitations of one or more embodiments of the emotion-focused multi-agent emotional intervention method provided below can be found in the above-described limitations of an emotion-focused multi-agent emotional intervention system, and will not be repeated here.

[0094] In one exemplary embodiment, an emotion-focused multi-agent emotional intervention method is provided, which is applied to the emotion-focused multi-agent emotional intervention system. The emotion-focused multi-agent emotional intervention method includes: Step 101: The proactive emotion perception module generates structured emotion quadruples based on the multimodal data published by the user, and stores the structured emotion quadruples in the user profile database.

[0095] Step 102: The multi-agent intervention module loads the target user's structured emotional quadruple from the user profile database, and controls multiple cooperating intelligent agent units to engage in dialogue with the target user according to the preset emotion-focused therapy process logic.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-agent sentiment intervention system based on emotion focusing, characterized in that, The emotion-focused multi-agent emotion intervention system includes an active emotion perception module, a user profile database, and a multi-agent intervention module. The multi-agent intervention module includes multiple agent units; The proactive emotion perception module is used to generate structured emotion quadruples based on the multimodal data published by the user, and store the structured emotion quadruples in the user profile database. The multi-agent intervention module is used to load the target user's structured emotional quadruple from the user profile database, and control multiple cooperative agent units to conduct dialogue with the target user according to the preset emotion-focused therapy process logic; each agent unit is responsible for a specific stage of the dialogue.

2. The emotion-focused based multi-agent sentiment intervention system of claim 1, wherein, The multi-agent intervention module includes five agent units, namely, a coordinator, an empathy connector, an emotion explorer, an emotion converter, and a narrative integrator. The coordinator is used to load the target user's structured emotional quadruple from the user profile database and activate the empathy connector; The empathic connector is used to engage in dialogue with the target user after being activated, and to determine whether the healing alliance has been successfully established after each round of empathic dialogue. If the healing alliance is successfully established, a first completion signal is sent to the coordinating commander; otherwise, the next round of dialogue with the target user continues. The coordinating commander is also used to activate the emotion explorer and send the structured emotion quadruple to the emotion explorer upon receiving the first completion signal. The emotion explorer, after being activated, uses the structured emotion quadruple and a built-in emotion-focused therapy questioning strategy to engage in dialogue with the target user, extract the target user's emotion markers, and when it is determined that the emotion markers are recognized by the target user, records the emotion markers in the emotion exploration record sheet, and sends a second completion signal and the emotion exploration record sheet to the coordinator. The coordinating commander is also used to activate the emotion converter and send the emotion exploration record form to the emotion converter upon receiving the second completion signal. The emotion conversion therapist, after being activated, engages in dialogue with the target user and performs an intervention task of emotion-focused therapy based on the emotional markers in the emotion exploration record table. During the execution of the intervention task, the extracted emotion conversion markers of the target user are recorded in the emotion conversion record table, and the emotional conversion of the target user is detected based on the emotion conversion markers. If the emotional conversion is achieved, a third completion signal is sent to the coordinator and the emotion conversion record table is transmitted; otherwise, the intervention task of emotion-focused therapy continues to be executed. The coordinating commander is also used to activate the narrative integrator and send the emotion transformation record form to the narrative integrator upon receiving the third completion signal. The narrative integrator, once activated, engages in dialogue with the target user based on the emotion transformation record form. When it is determined that the target user has completed the construction of a new personal narrative, the new personal narrative is sent as a final report to the coordinating commander.

3. The multi-agent emotion intervention system based on emotion focusing according to claim 2, characterized in that, The emotional markers include four categories: conflict and division markers, unresolved complex markers, problem reaction points, and vulnerability markers. Based on the emotion markers described in the emotion exploration record form, perform the intervention task of emotion-focused therapy, specifically including: When the emotion marker is a conflict / split marker, the intervention task is to conduct a dialogue with the target user using two-chair dialogue technology; When the emotion marker is an incomplete complex marker, the intervention task is to conduct a dialogue with the target user using empty chair dialogue technology; When the emotional marker is a problem response point, the intervention task is to use system arousal unfolding technology to have a dialogue with the target user; When the emotional marker is a vulnerability marker, the intervention task is to engage in dialogue with the target user using empathic affirmation techniques.

4. The multi-agent emotion intervention system based on emotion focusing according to claim 2, characterized in that, The empathic connector was obtained by fine-tuning the parameters of a large language model using a fine-tuning dataset, which is a public psychological dialogue dataset. The emotion converter is a hybrid architecture that combines a target large language model with retrieval-enhanced generation; the target large language model is a fine-tuned large language model.

5. The multi-agent emotion intervention system based on emotion focusing according to claim 1, characterized in that, The multimodal data includes at least one of text, images, and videos.

6. The multi-agent emotion intervention system based on emotion focusing according to claim 5, characterized in that, The active emotion perception module includes a data preprocessing unit and a data analysis unit, and the data analysis unit has multiple large models built in. The data preprocessing unit is used to perform word segmentation and stop word removal on the text to obtain preprocessed text, and to extract key frame images from the video according to a preset time interval or scene change detection. The data analysis unit is used for: The semantic consistency score between the preprocessed text and the input image is calculated. If the semantic consistency score is greater than a set value, the various models are controlled to select the multimodal enhancement instruction. If the semantic consistency score is less than or equal to the set value, the various models are controlled to select the text-dominant instruction. The input image includes images from the multimodal data and keyframe images extracted from the video. The various models output one or more initial structured sentiment quadruples through the multimodal enhancement instruction or the text-dominant instruction. The structured sentiment quadruple is obtained by weighted integration of multiple initial structured sentiment quadruples output by multiple large models.

7. The multi-agent emotion intervention system based on emotion focusing according to claim 6, characterized in that, The structured sentiment quadruple is obtained by weighted integration of multiple initial structured sentiment quadruples output by the large model, specifically including: The weighted Boda counting method is used to fuse multiple initial structured sentiment quadruples output by multiple large models to obtain the structured sentiment quadruples.

8. The multi-agent emotion intervention system based on emotion focusing according to claim 6, characterized in that, The large models include Qwen-VL and Deepseek-VL.

9. The multi-agent emotion intervention system based on emotion focusing according to claim 1, characterized in that, The structured emotion quadruple is represented as {aspect, category, emotion, opinion}, where the aspect is the entity or event discussed by the user, the category is the mental health dimension corresponding to the aspect, the emotion is the emotional polarity corresponding to the aspect, and the opinion is the vocabulary used by the user to express the emotion; the mental health dimension includes occupational stress and self-awareness.

10. A multi-agent emotion intervention method based on emotion focusing, characterized in that, The emotion-focused multi-agent emotional intervention method is applied to the emotion-focused multi-agent emotional intervention system according to any one of claims 1-9, and the emotion-focused multi-agent emotional intervention method includes: The proactive emotion perception module generates structured emotion quadruples based on the multimodal data posted by the user and stores the structured emotion quadruples in the user profile database. The multi-agent intervention module loads the target user's structured emotional quadruple from the user profile database and controls multiple cooperating intelligent agent units to engage in dialogue with the target user according to the preset emotion-focused therapy process logic.

Images