Multi-strategy emotion regulation and hybrid cognitive decision-making method based on pre-training model

Abstract

A multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-training model is used for efficiently converting negative texts and comprises the following steps: firstly, receiving and quantifying the negative texts and situational features through an input processing step; then, a HyperNetwork module is utilized to generate a dynamic strategy weight based on a situation feature; then, through a plurality of fine-tuned Adapter strategy units, candidate texts of various positive psychological intervention strategies are generated according to the dynamic weights; performing weighted fusion on the candidate texts to form a preliminary rewritten text; furthermore, through a hybrid cognitive dynamic decision module, a'thinking-generation 'iteration mechanism is adopted to carry out self-evaluation and optimization on the text until a final rewritten text conforming to a positive psychological principle is output. The method solves the problems that a traditional text rewriting system is single in strategy, insufficient in dynamic adaptability and low in output consistency, multi-angle positive rewriting of negative texts is achieved, and it is ensured that the output texts conform to the positive psychological principle and have the high emotion guiding effect and natural language expression.

Description

Technical Field

[0001] The present invention relates to the cross - field of natural language processing, positive psychology, and cognitive science, and particularly to a multi - strategy emotion regulation and hybrid cognitive decision - making method based on a pre - trained model. Background Art

[0002] In recent years, mental health problems among teenagers have shown a trend of getting younger and more concealed. Negative emotional expressions (such as "I will never succeed" and "The future is hopeless") are widespread in social media and online communication. Traditional psychological intervention methods and text rewriting systems usually rely on rule templates or single - generation models for processing. Their main defects are single - strategy, rigid expression, easy to cause semantic distortion or emotional conflict, which may exacerbate negative mental states. When facing sensitive situations of teenagers such as social anxiety and academic pressure, existing technologies often cannot take into account both the emotional delicacy of language and the diverse needs of personalized expression, thus having great limitations in the intervention effect.

[0003] With the development of pre - trained language models (such as GPT, BERT, T5, etc.), the adapter method based on lightweight fine - tuning technology provides a new technical idea for multi - strategy integration. Existing research has achieved independent expression of different psychological intervention strategies (such as growth mindset, social support, humorous perspective, etc.) through fine - tuning pre - trained models with Adapters. However, these methods still have deficiencies in dynamic weight allocation, multi - strategy collaborative generation, and ensuring the consistency of output text. Specifically:

[0004] Insufficient dynamic allocation of strategy weights Traditional methods usually adopt fixed weights or simple linear combinations, with limited ability to analyze input context features (such as emotion intensity, dialogue context, user profile, etc.), and it is difficult to dynamically adjust the contribution ratio of each strategy according to real - time changes, resulting in rigidity and limitations in the emotional guidance of the generated text.

[0005] Difficulty in multi - strategy collaborative generation Rewriting systems under single or fixed strategies are prone to semantic conflicts and cannot fully integrate the advantages of multiple positive psychological intervention strategies. For example, when rewriting "I will never succeed", a single humorous expression may be misunderstood by users, while a single growth mindset strategy may lack sufficient emotional support.

[0006] Lack of generation consistency and iterative optimization Most existing rewriting systems adopt a one - time generation method and cannot perform self - evaluation and dynamic adjustment after generation, resulting in the output text being difficult to reach the optimal state in terms of coherence, positive emotion expression, and language quality. The traditional generation method lacks an iterative process similar to "reflection - correction" in human cognition and is difficult to achieve continuous optimization.

[0007] It should be noted that the information disclosed in the above background art is only used for understanding the background of the present application, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0008] The main object of the present invention is to overcome the defects existing in the above background art, and provide a multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] A multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model, comprising the following steps:

[0011] S1. Input processing: Receive negative text and situational features, and use a pre-trained language model to vectorize the text to generate a text vector representation;

[0012] S2. Dynamic weight generation: Through a HyperNetwork module, based on the situational features, use a multi-layer perceptron to generate a strategy weight vector and perform normalization processing to obtain a dynamic strategy weight distribution;

[0013] S3. Multi-strategy text generation: Based on the dynamic strategy weights, use multiple independently fine-tuned Adapter strategy adaptation units to fine-tune the pre-trained model, and independently generate candidate texts corresponding to multiple different positive psychological intervention strategies;

[0014] S4. Text fusion: According to the dynamic strategy weights, perform weighted fusion on the candidate texts to form a preliminary rewritten text.

[0015] Further, the method further comprises the following steps:

[0016] S5. Hybrid cognitive decision-making: Through a hybrid cognitive dynamic decision-making module, adopt a "think-generate" alternating iteration mechanism, and perform self-evaluation and feedback optimization on the rewritten text according to a preset evaluation function, including strategy weight adjustment or local text correction, until a final rewritten text that meets the preset criteria is generated.

[0017] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model.

[0018] A computer program product includes a computer program, and when the computer program is executed by a processor, it implements the multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model.

[0019] The present invention has the following beneficial effects:

[0020] The present invention proposes a multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model to achieve efficient and accurate transformation of negative texts. Specifically, the present invention uses a HyperNetwork module to deeply analyze the contextual features of the input text, thereby generating a dynamic policy weight distribution, which not only ensures the lightweight of the model but also enhances the scalability of the strategy. Compared with traditional methods, the present invention can dynamically adjust the contribution ratio of each strategy according to the real-time changing contextual features, thus avoiding the rigidity and limitations in emotional guidance of the generated text. Multiple independently fine-tuned Adapter strategy modules are used to generate candidate rewritten texts in parallel, and a candidate text optimization algorithm guided by policy weights is used for weighted fusion. This process ensures that the final output text has obvious advantages in emotional guidance, expression coherence, and naturalness. This method effectively solves the semantic conflict problem that may occur during multi-strategy fusion and ensures the logical consistency and positive emotion of the generated results. Further, the present invention innovatively introduces an alternating iterative hybrid cognitive process of thinking (T) and generation (G). The system conducts self-evaluation after each round of generation, and adjusts the policy weights or makes local corrections to the text through a feedback mechanism, thereby gradually improving the generation effect. This iterative optimization process effectively makes up for the deficiencies of traditional one-time generation methods, making the output results more in line with the principles of positive psychology.

[0021] The present invention not only breaks through the technical bottlenecks of traditional methods in terms of strategy scalability, generation consistency, and dynamic adaptability, but also provides a standardized, low-cost, and easy-to-sustainably-expand positive psychological intervention technical support for multiple scenarios such as psychological counseling, educational guidance, and social media intervention. Through multi-strategy fusion and hybrid cognitive dynamic decision-making, the present invention can achieve precise transformation for typical negative expressions. For example, dynamically rewrite "I will never succeed" as "By setting phased goals, I can gradually break through", thereby significantly improving the psychological adaptability and emotional guidance effect of negative texts for teenagers.

[0022] In summary, the present invention provides an efficient, flexible, and personalized emotion intervention method, which can achieve positive transformation of negative texts in a variety of application scenarios, and has broad application prospects and practical value.

[0023] Other beneficial effects in the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flowchart of the multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is merely exemplary and not intended to limit the scope of the present invention and its applications.

[0026] In addition, the terms "first" and "second" are only used for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present invention, the meaning of "a plurality" is two or more, unless otherwise specifically defined.

[0027] The present invention dynamically allocates a variety of positive psychological intervention strategies according to situational features such as emotion tags and dialogue turns, and proposes an improvement scheme for the following problems existing in the prior art:

[0028] Single strategy and limited expression: Existing rewriting systems based on rule templates or single generation models are prone to semantic distortion or emotional conflict when facing complex situations (such as adolescent social anxiety, academic pressure, etc.), and cannot achieve a comprehensive transformation of negative texts.

[0029] Defects in static weight allocation: Traditional methods often assign fixed or simple linear combinations of weights to different psychological intervention strategies, lacking the ability to adjust dynamically in real time according to the situation, resulting in the rewritten results being difficult to balance emotional fineness and personalized expression.

[0030] Insufficient optimization due to one-time generation: Most existing rewriting systems adopt the method of outputting text once, lacking a self-evaluation and iterative correction mechanism, and unable to continuously optimize aspects such as text coherence and positive emotional expression after generation.

[0031] Semantic conflict problems in multi-strategy fusion: When using multiple strategies to generate candidate texts in parallel, how to avoid semantic conflicts between candidate texts and how to perform effective weighted fusion to ensure that the generated results are logically consistent and emotionally positive still lack systematic solutions.

[0032] The present invention aims to solve the problems of single strategy, insufficient dynamic adaptability and low output consistency in traditional rewriting systems. Through the dynamic generation of strategy weights by HyperNetwork, the parallel generation of multiple strategies by Adapter and the hybrid cognitive dynamic decision-making mechanism, it realizes the multi-angle positive rewriting of negative texts, ensuring that the output text not only conforms to the principles of positive psychology, but also has a high emotional guidance effect and natural language expression.

[0033] Refer to Figure 1 , the embodiments of the present invention provide a multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model, including the following steps:

[0034] Step S1, Input Processing: Receive negative text and situational features, and use a pre-trained language model to vectorize the text to generate a text vector representation;

[0035] Step S2, Dynamic Weight Generation: Through the HyperNetwork module, based on the situational features, use a multi-layer perceptron to generate a policy weight vector and perform normalization processing to obtain a dynamic policy weight distribution;

[0036] Step S3, Multi-Strategy Text Generation: Based on the dynamic policy weights, use multiple independently fine-tuned Adapter policy adaptation units to fine-tune the pre-trained model to independently generate candidate texts corresponding to multiple different positive psychological intervention strategies;

[0037] Step S4, Text Fusion: According to the dynamic policy weights, perform weighted fusion on the candidate texts to form a preliminary rewritten text.

[0038] See Figure 1 , in a preferred embodiment, the method further includes the following steps:

[0039] Step S5, Hybrid Cognitive Decision-Making: Through a hybrid cognitive dynamic decision-making module, adopt a "think-generate" alternating iteration mechanism, and self-evaluate and feedback optimize the rewritten text according to a preset evaluation function, including policy weight adjustment or local text correction, until a final rewritten text that meets the preset criteria is generated.

[0040] The present invention provides a multi-strategy positive psychological rewriting system and a hybrid cognitive dynamic decision-making method based on a pre-trained language model. The core lies in addressing issues such as single strategy, static weight allocation, and insufficient generation consistency in the prior art. By introducing dynamic weight allocation, multi-strategy parallel generation, and a hybrid cognitive process of think-generate alternating iteration, it realizes the efficient transformation and continuous optimization of negative text, thereby achieving a text rewriting effect that better conforms to the principles of positive psychology. Through multi-strategy fusion, dynamic adjustment, and the hybrid cognitive dynamic decision-making method, the present invention achieves precise, flexible, and efficient rewriting of negative text, breaking through the limitations of traditional single-strategy and one-time decision-making methods, and generating more natural and personalized positive text.

[0041] In some embodiments, step S1 specifically includes:

[0042] S11. Receive the negative text and situational features input by the user, where the situational features include sentiment labels, dialogue context information, and user portraits; the situational features may include sentiment labels, dialogue context information, and user portraits;

[0043] S12. Parse and vectorize and encode the negative text and situational features.

[0044] In a preferred embodiment, step S2 specifically includes:

[0045] S21. Extract the emotional features, context association features, and user features in the input text vector through the HyperNetwork;

[0046] S22. Predict the initial weight values of each intervention strategy based on the extracted multi-dimensional features;

[0047] S23. Use the Softmax function to normalize the initial weight values to generate dynamic strategy weights that meet the requirements of the probability distribution.

[0048] In some embodiments, step S3 specifically includes:

[0049] S31. Configure an independent Adapter fine-tuning module for each preset psychological intervention strategy, and each module updates parameters efficiently based on the shared pre-trained language model;

[0050] S32. Each Adapter module independently generates candidate rewritten texts for the corresponding intervention strategy based on the input text vector representation;

[0051] S33. According to the dynamic strategy weights, perform weighted fusion calculation on the text generation probabilities output by each Adapter module.

[0052] In some embodiments, step S5 specifically includes:

[0053] S51. Establish an initial thinking state based on the preliminary rewritten text;

[0054] S52. Thinking-generation is executed alternately:

[0055] In the thinking stage, perform multi-dimensional evaluation on the current text and update the thinking state;

[0056] In the generation stage, generate an optimized text based on the updated thinking state;

[0057] S53. Determine whether the iteration termination condition is met through a preset evaluation mechanism. If not, repeat S52.

[0058] Through the dynamic generation of strategy weights by the HyperNetwork, parallel generation of multiple strategies by the Adapter, weighted fusion of candidate texts, and the hybrid cognitive dynamic decision-making mechanism, the embodiments of the present invention ensure that the output text has a significant positive emotional guidance effect while being semantically coherent and logically clear.

[0059] The following further describes the implementation process and algorithm examples of specific embodiments of the present invention.

[0060] 1. Input module: Receives negative text and situational features, and vectorizes the text using sentence embedding technology;

[0061] The input module is responsible for receiving the negative text input by the user, as well as situational features such as emotion labels and dialogue turns, and converting the text into a vector representation through sentence embedding technology (such as pre-trained models like BERT and T5), which serves as the basic data for subsequent processing.

[0062] Data input and preprocessing, receiving negative text and situational features, and vectorizing and encoding the text using sentence embedding technology. The input module includes text parsing and vectorization processing. Common implementation methods include using models like BERT and T5 to convert the text into vectors of a fixed dimension, providing basic data for subsequent policy weight generation and text generation.

[0063] 2. HyperNetwork module: Generates dynamic policy weights through a multi-layer perceptron based on situational features, and normalizes them using Soft-max;

[0064] The HyperNetwork module calculates and generates a policy weight vector using a multi-layer perceptron based on the situational features provided by the input module, and normalizes it through the Softmax function to obtain a dynamic policy weight distribution. This module directly determines the focus of work for subsequent policy generation units.

[0065] Based on the preprocessed situational features, use the HyperNetwork module to generate a policy weight vector and normalize it through the Softmax function. This step adopts a multi-layer perceptron structure to generate a weight vector according to information such as emotion labels and dialogue turns. The normalization of the weights ensures the reasonable distribution of each intervention strategy during the subsequent candidate text generation process, directly affecting the quality and emotional orientation of the final rewritten text.

[0066] Dynamic policy weight allocation mechanism. Through the dynamic weight allocation method of the HyperNet-work module, it can adjust the weights of each intervention strategy according to the emotional features and context information of the input text, thus ensuring that the emotion of the rewritten text meets the user's needs.

[0067] 3. Adapter policy module: Consists of multiple independently fine-tuned adapter units, each unit corresponding to an active intervention strategy (such as growth mindset, humorous perspective), and uses LoRA / AdapterHub technology to achieve efficient parameter adjustment;

[0068] The Adapter strategy module consists of multiple independently fine-tuned adaptation units. Each unit generates candidate texts for a specific positive psychological intervention strategy (e.g., growth mindset, humorous perspective, social support, etc.). Implementation can use technologies such as LoRA and AdapterHub to ensure efficient parameter adjustment of each strategy module based on the pre-trained model, and to ensure the diversity and accuracy of the generated texts in terms of language style and content orientation.

[0069] In the Adapter strategy module, each adaptation unit independently generates candidate rewritten texts for the corresponding positive psychological intervention strategy. Each adaptation unit uses fine-tuning techniques (such as LoRA or AdapterHub) to generate candidate texts based on the pre-trained language model. Each unit can include multiple sub-steps, such as strategy feature extraction, text generation, and post-processing, etc., to ensure that the output texts can reflect their respective strategy styles.

[0070] Using the Adapter technology to fine-tune the pre-trained model can be achieved through different fine-tuning methods (such as LoRA, Prefix Tuning, etc.). Maintaining the diversity of adaptation strategies, each Adapter module independently and efficiently generates texts that meet the emotional needs.

[0071] 4. Text fusion module: Weighted fusion is performed based on the candidate texts generated by each strategy and their dynamic weights, and finally the rewritten result is output.

[0072] The text fusion module performs weighted fusion on the candidate texts output by each Adapter strategy module according to the dynamic weights generated by the HyperNetwork module to form the final rewritten result. This module is the core of the multi-strategy parallel generation mechanism, ensuring that the finally output text not only has a positive emotional guidance effect but also has high coherence and naturalness.

[0073] According to the dynamic strategy weights generated in step 2, weighted fusion is performed on each candidate text obtained in step 3 to form a preliminary rewritten text. The text fusion module integrates multiple candidate texts according to the corresponding strategy weights, and the specific method can be additive fusion, splicing, or more complex probability model calculations, so as to achieve the fusion of multi-strategy information and ensure the diversity and consistency of the text content.

[0074] The present invention can generate multiple candidate texts in parallel in the same system and generate the final rewritten text through weighted fusion. This feature ensures that the generated text has diversity and emotional guidance.

[0075] 5. The Hybrid Cognitive Dynamic Decision-making Module further enhances the system performance by self-evaluating and feedback-optimizing the preliminary generation results through an "Iterative Thinking-Generation" mechanism. This module can dynamically adjust the strategy weights or correct local text according to a preset evaluation function, thereby improving the positivity and coherence of the generation results.

[0076] Through the Hybrid Cognitive Dynamic Decision-making Module, an "Iterative Thinking-Generation" process is implemented to self-evaluate and feedback-optimize the initially generated rewritten text. This step evaluates the text by setting evaluation functions (such as positivity score, coherence score, etc.). If the evaluation does not meet the expectations, the strategy weights are dynamically adjusted or the text is locally corrected until the preset goal is achieved. This iterative process can adopt a multi-round feedback and optimization mechanism.

[0077] The Iterative Thinking-Generation mechanism continuously adjusts the strategy after each generation through self-evaluation and feedback-optimization to enhance the emotional guidance effect and language fluency of the generation results.

[0078] In a specific embodiment, the present invention includes the following processing procedures:

[0079] Input Module: Responsible for parsing and vectorizing the input negative text and situational features (such as emotional intensity, dialogue context, user profile, etc.), and providing input features for subsequent modules.

[0080] After the system starts, it first receives the input negative text and related situational information, and performs preprocessing and vectorization encoding.

[0081] Feature Extraction and Dynamic Weight Generation Module: For the input text embedding vector x (where x is the vector representation of the input text), the calculation formula of the HyperNetwork is:

[0082]

[0083] where n represents the number of intervention strategies. Subsequently, through normalization by the Softmax function, the dynamic weights of each strategy are obtained:

[0084]

[0085] Extract the features of the input text through the HyperNetwork, and generate a dynamic strategy weight distribution w = [w1,..., w n .

[0086] Adapter strategy module: For different psychological intervention strategies (such as growth mindset, self-acceptance, future hope, social support, humorous perspective, etc.), the pre-trained model is fine-tuned using adapter technology (for example, using LoRA to achieve efficient parameter updates). Each adapter corresponds to a strategy and generates candidate rewrite texts respectively. Suppose the probability distribution of candidate texts generated by the i-th strategy module is

[0087] p i (y|x),

[0088] Finally, the output of each module is weighted and fused to obtain the rewritten text:

[0089]

[0090] Candidate text generation: For each pre-tuned Adapter strategy module, candidate rewrite texts are generated. For example, the growth mindset module generates candidate text y1; the social support module generates candidate text y2; and the humor perspective module generates candidate text y3.

[0091] Hybrid cognitive dynamic decision module: This module uses the mechanism of alternating iteration of thinking (Think) and generating (Generate) to self-evaluate and optimize the initially generated text. Suppose the thinking state of the kth round is s k , the update formula is:

[0092] s k =T k (s k-1 ,y k-1 ),

[0093] And generate improved text based on the updated state:

[0094] y k =G k (s k ).

[0095] The iteration process continues until the preset evaluation function reaches the stopping threshold θ stop or the maximum number of iterations is reached.

[0096] Hybrid cognitive dynamic decision iteration: The system evaluates the initially generated text. If it finds that the text still does not meet expectations in some emotions or expressions, it adjusts the strategy weights and continuously corrects the text through thinking-generation alternating iterations. After each round of iteration, the text quality is judged by a preset evaluation function until it meets the predetermined standards.

[0097] As a preferred example, the preset evaluation function is defined as:

[0098]

[0099] Among them:

[0100] y k is the text generated in the k-th round; represents the fluency evaluation, which measures the naturalness of the text; represents the information coverage evaluation, which measures the relevance and information integrity between the generated text and the input question; represents the relevance evaluation, which measures the matching degree and consistency between the generated text and the input text.

[0101] Each evaluation item corresponds to a weight w1, w2, w3, and these weights can be adjusted according to the task requirements.

[0102] Fluency evaluation L(y k ):

[0103] The fluency evaluation is used to measure the naturalness and coherence of the generated text. The perplexity is used as a quantitative index, and its calculation formula is:

[0104]

[0105] Among them: N is the number of words in the text y k ; P(y i |y1,..., y i-1 ) is the prediction probability of the language model for the current word y i-1 based on the context y1,..., y i .

[0106] The lower the perplexity, the better the fluency of the generated text.

[0107] Information coverage evaluation

[0108] The information coverage evaluation measures the semantic relevance between the generated text and the input question. We use a pre-trained language model (such as BERT) to extract the vector representations of the input question q and the generated text yk, and then calculate the cosine similarity between the two:

[0109]

[0110] Among them: q and y k are the embedding vectors of the input question and the generated text respectively; ||q|| and ||y k || are the L2 norms of the corresponding vectors.

[0111] The closer the cosine similarity value is to 1, the higher the semantic relevance between the generated text and the input question.

[0112] Relevance evaluation

[0113] Relevance assessment is used to measure the content matching degree between the generated text and the input question or context. A TF-IDF-based method can be adopted for calculation, and its basic formula is:

[0114]

[0115] where: q i and y ki are respectively the terms in the input question and the generated text; TF-IDF(q i ) and TF-IDF(y ki ) are the TF-IDF weights of the corresponding terms.

[0116] A higher relevance score means that the generated text is more content-matched with the input question.

[0117] Final output: After the iteration termination condition is met, the system outputs the finally generated positive text. For example, it rewrites "I will never succeed" as "Through phased goals, I can make breakthroughs step by step", achieving the purpose of positive guidance and psychological counseling in terms of emotion.

[0118] Software and model training can rely on deep learning frameworks such as TensorFlow or PyTorch to implement pre-trained language models and Adapter fine-tuning techniques. The training data includes a large number of real negative texts and positively rewritten texts annotated by experts, ensuring that the model can learn multi-strategy expressions and the dynamic iterative optimization process.

[0119] The system can run based on common servers or cloud platforms, relying on GPU acceleration to achieve efficient computing of large-scale pre-trained models and dynamic weight generation modules. A client / server collaborative architecture can be adopted to support the efficient operation and real-time data interaction of the system in actual deployment, which is suitable for online application scenarios.

[0120] The present invention combines pre-trained language models, Adapter fine-tuning, HyperNetwork dynamic weight allocation, and hybrid cognitive dynamic decision-making, and realizes the multi-strategy collaboration and dynamic optimization problems that are difficult to solve by traditional methods through a systematic integration and alternating iteration mechanism. The dynamic policy weight allocation mechanism uses HyperNetwork to automatically generate the weights of each intervention strategy according to the input text and context features, realizing dynamic and real-time policy adjustment. The independent fine-tuning and parallel generation of the Adapter policy module use Adapter technology (such as LoRA) to achieve the independent expression of multiple psychological intervention strategies on the basis of the pre-trained model, and parallelly generate candidate texts, providing diverse inputs for subsequent weighted fusion. The weighted fusion of candidate texts weights and fuses the candidate texts generated by each strategy at the Token or sentence level to ensure that the final output text is superior to the single-strategy generation method in terms of emotional guidance, semantic coherence, and natural expression. The hybrid cognitive dynamic decision-making iteration mechanism introduces an alternating iteration process of thinking (T) and generation (G), and continuously self-reflects and corrects the generated text in combination with a preset evaluation function, significantly improving the overall quality and positive effect of the output text.

[0121] Advantages of the present invention compared with the prior art:

[0122] Traditional rewriting systems usually rely on fixed rules or single-generation models, lack dynamic weight allocation and multi-strategy collaborative generation capabilities, and do not have a self-correction mechanism after generation, which easily leads to insufficiently delicate emotional expressions or semantic conflicts in the output text. The present invention precisely addresses this problem. Through the dynamic generation of policy weights by HyperNetwork, parallel generation of multiple strategies by Adapter, weighted fusion of candidate texts, and a hybrid cognitive dynamic decision-making mechanism, it realizes the accurate transformation of typical negative texts. For example, it rewrites "I will never succeed" as "By setting phased goals, I can gradually break through", thus ensuring that the output text has significant positive emotional guidance effects while being semantically coherent and logically clear.

[0123] The application scenarios of the present invention include but are not limited to:

[0124] 1. Intelligent customer service and customer support systems

[0125] The multi-strategy emotional intervention method of the present invention can be widely applied in intelligent customer service and customer support systems. In these systems, customers often express dissatisfaction, anxiety, or anger. By applying the technology of the present invention, the system can effectively transform the negative emotions of customers and provide more gentle and positive feedback, thereby enhancing the customer experience and satisfaction.

[0126] 2. Mental health and psychotherapy

[0127] The present invention can be used as an effective tool in the field of mental health, especially in psychotherapy and online psychological counseling platforms. By transforming negative emotions into positive emotions and providing psychological counseling and support to users through personalized intervention strategies, the system can help people alleviate psychological problems such as anxiety, depression and stress, thereby promoting mental health.

[0128] 3. Education and study guidance

[0129] The present invention can be used in online learning platforms, educational counseling and teacher support tools. When students face negative emotions such as learning pressure and test anxiety, the present invention can provide positive emotional guidance, help students maintain a positive learning attitude, improve their learning efficiency and self-confidence, and thus promote their academic performance.

[0130] 4. Social media and public opinion management

[0131] The present invention can be applied in social media platforms and public opinion management systems, especially when dealing with a large number of user comments, posts and interactions. By automatically analyzing and converting negative emotional information, the platform can ensure the harmony of the public opinion environment, avoid excessive negative speech and emotional spread, and improve the quality of interaction among platform users.

[0132] 5. Advertising and Marketing In the field of advertising and marketing

[0133] The present invention can help enterprises adopt positive emotional strategies in their interactions with customers and promote brand image and emotional marketing. For example, enterprises can transform negative feedback or customer dissatisfaction into positive promotion of the brand, thereby improving customer loyalty and brand identity.

[0134] 6. Social services and crisis intervention

[0135] The present invention can be applied in the field of social services, especially in crisis intervention and emergency assistance services. When users show emotions such as sadness, anger or despair, the system can provide timely emotional support and guidance through emotional intervention strategies to help users overcome emotional crises, thereby promoting social stability and public safety.

[0136] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0137] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0138] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0139] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a ferromagnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memories.

[0140] In several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed with each other can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical, or other forms.

[0141] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0142] In addition, each functional unit in the embodiments of the present invention can be fully integrated in a processing unit, or each unit can be separately used as a unit, or two or more units can be integrated in one unit; the above integrated units can be implemented in the form of hardware, or in the form of hardware plus software functional units.

[0143] Those of ordinary skill in the art can understand that all or part of the steps for implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including those of the above method embodiments; and the foregoing storage medium includes: various media that can store program codes, such as removable storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0144] Alternatively, if the above integrated units of the present invention are implemented in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solutions of the embodiments of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. And the foregoing storage medium includes: various media that can store program codes, such as removable storage devices, ROM, RAM, magnetic disks, or optical discs.

[0145] The methods disclosed in several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0146] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0147] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0148] The above content is a further detailed description of the present invention in combination with specific preferred implementation manners. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those skilled in the technical field to which the present invention belongs, without departing from the concept of the present invention, several equivalent substitutions or obvious variations can be made, and if the performance or use is the same, they should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model, characterized in that, It includes the following steps: S1. Input processing: Receive negative text and situational features, and use a pre-trained language model to vectorize the text to generate a text vector representation. S2. Dynamic weight generation: Through the HyperNetwork module, based on the situational features, use a multi-layer perceptron to generate a policy weight vector and perform normalization processing to obtain a dynamic policy weight distribution. S3. Multi-strategy text generation: Based on the dynamic policy weights, use multiple independently fine-tuned Adapter policy adaptation units to fine-tune the pre-trained model and independently generate candidate texts corresponding to multiple different positive psychological intervention strategies. S4. Text fusion: According to the dynamic policy weights, perform weighted fusion on the candidate texts to form a rewritten text.

2. The method according to claim 1, wherein It also includes the following steps: S5. Hybrid cognitive decision-making: Through the hybrid cognitive dynamic decision-making module, adopt an "alternating iteration mechanism of thinking - generation", and self-evaluate and feedback-optimize the rewritten text according to a preset evaluation function, including adjusting the policy weights or correcting local text, until a final rewritten text that meets the preset criteria is generated.

3. The method according to claim 1, wherein Step S1 specifically includes: S11. Receive the negative text and situational features input by the user, and the situational features include emotion labels, dialogue context information, and user portraits. S12. Parse and vectorize and encode the negative text and situational features.

4. The method according to claim 3, wherein The situational features include emotion labels, dialogue context information, and user portraits.

5. The method according to claim 1, wherein Step S2 specifically includes: S21. Extract emotional features, context correlation features, and user features from the input text vector through the HyperNetwork. S22. Based on the extracted multi-dimensional features, predict the initial weight values of each intervention strategy. S23. Perform normalization processing on the initial weight values to generate dynamic policy weights that meet the requirements of the probability distribution.

6. The method according to claim 5, wherein The Softmax function is used to perform normalization processing on the initial weight values.

7. The method according to claim 1, wherein Step S3 specifically includes: S31. Configure an independent Adapter fine-tuning module for each preset psychological intervention strategy, and each module performs efficient parameter updates based on the shared pre-trained language model. S32. Each Adapter module independently generates a candidate rewritten text corresponding to the intervention strategy based on the input text vector representation. S33. According to the dynamic policy weights, perform weighted fusion calculation on the text generation probabilities output by each Adapter module.

8. The method according to claim 2, characterized in that Step S5 specifically includes: S51. Establish an initial thinking state based on the preliminary rewritten text. S52. Alternate execution of thinking - generation: In the thinking stage, perform multi-dimensional evaluation on the current text and update the thinking state; in the generation stage, generate an optimized text based on the updated thinking state. S53. Determine whether the iteration termination condition is met through a preset evaluation mechanism. If not, repeat the execution of S52.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model according to any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the multi-strategy emotion regulation and hybrid cognitive decision-making method based on a pre-trained model according to any one of claims 1 to 8.

Images