Active listening emotion support reply generation method and system facing implicit demand perception

Abstract

The invention discloses an active listening emotion support reply generation method and system oriented to implicit demand perception, relates to the technical field of natural language processing, and aims to solve the problem that an existing method usually only pays attention to identification and pacification of dominant emotions of a user and lacks in-depth understanding and active response ability to implicit demands. In order to solve the problem of difficulty in obtaining accurate emotion support reply due to the fact that the emotion support reply is difficult to obtain, the invention applies to gradually mine deep psychological causes of a user and extract potential demand information by performing multi-level reasoning on emotion expression and semantic clues input by the user. The overall framework comprises a listening stage and a suggestion stage, the listening stage guides a user to more truly express emotion and intention through an active listening strategy, and the suggestion stage generates targeted and common emotion support reply according to the identified implicit demand. According to the method, dynamic processing from surface emotion to deep demand is realized in combination with an inference chain mechanism, and the method has continuous emotional sharing ability and personalized response characteristics. And accurate emotion support reply can be obtained.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, specifically to a method and system for generating proactive listening emotion support responses based on implicit needs perception. Background Technology

[0002] In recent years, conversational emotion support has become a research hotspot in the field of natural language processing. Existing work largely focuses on passive emotion support scenarios, assuming users will actively express their emotional distress, with the model identifying the emotion and generating a response accordingly. While this approach has been applied in some psychological counseling and companionship conversations, it heavily relies on the user's willingness to proactively disclose information. However, in real-world interactions, nearly half of potential help seekers, due to limited self-awareness or reluctance to actively seek help, often use vague or semantically ambiguous language to mask their true emotions. This leads to a significant discrepancy between their surface expression and their underlying emotional needs, making it difficult for the model to capture the core psychological state and causing missed opportunities for effective support.

[0003] In summary, existing methods often focus only on identifying and soothing users' explicit emotions, lacking a deep understanding of implicit needs and the ability to proactively respond. Therefore, it is difficult to obtain accurate emotional support responses. Summary of the Invention

[0004] The purpose of this invention is to address the problem that existing methods often focus only on identifying and soothing users' explicit emotions, lacking a deep understanding of implicit needs and the ability to proactively respond. Consequently, accurate emotional support responses are difficult to obtain. This invention provides a method and system for generating proactive listening-based emotional support responses that are oriented towards the perception of implicit needs.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] A proactive listening and emotion support response generation method based on implicit needs perception includes the following steps:

[0007] Step 1: Construct a user profile covering three dimensions: feelings, needs, and memories;

[0008] Step 2: Based on the user profile, generate dialogues, segment the generated dialogues, and construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples;

[0009] Step 3: Train a large language model using training samples to obtain a trained large language model, i.e., a support model;

[0010] Step 4: Receive the natural language text containing implicit needs from the user, and input the natural language text into the support model to obtain the emotional support response output by the model.

[0011] Furthermore, the specific steps of step 1 are as follows:

[0012] Step 11: Obtain typical social situations in the context of emotional support, and construct multiple initial demand seeds based on each type of social situation, denoted as the seed set;

[0013] Step 12: Using each initial requirement seed in the seed set as an example, drive the large language model to generate user profile requirement candidates, and deduplicate all user profile requirement candidates to obtain the user profile requirement set.

[0014] Step 13: For each candidate user profile requirement in the user profile requirement set, generate a directed chain with a length not exceeding 3. Among them, the starting point It indicates explicit feelings coupled with emotions. This indicates the true potential demand after contextual validation;

[0015] Step 14: Filter all directed chains and retain only those chains that meet the semantic relevance threshold;

[0016] Step 15: Randomly obtain emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. and by the end node of the chain As a dimension of user profiling;

[0017] Step 16: Based on the directed chain, use the large model to perform inverse completion and... Related memory events are used to construct the memory dimensions of user profiles;

[0018] Step 17: Use the user profile's perception dimension, user profile's needs dimension, and user profile's memory dimension to obtain the user profile.

[0019] Furthermore, in step 11, 5-8 initial demand seeds are constructed based on each type of social context.

[0020] Furthermore, the specific steps of step 2 are as follows:

[0021] Step 21: Based on the user profile obtained in Step 1, use a large model to simulate users and supporters to generate listening stage dialogues. When generating listening stage dialogues, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies.

[0022] Step 22: Extract summaries from the listening phase dialogue, and use a large model to simulate users and supporters based on the extracted summaries to generate the suggestion phase dialogue;

[0023] Step 23: Combine the listening stage dialogue and the suggestion stage dialogue to obtain the completed support dialogue, and automatically score it according to the Active-Empathic Listening Scale. At the same time, check the number of dialogue rounds and eliminate dialogues with a score of less than 4 or fewer than 18 rounds to obtain the final supporter dialogue.

[0024] Step 24: Utilize a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases The dialogue is labeled and segmented to construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples.

[0025] Furthermore, the support model is based on LLAMA-3-8B-INSTRUCT and is obtained by minimizing the joint supervision loss;

[0026] The joint supervision loss is expressed as:

[0027]

[0028] in, For the input parameters of the large model, This is the training sample library.

[0029] An active listening and emotion support response generation system for implicit needs perception, the system includes a user profile construction module, a dialogue generation module, a model training module, and an emotion support response generation module;

[0030] The user profile building module is used to build user profiles covering three dimensions: feelings, needs, and memories.

[0031] The dialogue generation module generates dialogues based on user profiles, segments the generated dialogues, and constructs a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples;

[0032] The model training module uses training samples to train a large language model, resulting in a trained large language model, i.e., a support model.

[0033] The emotion support response generation module is used to receive natural language text containing implicit needs input by the user, input the natural language text into the support model, and obtain the emotion support response output by the model.

[0034] Furthermore, the user profile building module specifically performs the following steps:

[0035] Step 11: Obtain typical social situations in the context of emotional support, and construct multiple initial demand seeds based on each type of social situation, denoted as the seed set;

[0036] Step 12: Using each initial requirement seed in the seed set as an example, drive the large language model to generate user profile requirement candidates, and deduplicate all user profile requirement candidates to obtain the user profile requirement set.

[0037] Step 13: For each candidate user profile requirement in the user profile requirement set, generate a directed chain with a length not exceeding 3. Among them, the starting point It indicates explicit feelings coupled with emotions. This indicates the true potential demand after contextual validation;

[0038] Step 14: Filter all directed chains and retain only those chains that meet the semantic relevance threshold;

[0039] Step 15: Randomly obtain emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. and by the end node of the chain As a dimension of user profiling;

[0040] Step 16: Based on the directed chain, use the large model to perform inverse completion and... Related memory events are used to construct the memory dimensions of user profiles;

[0041] Step 17: Use the user profile's perception dimension, user profile's needs dimension, and user profile's memory dimension to obtain the user profile.

[0042] Furthermore, in step 11, 5-8 initial demand seeds are constructed based on each type of social context.

[0043] Furthermore, the dialogue generation module specifically performs the following steps:

[0044] Step 21: Based on the user profile obtained in Step 1, use a large model to simulate users and supporters to generate listening stage dialogues. When generating listening stage dialogues, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies.

[0045] Step 22: Extract summaries from the listening phase dialogue, and use a large model to simulate users and supporters based on the extracted summaries to generate the suggestion phase dialogue;

[0046] Step 23: Combine the listening stage dialogue and the suggestion stage dialogue to obtain the completed support dialogue, and automatically score it according to the Active-Empathic Listening Scale. At the same time, check the number of dialogue rounds and eliminate dialogues with a score of less than 4 or fewer than 18 rounds to obtain the final supporter dialogue.

[0047] Step 24: Utilize a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases The dialogue is labeled and segmented to construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples.

[0048] Furthermore, the support model is based on LLAMA-3-8B-INSTRUCT and is obtained by minimizing the joint supervision loss;

[0049] The joint supervision loss is expressed as:

[0050]

[0051] in, For the input parameters of the large model, This is the training sample library.

[0052] The beneficial effects of this invention are:

[0053] This application employs multi-level reasoning based on user input of emotional expressions and semantic cues to gradually uncover underlying psychological motivations and extract potential needs. The overall framework comprises a "listening stage" and a "suggestion stage." The former guides users to express their emotions and intentions more authentically through proactive listening strategies, while the latter generates targeted and empathetic emotional support responses based on identified implicit needs. This application combines a reasoning chain mechanism to achieve dynamic processing from surface emotions to deep-seated needs, possessing continuous empathy and personalized response characteristics. This results in accurate emotional support responses. Attached Figure Description

[0054] Figure 1 A diagram illustrating the impact of removing the needs and memory components from user profiles on user simulation effectiveness across roles, degree of anthropomorphism, and psychological complexity.

[0055] Figure 2 This is the overall flowchart of this application. Detailed Implementation

[0056] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0057] Specific Implementation Method 1: The proactive listening emotion support response generation method based on implicit need perception described in this implementation method includes the following steps:

[0058] Step 1: Based on psychological theories, construct a user profile covering three dimensions: feelings, needs, and memories;

[0059] Step 2: Based on the user profile, generate dialogues, segment the generated dialogues, and construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples;

[0060] Step 3: Train a large language model using training samples to obtain a trained large language model, i.e., a support model;

[0061] Step 4: Receive the natural language text containing implicit needs from the user, and input the natural language text into the support model to obtain the emotional support response output by the model.

[0062] This invention can be trained using eight NVIDIA A100 80GB GPUs under the PyTorch framework and can be deployed in a GPU-accelerated inference environment, with wide applications in scenarios such as mental health counseling, emotional support, and intelligent interaction.

[0063] Specific Implementation Method Two: This implementation method is a further explanation of Specific Implementation Method One. The difference between this implementation method and Specific Implementation Method One is that the specific steps of step 1 are as follows:

[0064] Step 11: Obtain typical social situations in the context of emotional support, and construct multiple initial demand seeds based on each type of social situation, denoted as the seed set;

[0065] Step 12: Using each initial requirement seed in the seed set as an example, drive the large language model to generate user profile requirement candidates, and deduplicate all user profile requirement candidates to obtain the user profile requirement set.

[0066] Step 13: For each candidate user profile requirement in the user profile requirement set, generate a directed chain with a length not exceeding 3. Among them, the starting point It indicates explicit feelings coupled with emotions. This indicates the true potential demand after contextual validation;

[0067] Step 14: Filter all directed chains and retain only those chains that meet the semantic relevance threshold;

[0068] Step 15: Randomly obtain emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. and by the end node of the chain As a dimension of user profiling;

[0069] Step 16: Based on the directed chain, use the large model to perform inverse completion and... Related memory events are used to construct the memory dimensions of user profiles;

[0070] Step 17: Use the user profile's perception dimension, user profile's needs dimension, and user profile's memory dimension to obtain the user profile.

[0071] Specific Implementation Method 3: This implementation method is a further explanation of Specific Implementation Method 2. The difference between this implementation method and Specific Implementation Method 2 is that in step 11, 5-8 initial demand seeds are constructed based on each type of social situation.

[0072] Specific Implementation Method Four: This implementation method is a further explanation of Specific Implementation Method Three. The difference between this implementation method and Specific Implementation Method Three is that the specific steps of step 2 are as follows:

[0073] Step 21: Based on the user profile obtained in Step 1, use a large model to simulate users and supporters to generate listening stage dialogues. When generating listening stage dialogues, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies.

[0074] Step 22: Extract summaries from the listening phase dialogue, and use a large model to simulate users and supporters based on the extracted summaries to generate the suggestion phase dialogue;

[0075] Step 23: Combine the listening stage dialogue and the suggestion stage dialogue to obtain the completed support dialogue, and automatically score it according to the Active-Empathic Listening Scale. At the same time, check the number of dialogue rounds and eliminate dialogues with a score of less than 4 or fewer than 18 rounds to obtain the final supporter dialogue.

[0076] Step 24: Utilize a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases The dialogue is labeled and segmented to construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples.

[0077] Specific Implementation Method 5: This implementation method is a further explanation of Specific Implementation Method 4. The difference between this implementation method and Specific Implementation Method 4 is that the support model is based on LLAMA-3-8B-INSTRUCT and is obtained by minimizing the joint supervision loss.

[0078] The joint supervision loss is expressed as:

[0079]

[0080] in, For the input parameters of the large model, This is the training sample library.

[0081] Specific Implementation Method Six: The Active Listening Emotional Support Response Generation System for Implicit Needs Perception described in this implementation method includes a user profile construction module, a dialogue generation module, a model training module, and an emotional support response generation module.

[0082] The user profile building module is used to construct user profiles covering three dimensions: feelings, needs, and memories, based on psychological theories.

[0083] The dialogue generation module generates dialogues based on user profiles, segments the generated dialogues, and constructs a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples;

[0084] The model training module uses training samples to train a large language model, resulting in a trained large language model, i.e., a support model.

[0085] The emotion support response generation module is used to receive natural language text containing implicit needs input by the user, input the natural language text into the support model, and obtain the emotion support response output by the model.

[0086] Specific Implementation Method Seven: This implementation method is a further explanation of Specific Implementation Method Six. The difference between this implementation method and Specific Implementation Method Six is ​​that the user profile construction module specifically performs the following steps:

[0087] Step 11: Obtain typical social situations in the context of emotional support, and construct multiple initial demand seeds based on each type of social situation, denoted as the seed set;

[0088] Step 12: Using each initial requirement seed in the seed set as an example, drive the large language model to generate user profile requirement candidates, and deduplicate all user profile requirement candidates to obtain the user profile requirement set.

[0089] Step 13: For each candidate user profile requirement in the user profile requirement set, generate a directed chain with a length not exceeding 3. Among them, the starting point It indicates explicit feelings coupled with emotions. This indicates the true potential demand after contextual validation;

[0090] Step 14: Filter all directed chains and retain only those chains that meet the semantic relevance threshold;

[0091] Step 15: Randomly obtain emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. and by the end node of the chain As a dimension of user profiling;

[0092] Step 16: Based on the directed chain, use the large model to perform inverse completion and... Related memory events are used to construct the memory dimensions of user profiles;

[0093] Step 17: Use the user profile's perception dimension, user profile's needs dimension, and user profile's memory dimension to obtain the user profile.

[0094] Specific Implementation Method Eight: This implementation method is a further explanation of Specific Implementation Method Seven. The difference between this implementation method and Specific Implementation Method Seven is that in step 11, 5-8 initial demand seeds are constructed based on each type of social situation.

[0095] Specific Implementation Method Nine: This implementation method is a further explanation of Specific Implementation Method Eight. The difference between this implementation method and Specific Implementation Method Eight is that the dialogue generation module specifically performs the following steps:

[0096] Step 21: Based on the user profile obtained in Step 1, use a large model to simulate users and supporters to generate listening stage dialogues. When generating listening stage dialogues, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies.

[0097] Step 22: Extract summaries from the listening phase dialogue, and use a large model to simulate users and supporters based on the extracted summaries to generate the suggestion phase dialogue;

[0098] Step 23: Combine the listening stage dialogue and the suggestion stage dialogue to obtain the completed support dialogue, and automatically score it according to the Active-Empathic Listening Scale. At the same time, check the number of dialogue rounds and eliminate dialogues with a score of less than 4 or fewer than 18 rounds to obtain the final supporter dialogue.

[0099] Step 24: Utilize a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases The dialogue is labeled and segmented to construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples.

[0100] Specific Implementation Method Ten: This implementation method is a further explanation of Specific Implementation Method Nine. The difference between this implementation method and Specific Implementation Method Nine is that the support model is based on LLAMA-3-8B-INSTRUCT and is obtained by minimizing the joint supervision loss.

[0101] The joint supervision loss is expressed as:

[0102]

[0103] in, For the input parameters of the large model, This is the training sample library.

[0104] Step 1: Based on psychological theories, construct a user profile covering three dimensions: feelings, needs, and memories. Use multi-dimensional modeling to simulate real emotional interaction scenarios to ensure that the profile has psychological consistency and dynamic response characteristics.

[0105] Step Two: Employing a two-stage scripted dialogue generation mechanism. In the listening stage, reflective responses, emotional mirroring, and gentle guidance replace frequent questioning, gradually bringing users' implicit needs to the surface. In the suggestion stage, supportive suggestions tailored to the user's needs are generated based on a thorough understanding of those needs. Subsequently, an automated scoring mechanism is constructed based on the Active-Empathic Listening Scale to filter out low-quality samples. On this basis, fine-grained cognitive reasoning annotation is performed, precisely labeling the cognitive processes and strategic phrases of supporters in each round of dialogue, forming an intermediate "thinking-strategy" monitoring signal.

[0106] Step 3: Using LLAMA-3-8B-INSTRUCT as the base model, conduct supervised fine-tuning using the data constructed in Step 2 to enhance the model's comprehensive capabilities in sentiment understanding, policy generation, and empathy expression.

[0107] Step 4: Construct a user simulation environment based on the large model and evaluate the model obtained in Step 3. Using the success rate of demand discovery as the core evaluation metric, conduct a multi-dimensional evaluation combining general performance indicators and psychological scales to comprehensively measure the model's overall performance in emotional support tasks.

[0108] First, in the user persona construction phase, this application focuses on evaluating the role of the needs module and the memory module in improving the user simulation effect. Specifically, a standard persona containing complete needs and memory modules is constructed, and one or two components are removed to form a control group persona. The persona is then compared and evaluated on three core dimensions: 1) Role Consistency, which measures whether the simulated user can maintain a coherent, reasonable, and stable persona in multiple rounds of dialogue; 2) Human-likeness, which assesses whether the simulated user possesses naturalness and rationality in language expression and behavioral logic similar to real users; and 3) Psychological Complexity, which judges whether the model can present multi-layered, implicit emotional states and accurately reflect the psychological characteristics of users indirectly expressing their needs. During the evaluation process, this application combines human expert annotation with automated model evaluation (using GPT-4o) to ensure the objectivity and consistency of the evaluation results. Experimental results show that after adding the demand and memory modules, the user simulation effect was significantly improved in all three dimensions, especially in the complex emotional logic caused by the unmet implicit needs, which provided a user base that closely resembled real-life scenarios for the subsequent generation of high-quality dialogues.

[0109] Table 1 Data Quality Assessment Results

[0110]

[0111] Table 1 shows the process of evaluating six basic indicators through a combination of manual assessment and the GPT-4o model. At the same time, GPT-4o is used to evaluate two dimensions of the comfort scale and two dimensions of the conversational support scale.

[0112] Based on the aforementioned high-fidelity user profile, this application further synthesizes emotional support dialogue data and systematically evaluates its quality. This application compares the constructed COCOON dataset with current mainstream emotional support dialogue datasets, with evaluation indicators divided into two main categories. The first category consists of basic dialogue quality indicators, including fluency, diversity, empathy, information content, humanoid behavior, and skillfulness. Fluency measures the naturalness and logical coherence of language expression; diversity reflects the richness of language and content; empathy assesses the accuracy of understanding and responding to user emotions; information content examines the relevance and sufficiency of suggestions; humanoid behavior judges whether dialogue behavior conforms to human interaction habits; and skillfulness reflects the model's comprehensive ability in empathic expression, positive guidance, and suggestion provision. The second category comprises professional evaluation scales designed based on psychological theories, including the ComfortScale and the Relationship Supportive Conversation Scale (RACScale). The former measures the effectiveness of emotional improvement and the user's negative feedback to the assistant, while the latter measures the supportiveness of the conversation and the overall conversation management ability. Evaluators include human experts with psychology backgrounds and automated scoring models to ensure the consistency and reproducibility of evaluation criteria across different scenarios. The evaluation results show that the COCOON dataset outperforms existing mainstream datasets on most metrics, especially in empathy and emotional improvement, demonstrating its superior ability to guide users to express their underlying emotions in proactive emotional support scenarios.

[0113] Table 2. Evaluation results of emotional support capabilities of different models

[0114]

[0115] Finally, this application trained the emotion support dialogue model COCOON-Llama3 on the COCOON dataset and verified its performance in proactive emotion support tasks. This application compared this model with models trained on other traditional datasets, and also constructed an ablation test version without fine-grained annotation information to evaluate the impact of intermediate inference processes and policy annotations on model performance. Building upon the original evaluation dimensions, this application further introduced a success rate (SR) metric to measure the model's ability to identify and respond to users' potential emotional needs. Evaluation results show that COCOON-Llama3 significantly outperforms all compared models in the success rate metric and maintains a leading position in multiple core dimensions such as fluency, empathy, and dialogue support. This indicates that the model can effectively identify the true needs hidden in users' expressions, guide them to gradually reveal their emotions, and provide targeted support. Furthermore, the model with fine-grained annotations outperformed its ablation test version in all metrics, further validating the significant role of inference chains and policy guidance in improving the quality of proactive emotion support.

[0116] Furthermore, the specific steps of step one are as follows:

[0117] Step 11: Obtaining typical social situations within an emotionally supportive context For each scenario, construct 5–8 high-quality initial requirement seeds, denoted as the seed set.

[0118] Steps one and two: In the same type of social context Below, with Using the seed in the example, the large language model is driven to generate user profile requirement candidates; all candidates are deduplicated to obtain the user profile requirement set used for subsequent inference chain construction. ;

[0119] Step 13: Targeting candidate user profile requirements within the set Generate directed chains of length no more than 3. ( ), where the starting point Indicates overt feelings coupled with emotions, endpoint This indicates the actual potential demand after contextual validation.

[0120] Step 14: Filter all directed chains, retaining only those that meet the semantic relevance threshold. The chain.

[0121] Step 15: Randomly Obtain Emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. This ensures the semantic anchor and psychological authenticity. (Based on the end node of the chain.) As a dimension of user profile requirements, it serves as an alignment target when generating subsequent suggestions.

[0122] Step 16: Based on the directed chain, use the large model for inverse completion and... Related memory events are used to construct the memory dimensions of user profiles;

[0123] Step 17: Utilize the user profile's perceptual dimension, user profile's demand dimension, and user profile's memory dimension to obtain the user profile.

[0124] Furthermore, the specific steps of step two are as follows:

[0125] Step 21: Based on the user profile obtained in Step 1, use a large model to simulate the user and supporter to generate a listening stage dialogue. When generating the listening stage dialogue, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies.

[0126] Step 22: Based on the dialogue in the listening phase, extract summaries and use a large model to simulate users and supporters to generate dialogue in the suggestion phase;

[0127] Steps 2 and 3: Combine the listening and suggestion phase dialogues to obtain a complete supportive dialogue. Automatically score the dialogue using the Active-Empathic Listening Scale, while also checking the number of dialogue rounds. Discard low-quality samples scoring less than 4 points and dialogues with fewer than 18 rounds. This yields the final supporter dialogue.

[0128] Step Two Four: Utilizing a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases Add annotations.

[0129] The dialogue is segmented to construct training samples, which include the dialogue history. Thinking about the text Strategy phrases Next round of remarks This leads to a corpus containing the final supporter dialogue after multiple annotations. .

[0130] Furthermore, the specific steps of step three are as follows:

[0131] Using LLAMA-3-8B-INSTRUCT as the base model, the support model is obtained by minimizing the joint supervision loss.

[0132]

[0133] The model input is the dialogue history. The output is "Thinking Text" +Strategy phrase +Next round of remarks The goal of joint sequence training is to enable the model to learn the cognitive processes of supporters and their final responses simultaneously during the unified generation process. Input parameters for a large model;

[0134] It receives natural language text containing implicit needs from user input and feeds it into a trained support model to obtain emotional support responses.

[0135] Furthermore, the specific steps of step four are as follows:

[0136] Step 41: Use a large model to act as a user and create a four-tuple profile for each user. Injecting and locking in the persona and expression style to form a simulated user agent. The initial state.

[0137] Step 42: Using the support model COCOON-Llama trained in Step 3, for each user profile Execute multiple rounds of scripted interactions to generate a session sequence. ;in and They represent the first The wheel's users and supporters responded.

[0138] Step 43: For each dialogue segment Call the evaluation model to calculate the comprehensive result. ;in This indicates whether the demand discovery was successful. This is the overall quality score for the dialogue. The final average score is obtained by averaging all dialogues.

[0139] This application is applicable to the task of generating proactive listening-based emotional support based on the perception of implicit needs. The model uses multi-level reasoning on the emotional expressions and semantic cues inputted by the user to gradually uncover their deep psychological motivations and extract potential needs. The overall framework includes a "listening stage" and a "suggestion stage." The former guides users to express their emotions and intentions more authentically through proactive listening strategies, while the latter generates targeted and empathetic emotional support responses based on the identified implicit needs. This method combines an inference chain mechanism to achieve dynamic modeling from surface emotions to deep needs, enabling the model to possess continuous empathy and personalized response characteristics. This application can be trained using eight NVIDIA A100 80GB GPUs under the PyTorch framework and deployed in a GPU-accelerated inference environment, with wide applications in scenarios such as mental health counseling, emotional companionship, and intelligent interaction.

[0140] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solution of the present invention and should not be used to limit the scope of protection. Any modifications made in accordance with the claims and specification of the present invention that are only partial should still fall within the protection scope of the present invention.

Claims

1. A method for generating emotionally supportive responses through proactive listening, oriented towards implicit needs perception, characterized in that... Includes the following steps: Step 1: Construct a user profile covering three dimensions: feelings, needs, and memories; Step 2: Based on the user profile, generate dialogues, segment the generated dialogues, and construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples; Step 3: Train a large language model using training samples to obtain a trained large language model, i.e., a support model; Step 4: Receive the natural language text containing implicit needs from the user, and input the natural language text into the support model to obtain the emotional support response output by the model.

2. The method for generating proactive listening emotion support responses based on implicit needs perception as described in claim 1, characterized in that... The specific steps of step 1 are as follows: Step 11: Obtain typical social situations in the context of emotional support, and construct multiple initial demand seeds based on each type of social situation, denoted as the seed set; Step 12: Using each initial requirement seed in the seed set as an example, drive the large language model to generate user profile requirement candidates, and deduplicate all user profile requirement candidates to obtain the user profile requirement set. Step 13: For each candidate user profile requirement in the user profile requirement set, generate a directed chain with a length not exceeding 3. , Among them, the starting point It indicates explicit feelings coupled with emotions. This indicates the true potential demand after contextual validation; Step 14: Filter all directed chains and retain only those chains that meet the semantic relevance threshold; Step 15: Randomly obtain emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. and by the end node of the chain As a dimension of user profiling; Step 16: Based on the directed chain, use the large model to perform inverse completion and... Related memory events are used to construct the memory dimensions of user profiles; Step 17: Use the user profile's perception dimension, user profile's needs dimension, and user profile's memory dimension to obtain the user profile.

3. The method for generating proactive listening emotion support responses based on implicit needs perception as described in claim 2, characterized in that... In step 11, 5-8 initial demand seeds are constructed based on each type of social situation.

4. The method for generating emotional support responses through proactive listening based on implicit needs perception, as described in claim 3, is characterized in that... The specific steps of step 2 are as follows: Step 21: Based on the user profile obtained in Step 1, use a large model to simulate users and supporters to generate listening stage dialogues. When generating listening stage dialogues, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies. Step 22: Extract summaries from the listening phase dialogue, and use a large model to simulate users and supporters based on the extracted summaries to generate the suggestion phase dialogue; Step 23: Combine the listening stage dialogue and the suggestion stage dialogue to obtain the completed support dialogue, and automatically score it according to the Active-Empathic Listening Scale. At the same time, check the number of dialogue rounds and eliminate dialogues with a score of less than 4 or fewer than 18 rounds to obtain the final supporter dialogue. Step 24: Utilize a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases The dialogue is labeled and segmented to construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples.

5. The method for generating emotional support responses through proactive listening based on implicit needs perception, as described in claim 4, is characterized in that... The support model is based on LLAMA-3-8B-INSTRUCT and is obtained by minimizing the joint supervision loss. The joint supervision loss is expressed as: in, For the input parameters of the large model, This is the training sample library.

6. A proactive listening emotion support response generation system for implicit needs perception, characterized in that... The system includes a user profile building module, a dialogue generation module, a model training module, and an emotion support response generation module. The user profile building module is used to build user profiles covering three dimensions: feelings, needs, and memories. The dialogue generation module generates dialogues based on user profiles, segments the generated dialogues, and constructs a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples; The model training module uses training samples to train a large language model, resulting in a trained large language model, i.e., a support model. The emotion support response generation module is used to receive natural language text containing implicit needs input by the user, input the natural language text into the support model, and obtain the emotion support response output by the model.

7. The proactive listening emotion support response generation system for implicit need perception according to claim 6, characterized in that... The user profile building module specifically performs the following steps: Step 11: Obtain typical social situations in the context of emotional support, and construct multiple initial demand seeds based on each type of social situation, denoted as the seed set; Step 12: Using each initial requirement seed in the seed set as an example, drive the large language model to generate user profile requirement candidates, and deduplicate all user profile requirement candidates to obtain the user profile requirement set. Step 13: For each candidate user profile requirement in the user profile requirement set, generate a directed chain with a length not exceeding 3. , Among them, the starting point It indicates explicit feelings coupled with emotions. This indicates the true potential demand after contextual validation; Step 14: Filter all directed chains and retain only those chains that meet the semantic relevance threshold; Step 15: Randomly obtain emotions and emotions With the head node Combined, we obtain the perception dimensions used for user profiling. and by the end node of the chain As a dimension of user profiling; Step 16: Based on the directed chain, use the large model to perform inverse completion and... Related memory events are used to construct the memory dimensions of user profiles; Step 17: Use the user profile's perception dimension, user profile's needs dimension, and user profile's memory dimension to obtain the user profile.

8. The proactive listening emotion support response generation system for implicit need perception as described in claim 7, characterized in that... In step 11, 5-8 initial demand seeds are constructed based on each type of social situation.

9. The proactive listening emotion support response generation system for implicit need perception according to claim 8, characterized in that... The dialogue generation module specifically performs the following steps: Step 21: Based on the user profile obtained in Step 1, use a large model to simulate users and supporters to generate listening stage dialogues. When generating listening stage dialogues, based on the principle of active listening, reflective response, emotional mirroring and gentle guidance are used as generation strategies. Step 22: Extract summaries from the listening phase dialogue, and use a large model to simulate users and supporters based on the extracted summaries to generate the suggestion phase dialogue; Step 23: Combine the listening stage dialogue and the suggestion stage dialogue to obtain the completed support dialogue, and automatically score it according to the Active-Empathic Listening Scale. At the same time, check the number of dialogue rounds and eliminate dialogues with a score of less than 4 or fewer than 18 rounds to obtain the final supporter dialogue. Step 24: Utilize a large model to analyze the cognitive processes during the dialogue with the ultimate supporter. With strategy phrases The dialogue is labeled and segmented to construct a dialogue history. Thinking about the text Strategy phrases And the next round of discourse The training samples.

10. The proactive listening emotion support response generation system for implicit need perception according to claim 9, characterized in that... The support model is based on LLAMA-3-8B-INSTRUCT and is obtained by minimizing the joint supervision loss. The joint supervision loss is expressed as: in, For the input parameters of the large model, This is the training sample library.

Images