Method for classifying psychological stigma based on multi-agent self-correction and thought chain distillation

Abstract

The application discloses a psychological stigma classification method based on multi-agent self-correction and thought chain distillation, relates to the technical field of artificial intelligence natural language processing, and generates a generator to output a preliminary classification result and reasoning principle according to interview content; an evaluator carries out fine-grained quality evaluation in combination with a multi-dimensional scoring scale and a major error check and generates a modification suggestion; a fine tuner modifies the classification result and reasoning principle based on the feedback of the evaluator and in combination with the interview content until a preset exit condition is met to obtain a final classification result and reasoning principle; a compressor carries out token constraint compression on the reasoning principle, splices the interview content and the simplified reasoning principle as a supervision signal, and trains a student model; and finally, a student model is used to realize psychological stigma category prediction of target interview content. The application adopts a multi-agent system to overcome the unreliable single-round reasoning problem of a large language model, and combines token compression and knowledge distillation to migrate the complex multi-agent system capability to a lightweight student model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence natural language processing technology, and in particular to a psychological stigma classification method based on multi-agent self-correction and thought chain distillation. Background Technology

[0002] Mental health stigma defines mental illness as an abnormal state that deviates from social norms, treating affected individuals as a stigmatized group. This includes negative stereotypes and discriminatory behaviors, and is a core factor hindering mental illness patients from receiving treatment and achieving recovery. With social communication increasingly extending to online communities, peer communication platforms, and human-computer interaction scenarios, accurate analysis of stigmatizing language sensitivity is crucial for various social computing system applications such as public opinion monitoring, needs triage, and psychological support.

[0003] The stigma surrounding mental health is multifaceted and complex. Psychological and sociological theories can effectively distinguish stigmatizing expressions from ordinary negative speech. Corrigan et al.'s social conceptual framework breaks down stigma into cognitive judgment, emotional response, and behavioral response. Based on attribution models, Meng et al. introduced MHStigmaInterview, a deeply annotated expert interview corpus, providing a benchmark dataset for the identification, detection, and fine-grained classification of mental health stigma. This dataset categorizes non-stigmatizing categories and, combined with attribution models, identifies seven explicit categories: Responsibility, Social Distance, Anger, Helping, Pity, Coercive Segregation, and Fear.

[0004] While neural network classifiers have made significant progress in mental health research, their practical applications face bottlenecks due to their limited contextual understanding and inherent black-box nature. In recent years, large language models (LLMs) have been widely used in health education and mental illness diagnosis due to their powerful contextual parsing capabilities. However, most existing large language model solutions generally employ unstable single-round reasoning patterns. While the output explanations may seem reasonable, they fail to reflect the deeper logic behind the decisions, exhibiting particularly pronounced limitations in complex scenarios requiring a comprehensive judgment combining details and reason. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this application provides a psychological stigma classification method based on multi-agent self-correction and thought chain distillation. It uses a multi-agent system to overcome the unreliability of single-round reasoning in large language models, and combines token compression and knowledge distillation to transfer the capabilities of complex multi-agent systems to lightweight student models.

[0006] To achieve the above objectives, this application adopts the following technical solution, including: A psychological stigma classification method based on multi-agent self-correction and thought chain distillation includes the following steps: S1, Obtain the interview dataset; the interview content consists of a sequence of utterances, and the interview label is the probability of the psychological stigma category corresponding to the interview content. S2, construct the StigReDi architecture, which is a comprehensive model architecture that integrates multi-agent self-correction and thought chain distillation, as shown below: In the protocol-guided candidate generation module, the generator produces a preliminary output based on the interview content, which includes the psychological stigma classification results and the corresponding reasoning principles. In the scale-based quality assessment module, the evaluator assesses the quality of the psychological stigma classification results and the corresponding reasoning principles based on the rating scale and major error checking, and generates modification suggestions. In the conditional feedback fine-tuning iteration module, the fine-tuner modifies the psychological stigma classification results and corresponding reasoning principles based on the feedback from the evaluator and in conjunction with the interview content. The modified psychological stigma classification results and corresponding reasoning principles are then evaluated by the evaluator and modified by the fine-tuner. The fine-tuning iteration process continues until a predetermined exit condition is met, resulting in the final psychological stigma classification results and corresponding reasoning principles. In the token-restricted knowledge distillation module, the compressor compresses the final reasoning principle to keep its total number of tokens within a preset threshold; the compressed, simplified reasoning principle and the interview content are used together as a supervision signal to train the student model, enabling the student model to predict the final psychological stigma classification result. S3 uses the trained student model to predict the psychological stigma category of the target interview content.

[0007] Preferably, the generator follows a preset six-step standardized protocol to progressively reason about the interview content, sequentially completing the abstraction of the interview theme, extraction of relevant evidence, verification of category rules, resolution of conflicting evidence, and soundness check of reasoning logic, and outputting preliminary psychological stigma classification results and corresponding reasoning principles.

[0008] Preferably, the evaluator performs a multi-dimensional, fine-grained quality assessment of the psychological stigma classification results and the corresponding reasoning principles, and generates modification suggestions, as follows: ; in, and and represent the mental stigma classification result and the corresponding reasoning principle input to the evaluator in the t-th fine-tuning iteration, respectively. When t=1, the evaluator input is the preliminary mental stigma classification result and the corresponding reasoning principle generated by the generator. This indicates the content of the interview; Represents the evaluator; Used to indicate whether a major error exists; This indicates the scoring across multiple evaluation dimensions; Descriptions indicating major and general errors; This indicates a suggestion for modification.

[0009] Preferably, the major errors include three categories: fabricated evidence, insufficient support, and violation of rules; the evaluation dimensions include seven dimensions: procedural compliance, fidelity of evidence, rule compliance, category completeness, logical consistency, semantic precision, and explanatory efficiency.

[0010] Preferably, the modifications to the psychological stigma classification results and corresponding reasoning principles during the fine-tuning process are guided by the principles of complete citation, minimal modification, and label consistency.

[0011] Preferably, the exit conditions for fine-tuning iterations include: There are no major errors, and the evaluator's total score across all evaluation dimensions exceeds the set upper limit threshold τ for the total evaluation score; In consecutive N err During the next fine-tuning iteration, several major errors occurred. The improvement in the total assessment score that can be achieved after fine-tuning remains at a marginal level within M consecutive fine-tuning iterations; The number of fine-tuning iterations exceeds the set upper limit threshold N. iter .

[0012] Preferably, the compressor compresses the final reasoning principle: ; in, This represents the reasoning principle that is finally output after fine-tuning and iteration. This represents the final output of the psychological stigma classification result after fine-tuning and iteration; Indicates compressor; This represents the simplified reasoning principle after compression. RoBERTa-base was used as the backbone network of the student model; interview content was concatenated and fused with compressed and simplified reasoning principles as input to the student model. ; in, This indicates the content of the interview; Indicates splicing; Represents the input to the student model; introduces dedicated special identifiers. As a separator / connector segment.

[0013] Preferably, the student model uses cross-entropy loss. Training is performed to approximate the final output of the psychological stigma classification result after fine-tuning iterations, using cross-entropy loss. As shown below: ; in, This represents the number of samples in the training set. This represents the final output of the psychological stigma classification result after fine-tuning iterations for the i-th training sample. For indicator functions, if ,but ,otherwise ; This represents the k-th psychological stigma category; The input to the student model is composed of a simplified reasoning principle and interview content from the i-th training sample. The student model input is Time prediction category is The posterior probability.

[0014] This application also provides an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned psychological stigma classification method based on multi-agent self-correction and thought chain distillation.

[0015] This application also provides a computer program product comprising a computer program / instruction that, when executed by a processor, implements the aforementioned psychological stigma classification method based on multi-agent self-correction and thought chain distillation.

[0016] The advantages of this application are: (1) This application proposes the StigReDi architecture, a comprehensive model architecture that integrates multi-agent self-correction and thought chain distillation, which can achieve interpretable fine-grained classification of mental health stigma. The framework consists of two main parts operating in tandem: first, a multi-agent system (generator + evaluator + fine tuner), equipped with a multi-agent architecture with a built-in self-fine-tuning review mechanism, which is specifically designed to overcome the unreliability of single-round reasoning in large language models and significantly improve the accuracy of stigma discrimination; second, a principle-enhanced student model, which first uses a compressor to compress the reasoning principle output by the multi-agent system to keep the total number of tokens within a preset threshold, and then relies on the highly reliable reasoning basis generated by the preceding multi-agent system as an enhanced supervision signal to carry out knowledge distillation training, thereby comprehensively improving the context representation ability and overall prediction performance of the lightweight model.

[0017] (2) This application adopts multi-agent collaboration, with the generator, evaluator and fine tuner having different roles and closed-loop iterative optimization, which ensures the dual quality of classification results and reasoning principles.

[0018] (3) The fine-tuning process follows three principles: complete citation, minimal modification, and label consistency, ensuring that the revision does not deviate from the original interview semantics, does not deliberately tamper with evidence, and maintains the stability of stigma labels, while taking into account the rationality of the revision and the authenticity of the content.

[0019] (4) Four types of iteration termination conditions are set to adapt to different sample reasoning difficulties, which avoids the waste of computing power by invalid loops and ensures that high-difficulty samples are fully corrected, thus balancing accuracy and efficiency.

[0020] (5) This application achieves lightweight deployment. By combining token compression and knowledge distillation, the reasoning principle is compressed with token-limited compression through a compressor. The text length is controlled while retaining the core reasoning logic, which solves the problems of excessively long reasoning text and high model reasoning overhead, and adapts to the requirements of lightweight deployment. RoBERTa-base is selected as the backbone network of the student model. Interview text + simplified reasoning principle splicing is used as the supervision signal. A special delimiter is introduced to fuse the input. The cross-entropy loss function is used to accurately fit the optimal classification result, and the complex reasoning ability of multi-agent is distilled into the lightweight student model, taking into account both reasoning interpretability and deployment efficiency.

[0021] (6) This application adopts strong rule constraints, relies on codebooks, scale scoring, and error checking to reduce large model illusions and subjective biases.

[0022] (7) This application is highly interpretable, adopts the structured reasoning principle of the whole process, and the evidence is traceable and the logic is retrievable. Attached Figure Description

[0023] Figure 1This is a flowchart of the psychological stigma classification method based on multi-agent self-correction and thought chain distillation of the present invention.

[0024] Figure 2 This is a schematic diagram of the StigReDi architecture of the present invention. Detailed Implementation

[0025] 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.

[0026] like Figure 1 As shown, the psychological stigma classification method based on multi-agent self-correction and thought chain distillation includes the following steps: S1, Obtain the interview dataset.

[0027] , Indicates by A dataset consisting of interviews, in which... This represents the content of the i-th interview. , It consists of a sequence of m sentences exchanged between the chatbot and the participant in the i-th interview. This represents the m-th sentence exchanged between the chatbot and the participant in the i-th interview. This represents the psychological stigma category label (probability label) corresponding to the i-th interview. , This represents the probability that the i-th interview corresponds to the k-th psychological stigma category. There are a total of 8 psychological stigma categories, including non-stigmatizing, responsibility, social distance, anger, helping, pity, coercive segregation, and fear.

[0028] The definitions of the eight categories of psychological stigma are as follows: Non-stigmatization: A positive attitude towards mental illness that is inclusive, objective, and unbiased; Responsibility: It is believed that individuals should bear primary fault and responsibility for their own psychological problems; Social distancing: deliberately avoiding and distancing oneself from people with mental illnesses, and reducing the willingness to engage in social contact; Anger: Negative emotions that evoke blame, resentment, and hostility towards individuals with mental illness; Assistance: We are willing to provide care, support, and practical help to those suffering from mental health issues; Compassion: Viewing people with mental illness as vulnerable and feeling sympathy and pity; Forced isolation: Advocating for the isolation, restraint, and mandatory control of people with mental disorders; Fear: Influenced by stereotypes, individuals with mental illnesses develop fear and a sense of danger.

[0029] S2, constructing the StigReDi architecture, which is a comprehensive model architecture that integrates multi-agent self-correction and thought chain distillation.

[0030] like Figure 2 As shown, the StigReDi architecture is as follows: First, in the Protocol-Guided Candidate Generation module, the generator follows a predefined six-step protocol to produce an initial output based on the given interview content, including preliminary psychological stigma classification results and corresponding structured reasoning principles.

[0031] Subsequently, in the Rubric-Based Quality Evaluation module, the evaluator reviews the initial output based on the rating scale and major error check, and generates modification suggestions.

[0032] Next, in the Feedback-Conditioned Iterative Refinement module, the refiner modifies the psychological stigma classification results and corresponding reasoning principles based on the evaluator's feedback and the interview content. The modified psychological stigma classification results and corresponding reasoning principles are then evaluated by the evaluator and modified by the refiner. This iterative refinement process continues until a predetermined exit condition is met, resulting in the final output of the psychological stigma classification results and corresponding reasoning principles after the iterative refinement.

[0033] Finally, in the token-bound knowledge distillation module, a compressor is applied to the final output reasoning principle after fine-tuning iterations to keep its total number of tokens within a preset threshold. The compressed, simplified reasoning principle, along with the original interview content, serves as a supervisory signal to train the student model, enabling it to predict the final classification results derived from the multi-agent system (generator + evaluator + fine-tuner). The multi-agent system, consisting of the generator, evaluator, and fine-tuner, functions as the teacher model.

[0034] The detailed descriptions of each module are as follows.

[0035] 1. Protocol-guided candidate generation module This module introduces a generator to predict psychological stigma categories through a step-by-step reasoning process: ; in, and These represent the preliminary results of psychological stigma classification and the corresponding reasoning principles, respectively. This represents the content of an interview between a chatbot and a respondent; This represents a generator.

[0036] The generator follows a pre-defined six-step standardized protocol to progressively reason through the input interview text, successively completing the abstraction of the interview theme, extraction of relevant evidence, verification of category rules, resolution of conflicting evidence, and check of the soundness of reasoning logic, and outputting preliminary psychological stigma classification results and corresponding reasoning principles.

[0037] Specifically, the generator's progressive reasoning process is as follows: (1) The generator performs step-back abstraction to summarize the main points of the interview and extract the interviewee's position globally. (2) The generator extracts fragments from the interview as evidence, and each fragment corresponds to a certain psychological stigma category. (3) The category checking step screens and verifies the evidence and categories extracted in the previous step according to the rules of the codebook. The original intention of this design is to reduce the false illusions caused by the accumulation of inductive bias in the pre-training stage of the Big Prophecy Model (LLM). (4) Since an interview may have multiple different rational evidences leading to different label categories, a conflict resolution principle is defined to include measuring the confidence of evidence and social impact to select the most dominant psychological stigma category. (5) In the soundness check operation, its logical chain will be reviewed and checked to eliminate the situation of impaired reasoning consistency or disjointed evidence citations as much as possible. (6) The generator gives the final prediction result of the psychological stigma category, that is, the preliminary classification result. The above sequential steps are then encapsulated into a well-structured and clear reasoning principle. .

[0038] The progressive reasoning design emulates the deductive reasoning path of human experts who break down puzzles, and by weaving classification operations into a protocol-driven structured workflow, it solidifies the reliability and interpretability of decision-making arguments.

[0039] 2. Scale-based quality assessment module Large language models often struggle to produce satisfactory results in initial evaluations. In this task context, although the generator can predict mental stigma categories and provide explanations, its single-generation mechanism makes it prone to producing seemingly reasonable but ultimately unreliable results. To mitigate this issue, a self-fine-tuning mechanism is employed to iteratively optimize the generator's output, thereby enhancing the reliability of predictions.

[0040] This application develops an evaluator to detect the presence of major errors and perform multi-dimensional fine-grained quality assessment, the execution process of which is as follows: ; in, and Let represent the mental stigma classification result and the corresponding reasoning principle input in the t-th fine-tuning iteration, respectively. When t=1, the evaluator input is the preliminary mental stigma classification result and the corresponding reasoning principle generated by the generator; the evaluator output is... Used to indicate whether a major error exists; This indicates the scoring across multiple evaluation dimensions; Provide descriptive tags for critical and general errors; Provide suggestions for modification; This represents the evaluator.

[0041] This embodiment summarizes three types of major errors: the evidence cited in the reasoning principle did not appear in the actual interview (fabricated evidence); the reasoning principle lacks sufficient evidence to support the selected psychological stigma category (insufficient support); and the codebook rules were not followed when reasoning about the category (violation of rules). Once any of the above errors occur, a major error exists, indicating that the current classification result needs further optimization to improve its judgment.

[0042] In addition, another key factor in determining whether further optimization is needed is the comprehensive score for the following evaluation dimensions: procedural adherence, grounding fidelity, rule conformity, categorical completeness, logical consistency, semantic precision, and explanatory efficiency. For each evaluation dimension, the evaluator is prompted to assign a score from 0 to 4.

[0043] The definitions of the seven evaluation dimensions are as follows: Procedural compliance: Adhere to standardized reasoning processes and protocol steps; Fidelity of evidence: The arguments are completely anchored to the original text, with no fabrication or alteration; Rule compliance: Strictly adhere to domain codebooks and expert judgment rules; Category completeness: Fully cover candidate categories to avoid omissions; Logical consistency: The reasoning chain is self-consistent and coherent, and the arguments match the conclusion; Semantic precision: Accurately understand the semantics of text, reducing misinterpretation and misjudgment; Explanation efficiency: The explanations are concise and efficient, eliminating redundant expressions.

[0044] At the same time, the evaluator also needs to provide descriptive error labels and targeted improvement suggestions to assist in subsequent fine-tuning steps.

[0045] 3. Fine-tuning iterative module with conditional feedback After completing the quality assessment, the fine-tuner will perform targeted modifications based on the feedback provided by the evaluator. The modification process is formalized as follows: ; ; in, Indicates the fine-tuning mechanism; This represents the mental stigma classification result and corresponding reasoning principle output by the t-th fine-tuning iteration, which is also the mental stigma classification result and corresponding reasoning principle input by the (t+1)-th fine-tuning iteration. This represents the feedback from the evaluator during the t-th fine-tuning iteration, i.e., the output of the multi-dimensional fine-grained quality evaluation result.

[0046] The fine-tuning process is guided by the following three principles: First, the principle of complete citation: all citations used as evidence must consist of continuous text from the interview, and any modification or summarization is strictly prohibited. Second, the principle of minimal modification: the fine-tuner is required to perform localized, restrictive adjustments only to specific reasoning steps based on the evaluator's feedback, to avoid radical rewriting. Third, the principle of label consistency: it is explicitly stipulated that the final output of the predicted label results must maintain strong consistency with its modified overall reasoning logic.

[0047] The fine-tuning process will terminate and exit immediately when any of the following conditions are met (exit mechanism): (1) There are no major errors in the entire deduction conclusion, and the total evaluation score of the evaluator in all evaluation dimensions exceeds the set upper limit threshold τ (τ=23). (2) Major errors occur consecutively and in consecutive N events. err In the next fine-tuning iteration (N) err =3); (3) The improvement in the total evaluation score that can be achieved after fine-tuning is at a negligible marginal level (Δ<2) in the M consecutive fine-tuning iterations. Among them, the improvement in evaluation score that can be achieved in the current iteration (the t-th fine-tuning iteration) is , =∑ This represents the total evaluation score for the current iteration across all dimensions. This indicates the highest total score ever achieved up to the current iteration. This represents the improvement in the evaluation score during the current iteration; (4) The number of fine-tuning iterations performed exceeds the set upper limit threshold N for the number of iterations. iter (N) iter =5).

[0048] The initial purpose of setting the first exit condition is to ensure that the output category prediction is accurate and that the reasoning behind it is also of excellent quality; the remaining exit conditions are intended to control overall waste by intercepting and terminating the process in advance. The combination that achieves the highest and best score evaluation by integrating the performance of each loop in the entire process will be selected as the final output and result example of the entire process.

[0049] 4. Token-restricted knowledge distillation module

[0050] The token-restricted knowledge distillation module guides the student model built on RoBERTa-base to learn the structured reasoning principles optimized through multiple rounds of fine-tuning and iteration. This enables the student model to jointly understand the deep semantics of interviews and the criteria for stigmatization, thereby effectively improving the classification and recognition performance of mental health stigmatization texts.

[0051] During the training phase, the evaluator uses real mental stigma categories to generate inference principles consistent with the actual labels; conversely, during the testing phase, the actual labels for mental stigma categories are hidden to prevent label leakage. Furthermore, considering that the evaluator can utilize actual labels to provide targeted revision suggestions, a more stringent exit mechanism (τ=27) is employed.

[0052] To overcome the issue of long text content exceeding the maximum input length limit of the student model, this application adds a compressor to achieve text simplification and redundancy filtering, condensing and reducing lengthy reasoning explanations after iterative optimization. This ensures that the overall number of tokens after combining the original interview text with the simplified reasoning principles is strictly controlled within the model's preset input threshold range, effectively avoiding the model input limitation problem caused by excessively long sequences.

[0053] ; in, This represents the reasoning principle that is finally output after fine-tuning and iteration. This represents the final output of the psychological stigma classification result after fine-tuning and iteration. Indicates compressor; This represents the simplified reasoning principle after compression. This application selects RoBERTa-base as the backbone network for the student model of knowledge distillation. Special identifiers are introduced. As a separator and connector, the original interview content is spliced ​​and fused with the compressed and simplified reasoning principles, and used as the input to the complete student model for subsequent training and feature learning.

[0054] ; in, Indicates splicing; This represents the input to the student model; The student model uses cross-entropy loss. Training is performed to closely approximate the output of the multi-agent system (generator + evaluator + fine-tuner): ; in, This represents the number of samples in the training set. This represents the final output of the psychological stigma classification result after fine-tuning iterations for the i-th training sample (interview content). For indicator functions, if ,but ,otherwise ; This represents the k-th psychological stigma category; The input to the student model is composed of a simplified reasoning principle and interview content from the i-th training sample. The student model input is Time prediction category is The posterior probability.

[0055] S3 uses the trained student model to predict the psychological stigma category of the target interview content.

[0056] Example 1 The experiments were primarily conducted on the MHStigmaInterview dataset, an expert-annotated corpus containing 4141 chatbot-respondent interviews designed for detecting and fine-grained classification of mental health stigma. Each interview was labeled either as a non-stigmatizing category or one of seven stigmatizing attribution categories derived from an attribution model. The dataset was split into training, validation, and test sets in a 6:2:2 ratio, and the performance of the multi-agent system and the principle-enhanced student model was evaluated on the test set.

[0057] In addition, the CBT-PC dataset from CBT-Bench, designed for multi-label classification of core beliefs, was introduced to verify the generalization ability and robustness of the proposed method. This dataset consists of 184 instances, taking patient narratives and related automatic thoughts as input, and labeling them with one or more categories from three subcategories: Helpless, Unlovable, and Worthless. The dataset was split into three groups (training, validation, and testing) in a 7:1.5:1.5 ratio. Evaluation of the multi-agent system was run directly on the overall dataset, while student models were evaluated on the test set.

[0058] The specific statistics for MHStigmaInterview and CBT-PC are shown in Table 1 below.

[0059] Table 1. Statistics of the original dataset

[0060] To verify the effectiveness of the method of this invention, comparative tests were conducted between the multi-agent system in the StigReDi architecture and the following LLMs: GPT-4o, Llama-3.1-8B, Llama-3.3-70B, Mistral Nemo-12B, and Mixtral8×7B. For the above LLM models, various prompting mechanisms were implemented in the experiments, including zero-shot architecture, one-shot architecture, and an architecture that provides a full-codebook approach to guide the model in performing discrimination and classification. Furthermore, a comparative experiment was conducted between the student model in the StigReDi architecture and the underlying RoBERTa classifier, which serves as the standard foundation, to assess and verify the superiority of the proposed solution.

[0061] Given the inherent class imbalance problem in the MHStigmaInterview dataset shown in Table 1, this embodiment uses metrics including macro precision (Macro-P), macro recall (Macro-R), macro F1 score (Macro-F1), and accuracy (Acc) to ensure a more comprehensive performance evaluation. Furthermore, Cohen's Kappa coefficient (Kappa) is introduced to quantify the consistency between the model's predictions and the true labels. For experiments on the CBT-PC dataset, standard evaluation protocols are followed, using macro recall (Macro-R), macro F1 score (Macro-F1), and accuracy (Acc) as evaluation metrics. All scores are the average of three independent runs.

[0062] Table 2 comprehensively presents a performance comparison of the proposed method with various baseline models on the test set. These results highlight the self-fine-tuning based on thought chains and its significant advantages in knowledge distillation. The comprehensive evaluation presented in Table 2 confirms the effectiveness of the student model in the StigReDi architecture in mental health stigma classification. Among pre-trained and LLM-based methods, the lightweight student model using RoBERTa-base as the backbone network and pre-trained with token-restricted knowledge distillation achieves the best overall performance. This model maintains the best performance on all evaluation metrics, including Macro-P, Macro-R, Macro-F1, Acc, and Kappa (as shown in bold in Table 2). By incorporating knowledge distillation, the student classifier inherits the powerful reasoning capabilities of multi-agent systems, even surpassing powerful models with large parameters such as GPT-4o in fine-grained mental stigma classification tasks. This highlights the criticality and advantages of combining multi-agent mechanisms (multi-agent mechanisms) with lightweight model architectures for efficient collaborative training. Furthermore, even without undergoing the distillation step, directly applying a multi-agent system equipped with self-fine-tuning components for evaluation and prediction demonstrates remarkable advantages, far exceeding those of various large-scale models used natively without intervention or adjustment. This proves that having self-correction capabilities can significantly improve the accuracy of all-dimensional recognition.

[0063] Table 2. Performance Comparison of StigReDi Architecture with Various Baseline Models on the MHStigmaInterview Dataset

[0064] To evaluate the generalization ability and system robustness of the StigReDi architecture, its performance was tested on the CBT-PC dataset, and the evaluation scores are shown in Table 3. Given that this benchmark is a multi-label coexistence classification task, the prompt settings were adjusted to guide the system to output all reasons supporting the correct category and their corresponding labels. The evaluation metrics in Table 3 confirm that the method provides excellent performance. The method combining iterative reflective correction and distillation architecture (StigReDi architecture) achieves the highest accuracy and leads in the core evaluation metrics of macro-recall (Macro-R) and macro-F1 (Macro-F1). This demonstrates that the classification and reasoning framework of this invention remains robust and has the potential for widespread application when moving beyond the initial stigmatization identification scenario to the more complex psychological cognitive multi-label domain. Furthermore, simply using the proxy loop system without distillation transformation to perform evaluation comparisons also outperforms the ultra-large-scale parameter model without the reflective system.

[0065] Table 3. Evaluation results of the StigReDi architecture and various baseline models on the CBT-PC dataset.

[0066] This application proposes the StigReDi architecture, a self-optimizing framework for thought chains based on multi-agent thinking that incorporates knowledge distillation, aiming to adapt large language models to fine-grained psychological stigma classification tasks in interviews. The multi-agent system utilizes a generator, evaluator, and fine-tuner to iteratively generate, evaluate, and modify classification reasoning processes based on psychological theories and expert codebooks. Furthermore, token-restricted reasoning process compression and knowledge distillation techniques transfer the rigorous analytical capabilities of the multi-agent system to a compact student model.

[0067] Experiments on the MHStigmaInterview benchmark and the CBT-PC transfer dataset demonstrate that the StigReDi architecture exhibits reliable and well-reasoned reasoning capabilities as well as excellent cross-dataset generalization. Compared to large-scale language models (LLMs), student models under the StigReDi architecture are not only parameter-efficient but also maintain analytical rigor while reducing computational costs.

[0068] The above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A psychological stigma classification method based on multi-agent self-correction and thought chain distillation, characterized in that, Includes the following steps: S1, Obtain the interview dataset; the interview content consists of a sequence of utterances, and the interview label is the probability of the psychological stigma category corresponding to the interview content. S2, construct a comprehensive model architecture that integrates multi-agent self-correction and thought chain distillation, as shown below: In the protocol-guided candidate generation module, the generator produces a preliminary output based on the interview content, including the psychological stigma classification results and the corresponding reasoning principles. In the scale-based quality assessment module, the evaluator assesses the quality of the psychological stigma classification results and the corresponding reasoning principles based on the rating scale and major error checking, and generates modification suggestions. In the fine-tuning iteration module of conditional feedback, the fine-tuner modifies the psychological stigma classification results and corresponding reasoning principles based on the feedback from the evaluator and in conjunction with the interview content. The modified mental stigma classification results and corresponding reasoning principles are further evaluated by the evaluator and modified by the fine-tuner. The fine-tuning iteration process continues until a predetermined exit condition is met, resulting in the final mental stigma classification results and corresponding reasoning principles. In the token-restricted knowledge distillation module, the compressor compresses the final reasoning principle to keep its total number of tokens within a preset threshold; the compressed, simplified reasoning principle and the interview content are used together as a supervision signal to train the student model, enabling the student model to predict the final psychological stigma classification result. S3 uses the trained student model to predict the psychological stigma category of the target interview content.

2. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1, characterized in that, The generator follows a preset six-step standardized protocol to progressively reason about the interview content, successively completing the abstraction of the interview theme, extraction of relevant evidence, verification of category rules, resolution of conflicting evidence, and soundness check of reasoning logic, and outputting preliminary psychological stigma classification results and corresponding reasoning principles.

3. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1, characterized in that, The evaluator performs a multi-dimensional assessment of the psychological stigma classification results and the corresponding reasoning principles, and generates modification suggestions, as follows: in, and and represent the mental stigma classification result and the corresponding reasoning principle input to the evaluator in the t-th fine-tuning iteration, respectively. When t=1, the evaluator input is the preliminary mental stigma classification result and the corresponding reasoning principle generated by the generator. This indicates the content of the interview; Represents the evaluator; Used to indicate whether a major error exists; This indicates the scoring across multiple evaluation dimensions; Descriptions indicating major and general errors; This indicates a suggestion for modification.

4. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1 or 3, characterized in that, Major errors include three categories: fabricated evidence, insufficient support, and violation of rules; the evaluation dimensions of the rating scale include seven dimensions: procedural compliance, fidelity of evidence, rule compliance, category completeness, logical consistency, semantic precision, and explanatory efficiency.

5. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1, characterized in that, The modifications to the psychological stigma classification results and reasoning principles during the fine-tuning process are guided by the principles of complete citation, minimal modification, and label consistency.

6. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1, characterized in that, The exit conditions for fine-tuning iterations include: There are no major errors, and the evaluator's total score across all evaluation dimensions exceeds the set upper limit threshold τ for the total evaluation score; In consecutive N err During the next fine-tuning iteration, several major errors occurred. The improvement in the total assessment score that can be achieved after fine-tuning remains at a marginal level within M consecutive fine-tuning iterations; The number of fine-tuning iterations exceeds the set upper limit threshold N. iter .

7. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1, characterized in that, The compressor compresses the final reasoning principle: in, This represents the reasoning principle that is finally output after fine-tuning and iteration. This represents the final output of the psychological stigma classification result after fine-tuning and iteration; Indicates compressor; This represents the simplified reasoning principle after compression. RoBERTa-base was used as the backbone network of the student model; interview content was concatenated and fused with compressed and simplified reasoning principles as input to the student model. in, This indicates the content of the interview; Indicates splicing; Represents the input to the student model; introduces dedicated special identifiers. As a separator / connector segment.

8. The psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in claim 1, characterized in that, The student model uses cross-entropy loss. Training is performed to approximate the final output of the psychological stigma classification result after fine-tuning iterations, using cross-entropy loss. As shown below: in, This represents the number of samples in the training set. This represents the final output of the psychological stigma classification result after fine-tuning iterations for the i-th training sample. For indicator functions, if ,but ,otherwise ; This represents the k-th psychological stigma category; The input to the student model is composed of a simplified reasoning principle and interview content from the i-th training sample. The student model input is Time prediction category is The posterior probability.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the psychological stigma classification method based on multi-agent self-correction and thought chain distillation as described in any one of claims 1 to 8.

Images