Incremental pre-training and emotional response optimization method for large language model in psychological field
By constructing a multi-level dataset and incremental pre-training optimization method, the problems of insufficient professional knowledge adaptation and empathy ability of large language models in the field of psychological services were solved, personalized and standardized response generation was achieved, and data and resource requirements were reduced.
Patent Information
- Application Number
- CN202511022385.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-24
- Publication Date
- 2025-09-12
AI Technical Summary
Existing general-purpose large language models lack a deep understanding of professional terminology and clinical intervention processes in the field of psychological services, have weak empathy capabilities, and find it difficult to generate responses that comply with industry standards and are personalized.
By constructing a multi-level psychological domain dataset, adopting incremental pre-training and knowledge module isolation strategies, combining emotional feature extraction and empathy strategy generation, dynamic adaptation and real-time optimization are performed to achieve efficient absorption of professional knowledge and personalized response of the model.
The model can accurately understand professional terminology and intervention logic, generate personalized responses that comply with industry standards, reduce data volume and computing resource requirements, and avoid inappropriate expressions and repetitive responses.
Smart Images
Figure CN120636705A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intersectional technology of artificial intelligence and mental health, and specifically to a method for incremental pre-training and empathy response optimization of a large language model in the psychological field. Background Art
[0002] With the rapid development of big language model technology, its application potential in the field of psychological services is becoming increasingly prominent. However, existing general-purpose big language models have two core flaws: First, they are insufficiently adapted to domain knowledge. General-purpose models lack a deep understanding of psychological terminology (such as "cognitive restructuring" and "emotional dissociation"), clinical intervention processes (such as the stages of CBT therapy), and pathological characteristics (such as the somatization of anxiety disorders), which can easily lead to responses that do not meet industry standards. Second, their empathy is weak. Models have difficulty accurately identifying users' emotional metaphors (such as "I feel like a stone is weighing on my heart") and potential psychological needs (such as expressing emotional distress through complaining about physical symptoms). Responses often appear to be "mechanical comfort," failing to establish effective emotional connections.
[0003] Existing technologies require massive amounts of labeled data (psychological data is highly private and difficult to obtain) and computing resources, and can easily lead to "catastrophic forgetting" of general knowledge in the model. Adjusting model parameters using only a small amount of conversation data makes it difficult for the model to grasp complex psychological intervention logic, and no special optimization mechanism is designed for empathy. Relying on manually preset empathy response templates, it has poor flexibility, cannot cope with diverse user expressions, and is prone to repetitive and patterned responses. In view of this, we propose a method for incremental pre-training and empathy response optimization for large language models in the psychological field. Summary of the Invention
[0004] The purpose of the present invention is to provide a method for incremental pre-training and empathy response optimization of a large language model in the psychological field to solve the problems raised in the above background technology.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for incremental pre-training and empathy response optimization of a large language model in the psychological field, comprising the following steps:
[0006] S1. Construct a multi-level psychological domain dataset, which includes professional knowledge data, clinical case data, and empathy annotation data, and preprocess the data;
[0007] S2. Perform incremental pre-training optimization on the base model, including base model selection, incremental knowledge injection, and forgetting suppression training;
[0008] S3. Conduct empathy response optimization training, including emotion feature extraction, empathy strategy generation, and empathy response generation and iteration;
[0009] S4. Achieve dynamic adaptation and real-time optimization, including domain knowledge updates, empathy calibration, and real-time feedback adjustments.
[0010] Optionally, constructing a multi-level psychological domain dataset in S1 includes:
[0011] S11. Data collection: collecting professional knowledge data, clinical case data, and empathy annotation data;
[0012] S12. Data preprocessing: Use the federated learning framework for distributed data cleaning, perform structured text processing, add knowledge tags to professional knowledge data, annotate intervention techniques for case data, and build an emotion-response mapping library for empathy data.
[0013] Optionally, the professional knowledge data includes psychology textbooks, industry guidelines and academic papers; the clinical case data are psychological counseling records and psychological crisis intervention dialogue transcripts after desensitization processing; the empathy annotation data are user emotional expression texts annotated with empathy levels and reasons by senior psychological counselors.
[0014] Optionally, the incremental pre-training optimization of the base model in S2 includes:
[0015] S21. Select a large language model with strong conversational capabilities as the base model;
[0016] S22, adopting the knowledge module isolation strategy, adds a psychological knowledge encoding layer to the original model architecture, and inputs the data set constructed in S1 in stages for incremental knowledge injection;
[0017] S23. After each round of incremental training, a general knowledge verification set is introduced to calculate the performance decay rate of the model on general tasks. If the decay rate exceeds 5%, parameter distillation technology is used to freeze some parameters of the model on general tasks.
[0018] Optionally, the data sets are input in stages in S22 as follows: in the first stage, professional knowledge data is input, and the model is taught the exact meaning of professional vocabulary through term-context binding training; in the second stage, clinical case data is input, and the model is taught the relationship between intervention technology and effect through causal relationship learning.
[0019] Optionally, the S3 empathy response optimization training includes:
[0020] S31. Construct an emotion-semantics bimodal encoder to process user input text, identify core appeals at the semantic level, and extract emotion feature vectors at the emotional level using a fine-grained emotion classification model combined with emotion intensity scoring.
[0021] S32. Training an empathy strategy decision model based on the emotion-response mapping library. The model input is the user's emotion feature vector and conversation context, and the output is an empathy response strategy. The strategy generation must meet the psychological intervention compliance.
[0022] S33. Input the empathy strategy as an instruction to the model trained in S2 to generate a preliminary response. Use human feedback reinforcement learning optimization, use the counselor's empathy quality score for the response as a reward signal, adjust the model parameters through the proximal policy optimization algorithm and iterate training.
[0023] Optionally, the compliance of the psychological intervention specifically includes avoiding giving medical diagnosis and not promising absolute effects.
[0024] Optionally, the dynamic adaptation and real-time optimization in S4 include:
[0025] S41. Collect the latest psychological research results and clinical guidelines every month, and update the parameters of the psychological knowledge encoding layer through lightweight incremental training;
[0026] S42. Introduce new empathy annotation data every quarter and retrain the empathy strategy decision model;
[0027] S43. In actual applications, real-time monitoring of user feedback on responses is performed. When three consecutive negative feedbacks are received, an emergency optimization process is triggered, and manual intervention is performed to adjust the model parameters.
[0028] Compared with the existing technology, the present invention provides a method for incremental pre-training and empathy response optimization of a large language model in the psychological field, which has the following beneficial effects:
[0029] 1. This method of incremental pre-training and empathy response optimization for the large language model in the psychological field uses a strategy of incremental pre-training and knowledge module isolation to enable the model to efficiently absorb professional knowledge in the psychological field without compromising the model's general capabilities. It can accurately understand professional terminology and intervention logic, and respond in accordance with industry standards. At the same time, through the fine-grained extraction of emotional features and the dynamic generation of empathy strategies, the model can identify users' explicit and implicit emotions, generate personalized and warm responses, and avoid stereotyped comfort.
[0030] 2. The incremental pre-training and empathy response optimization method for the large language model in the psychological field does not require full retraining. Through incremental updates and targeted optimization, it reduces the demand for data volume and computing resources. At the same time, it adopts forgetting suppression technology to retain the original advantages of the model; compliance verification is embedded in the generation of empathy responses to avoid inappropriate expressions and reduce the risks of psychological services. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a structural schematic diagram of the present invention. DETAILED DESCRIPTION
[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0033] like Figure 1 As shown, the present invention provides a technical solution: a method for incremental pre-training and empathy response optimization of a large language model in the psychological field, comprising the following steps:
[0034] S1. Construct a multi-level psychological domain dataset, which includes professional knowledge data, clinical case data, and empathy annotation data, and pre-process the data. Constructing a multi-level psychological domain dataset includes:
[0035] S11. Data collection: collecting professional knowledge data, clinical case data, and empathy annotation data;
[0036] S12. Data preprocessing: Use the federated learning framework for distributed data cleaning, perform structured text processing, add knowledge tags to professional knowledge data, annotate intervention techniques for case data, and build an emotion-response mapping library for empathy data.
[0037] Among them, professional knowledge data includes psychology textbooks, industry guidelines and academic papers; clinical case data are psychological counseling records and psychological crisis intervention dialogue transcripts after desensitization processing; empathy annotation data are user emotional expression texts labeled with empathy levels and reasons by senior psychological counselors.
[0038] S2. Perform incremental pre-training optimization on the base model, including base model selection, incremental knowledge injection, and forgetting suppression training;
[0039] Among them, the incremental pre-training optimization of the base model includes:
[0040] S21. Select a large language model with strong conversational capabilities as the base model;
[0041] S22. Adopting the knowledge module isolation strategy, a new psychological knowledge encoding layer is added to the original architecture of the model, and the data set constructed by S1 is input in stages for incremental knowledge injection; the specific stages of inputting the data set are as follows: in the first stage, professional knowledge data is input, and the model is taught the accurate meaning of professional vocabulary through term-context binding training; in the second stage, clinical case data is input, and the model is taught to understand the relationship between intervention technology and effect through causal relationship learning.
[0042] S23. After each round of incremental training, a general knowledge verification set is introduced to calculate the performance decay rate of the model on general tasks. If the decay rate exceeds 5%, parameter distillation technology is used to freeze some parameters of the model on general tasks.
[0043] S3. Conduct empathy response optimization training, including emotion feature extraction, empathy strategy generation, and empathy response generation and iteration;
[0044] Empathic response optimization training includes:
[0045] S31. Construct an emotion-semantics bimodal encoder to process user input text, identify core appeals at the semantic level, and extract emotion feature vectors at the emotional level using a fine-grained emotion classification model combined with emotion intensity scoring.
[0046] S32. Train the empathy strategy decision model based on the emotion-response mapping library. The model input is the user's emotion feature vector and conversation context, and the output is the empathy response strategy. The strategy generation must meet the psychological intervention compliance. The psychological intervention compliance specifically means avoiding medical diagnosis and not promising absolute effects.
[0047] S33. Input the empathy strategy as an instruction to the model trained in S2 to generate a preliminary response. Use human feedback reinforcement learning optimization, use the counselor's empathy quality score for the response as a reward signal, adjust the model parameters through the proximal policy optimization algorithm and iterate training.
[0048] S4. Achieve dynamic adaptation and real-time optimization, including domain knowledge updates, empathy calibration, and real-time feedback adjustments;
[0049] Dynamic adaptation and real-time optimization include:
[0050] S41. Collect the latest psychological research results and clinical guidelines every month, and update the parameters of the psychological knowledge encoding layer through lightweight incremental training;
[0051] S42. Introduce new empathy annotation data every quarter and retrain the empathy strategy decision model;
[0052] S43. In actual applications, real-time monitoring of user feedback on responses is performed. When three consecutive negative feedbacks are received, an emergency optimization process is triggered, and manual intervention is performed to adjust the model parameters.
[0053] As an application of this embodiment:
[0054] This embodiment, through the strategy of incremental pre-training and knowledge module isolation, allows the model to efficiently absorb professional knowledge in the field of psychology without compromising the model's general capabilities. It can accurately understand professional terminology and intervention logic, and respond in accordance with industry standards. At the same time, through the fine-grained extraction of emotional features and the dynamic generation of empathy strategies, the model can identify users' explicit and implicit emotions, generate personalized and warm responses, and avoid stereotyped comfort. No full retraining is required. Through incremental updates and targeted optimization, the demand for data volume and computing resources is reduced. At the same time, the use of forgetting suppression technology retains the original advantages of the model. Compliance verification is embedded in the generation of empathy responses to avoid inappropriate expressions and reduce the risks of psychological services.
[0055] The above generally describes the present invention in detail. However, it is obvious to those skilled in the art that modifications or improvements may be made based on the present invention. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.
Claims
1. A method for incremental pre-training and empathy response optimization of a large language model in the psychological domain, characterized by: The following steps are involved: S1. Construct a multi-level psychological domain dataset, which includes professional knowledge data, clinical case data, and empathy annotation data, and preprocess the data; S2. Perform incremental pre-training optimization on the base model, including base model selection, incremental knowledge injection, and forgetting suppression training; S3. Conduct empathy response optimization training, including emotion feature extraction, empathy strategy generation, and empathy response generation and iteration; S4. Achieve dynamic adaptation and real-time optimization, including domain knowledge updates, empathy calibration, and real-time feedback adjustments.
2. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 1 is characterized in that: The multi-level psychological domain dataset constructed in S1 includes: S11. Data collection: collecting professional knowledge data, clinical case data, and empathy annotation data; S12. Data preprocessing: Use the federated learning framework for distributed data cleaning, perform structured text processing, add knowledge tags to professional knowledge data, annotate intervention techniques for case data, and build an emotion-response mapping library for empathy data.
3. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 2, characterized in that: The professional knowledge data includes psychology textbooks, industry guidelines and academic papers; the clinical case data are psychological counseling records and psychological crisis intervention dialogue transcripts after desensitization processing; the empathy annotation data are user emotional expression texts annotated with empathy levels and reasons by senior psychological counselors.
4. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 1, characterized in that: The incremental pre-training optimization of the base model in S2 includes: S21. Select a large language model with strong conversational capabilities as the base model; S22, adopting the knowledge module isolation strategy, adds a psychological knowledge encoding layer to the original model architecture, and inputs the data set constructed in S1 in stages for incremental knowledge injection; S23. After each round of incremental training, a general knowledge verification set is introduced to calculate the performance decay rate of the model on general tasks. If the decay rate exceeds 5%, parameter distillation technology is used to freeze some parameters of the model on general tasks.
5. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 4, characterized in that: The data sets are input in stages in S22 as follows: in the first stage, professional knowledge data is input, and the model is taught the exact meaning of professional vocabulary through term-context binding training; in the second stage, clinical case data is input, and the model is taught the relationship between intervention technology and effect through causal relationship learning.
6. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 1, characterized in that: The S3 Empathic Response Optimization Training includes: S31. Construct an emotion-semantics bimodal encoder to process user input text, identify core appeals at the semantic level, and extract emotion feature vectors at the emotional level using a fine-grained emotion classification model combined with emotion intensity scoring. S32. Training an empathy strategy decision model based on the emotion-response mapping library. The model input is the user's emotion feature vector and conversation context, and the output is an empathy response strategy. The strategy generation must meet the psychological intervention compliance. S33. Input the empathy strategy as an instruction to the model trained in S2 to generate a preliminary response. Use human feedback reinforcement learning optimization, use the counselor's empathy quality score for the response as a reward signal, adjust the model parameters through the proximal policy optimization algorithm and iterate training.
7. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 6, characterized in that: The compliance of psychological intervention specifically includes avoiding giving medical diagnosis and not promising absolute effects.
8. The incremental pre-training and empathy response optimization method for a large language model in the psychological domain according to claim 1, characterized in that: The dynamic adaptation and real-time optimization in S4 include: S41. Collect the latest psychological research results and clinical guidelines every month, and update the parameters of the psychological knowledge encoding layer through lightweight incremental training; S42. Introduce new empathy annotation data every quarter and retrain the empathy strategy decision model; S43. In actual applications, real-time monitoring of user feedback on responses is performed. When three consecutive negative feedbacks are received, an emergency optimization process is triggered, and manual intervention is performed to adjust the model parameters.