Empathic Intelligent Customer Service System Based on Multi-Agent and Explainable Distillation

Abstract

This invention relates to the fields of human-computer interaction and natural language processing technology, and discloses an empathic intelligent customer service system based on multi-agent and interpretable distillation, including: an input and memory module, an emotion determination module, an empathy judgment and attitude determination module, a knowledge retrieval module, a response generation module, and a log and data closed-loop module. During empathic customer service interaction, the emotion determination module identifies the user's emotional state from multiple dimensions, and combines an emotion type classification unit and an emotion intensity assessment unit to achieve accurate emotion analysis. This solves the problem of emotion misjudgment caused by single-modal analysis in traditional systems, ensuring the accuracy of empathic responses. Furthermore, by real-time detection of the consistency between emotion recognition results and the dialogue context, errors in empathic interaction are further reduced. Moreover, when the response strategy deviates, a dynamic adjustment mechanism can correct it in real time, ensuring the rationality and effectiveness of empathic decision-making.

Description

Technical Field

[0001] This invention relates to the fields of human-computer interaction and natural language processing, specifically to an empathic intelligent customer service system based on multi-agent and interpretable distillation. Background Technology

[0002] Human-computer interaction refers to the process of information exchange between humans and computers using a certain dialogue language and interactive methods to complete a specific task.

[0003] Currently, due to the diverse emotional expressions in customer service dialogues, the single-modal analysis model relied upon for user emotional state identification is insufficient in understanding complex emotional contexts and cannot detect the consistency between emotional identification results and dialogue context in real time. When emotional misjudgment occurs, it will cause directional errors in empathetic responses and cannot guarantee the accuracy of empathetic interaction.

[0004] Therefore, we propose an empathic intelligent customer service system based on multi-agent and interpretable distillation to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an empathic intelligent customer service system based on multi-agent and interpretable distillation, which solves the problem mentioned in the background technology of the inability to detect the consistency between emotion recognition results and dialogue context in real time.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an empathic intelligent customer service system based on multi-agent and interpretable distillation, comprising: The input and memory module uses the dialogue history acquisition unit to collect the user's current input and the history of the last N rounds of dialogue, and outputs structured dialogue memory data through the memory generation unit; The emotion determination module receives the structured dialogue memory data, identifies the user's emotion category through the emotion type classification unit, and outputs the emotion determination result using the emotion intensity assessment unit. The emotion categories include natural, happy, sad, angry, surprised, fearful, disgusted, frustrated, and excited, and the emotion intensity is a level from 1 to 5. The empathy judgment and attitude determination module receives the emotion judgment result and structured dialogue memory data, determines whether an empathy response is needed through the empathy decision unit, and outputs attitude tags using the attitude tag selection unit. The attitude tags include appeasement, apology, clarification, explanation, delay explanation, and price explanation. The knowledge retrieval module receives user input and the structured dialogue memory data, performs dual-channel retrieval through the character retrieval submodule and the semantic retrieval submodule, retrieves the Top-K candidate knowledge fragments from each, and performs weighted fusion using the deduplication and rearrangement unit to output the retrieved knowledge fragment; The response generation module receives the attitude tags and retrieved knowledge fragments, synthesizes the response content through the condition generation unit, and outputs the empathetic response through the strategy control unit; The log and data closed-loop module integrates end-to-end data, including user input, sentiment assessment results, attitude tags, retrieved knowledge fragments, and response content. Interaction information is stored through the log recording unit and used by the feedback collection unit for system optimization and data construction.

[0007] Preferably, the dialogue history acquisition unit of the input and memory module specifically includes a dialogue history caching subunit and a memory summary generation subunit. The dialogue history caching subunit is used to dynamically store the dialogue text of the last N rounds and update the historical data through a sliding window mechanism. The memory summary generation subunit uses a sequence-to-sequence model to compress the dialogue history, generate a summary memory representation, and output it to the sentiment determination module.

[0008] Preferably, the emotion type classification unit of the emotion determination module adopts a multi-label classification neural network model. The multi-label classification neural network model extracts the emotional keyword features in the dialogue text based on the attention mechanism and outputs a multi-dimensional emotion probability distribution. The emotion intensity assessment unit calculates the continuous value of emotion intensity through a regression model and discretizes it into levels from 1 to 5.

[0009] Preferably, the empathy decision-making unit of the empathy judgment and attitude determination module integrates a rule engine and a deep learning model. The rule engine triggers empathy decisions based on the emotional intensity threshold and historical empathy frequency. The deep learning model generates an empathy necessity score by fusing multimodal features and combining emotional type, intensity, and dialogue context. The attitude label selection unit selects the optimal label from a predefined attitude set based on the empathy necessity score.

[0010] Preferably, the character retrieval submodule of the knowledge retrieval module uses the BM25 algorithm for keyword matching, the semantic retrieval submodule uses a pre-trained embedding vector model to calculate the semantic similarity between the query and the knowledge base, and the deduplication and reordering unit integrates the character retrieval score and the semantic retrieval score through a weighted score fusion strategy, and sets an upper limit on the fragment length to control the output scale.

[0011] Preferably, the condition generation unit of the response generation module is based on the Transformer architecture. The condition generation unit uses attitude labels as prefix conditions and retrieves knowledge fragments as context constraints to generate response text through an autoregressive approach. The policy control unit introduces a reinforcement learning strategy to dynamically adjust the repetition penalty and diversity parameters in the generation process.

[0012] Preferably, the feedback collection unit of the log and data closed-loop module includes a user satisfaction rating interface and an error feedback labeling tool. The user satisfaction rating interface collects real-time rating data of users on responses, and the error feedback labeling tool allows administrators to manually correct and relabel the system output. The feedback collection unit sends the collected data to the data construction pipeline for incremental model training.

[0013] Preferably, the system further includes a data generation and filtering module, which synthesizes empathic dialogue data through a multi-agent simulated environment and introduces forward-looking information for the next round of questions during the response generation stage to enhance cross-round coherence. The data generation and filtering module integrates a reward model scoring mechanism to perform quality filtering on the synthesized data and retain highly consistent samples.

[0014] Preferably, the system further includes an interpretable knowledge distillation module, which organizes the output of the teacher model into structured thinking fragments, including emotional evidence, reasons for empathy judgment, basis for attitude selection, and knowledge fragment identifiers. The interpretable knowledge distillation module trains a lightweight student model through efficient parameter fine-tuning technology, enabling the student model to internalize the "think first, then respond" decision-making paradigm.

[0015] Preferably, the parameter fine-tuning technique of the interpretable knowledge distillation module adopts the LoRA method. The configurable parameters of the LoRA method include rank dimension, scaling factor, learning rate and training epochs, which are dynamically optimized according to the target hardware resources. The student model only needs to load adapter weights during the inference stage, reducing memory usage and inference latency.

[0016] Compared with existing technologies, this invention provides an empathic intelligent customer service system based on multi-agent and interpretable distillation, which has the following beneficial effects: 1. In this invention, when conducting empathic customer service interactions, the emotion judgment module identifies the user's emotional state from multiple dimensions. Combined with the emotion type classification unit and the emotion intensity assessment unit, accurate emotion analysis is achieved. This solves the problem of emotion misjudgment caused by single-modal analysis in traditional systems, ensuring the accuracy of empathic responses. At the same time, by detecting the consistency between the emotion recognition results and the dialogue context in real time, the error of empathic interaction is further reduced.

[0017] 2. In this invention, when making empathic decisions, the empathic judgment and attitude determination module integrates a rule engine and a deep learning model, combined with the domain knowledge support provided by the knowledge retrieval module, to generate attitude labels that match the user's questions in real time. This enables the system to reduce the disconnect between empathic responses and actual problems, and to correct deviations in response strategies in real time through a dynamic adjustment mechanism, ensuring the rationality and effectiveness of empathic decisions.

[0018] 3. In this invention, when generating empathic responses, the condition generation unit and strategy control unit of the response generation module, combined with the structured thinking fragments provided by the interpretable knowledge distillation module, enable multi-level and tiered empathic response generation. This allows for dynamic adjustment of response content based on the user's emotional state, solving the problem of lack of relevance and depth in traditional system responses, and improving the accuracy of empathic interaction and user experience. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of the empathic intelligent customer service system based on multi-agent and interpretable distillation of the present invention. Detailed Implementation

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

[0021] Please see Figure 1 This empathic intelligent customer service system, based on multi-agent and interpretable distillation, includes: The input and memory module uses the dialogue history acquisition unit to collect the user's current input and the history of the last N rounds of dialogue, and outputs structured dialogue memory data through the memory generation unit; The emotion determination module receives structured dialogue memory data, identifies the user's emotion category through the emotion type classification unit, and outputs the emotion determination result using the emotion intensity assessment unit. The emotion categories include natural, happy, sad, angry, surprised, fearful, disgusted, frustrated, and excited, and the emotion intensity is rated from 1 to 5. The empathy assessment and attitude determination module receives the emotion assessment results and structured dialogue memory data. It determines whether an empathic response is needed through the empathy decision-making unit and outputs attitude tags using the attitude tag selection unit. The attitude tags include appeasement, apology, clarification, explanation, delay explanation, and price explanation. The knowledge retrieval module receives user input and structured dialogue memory data. It performs dual-channel retrieval through the character retrieval submodule and the semantic retrieval submodule, respectively, and retrieves the Top-K candidate knowledge fragments. It then uses the deduplication and rearrangement unit to perform weighted fusion and outputs the retrieved knowledge fragments. The response generation module receives attitude tags and retrieved knowledge fragments, synthesizes response content through the condition generation unit, and outputs an empathetic response through the strategy control unit; The log and data closed-loop module integrates end-to-end data, including user input, sentiment assessment results, attitude tags, retrieved knowledge fragments, and response content. Interaction information is stored through the log recording unit and used by the feedback collection unit for system optimization and data construction.

[0022] The dialogue history acquisition unit of the input and memory module specifically includes a dialogue history caching subunit and a memory summary generation subunit. The dialogue history caching subunit is used to dynamically store the dialogue text of the last N rounds and update the historical data through a sliding window mechanism. The memory summary generation subunit uses a sequence-to-sequence model to compress the dialogue history, generate a summary memory representation, and output it to the sentiment determination module. The dialogue history cache subunit uses a queue data structure to implement the sliding window. Let the current dialogue sequence be represented in chronological order as follows: ,in Indicates the first The dialogue text is presented in rounds, with a fixed window size N. Initially, the window contains the most recent N rounds of dialogue, i.e., a subset. When a new round of dialogue When generated, the update mechanism is executed through the following steps: First, the new dialogue Add to the tail of the queue; Secondly, check if the queue length exceeds N: if the current length is greater than N, automatically remove the dialogue text from the head of the queue. This ensures that the window always contains the latest N rounds of data; This process can be described mathematically as follows: The updated window content is Its general form can be expressed as: ; in This represents the original window history set. The oldest element in the window. For new dialogue text; In actual code implementation, this is usually accomplished using loop pointers and dynamic arrays to optimize time complexity and ensure the efficiency of update operations; The core of the memory summarization generation subunit is a sequence-to-sequence model based on Transformer, and its specific implementation steps are as follows: Input representation: The dialogue history is used as the input sequence, and each round of dialogue is regarded as a sub-unit. The input text is first converted into a word vector sequence through the embedding layer, and positional encoding is added to preserve the order information. Encoder Stage: The encoder consists of stacked Transformer encoding blocks, each layer containing a multi-head self-attention mechanism and a feedforward neural network. The encoder processes the input sequence and outputs a sequence of hidden states, where each hidden state captures contextual information at its corresponding position. Finally, a context vector is generated through pooling operations. As a holistic representation of the dialogue history, the encoding process can be formalized as follows: ; in For encoder parameters, Indicates encoder, For sequence, For the encoder output, For encoder functions, For context vectors, This is the pooling function; Decoder stage: The decoder is based on the context vector The digest sequence is generated autoregressively, and the decoder is also composed of Transformer decoding blocks, generating one token at each step until an end token is produced. At time step... The decoder is based on the generated prefix. and context vector Predict the next token The decoding process uses an attention mechanism to dynamically focus on relevant parts of the encoder output, enhancing the accuracy of the summary. The probability distribution is calculated as follows: ; in For output layer weights, Indicates output. For decoder parameters, Indicates decoder, Let be a probability function. For time steps, For the next token, For the generated prefix, For context vectors, for Activation function Decoder function; Training and Optimization: The model is trained using supervised learning. The training data includes dialogue history-summary pairs. The loss function is cross-entropy loss, which maximizes the likelihood between the generated summary and the true summary. ; in, It is a loss function. The first true summary One token, This refers to the length of the abstract. For time steps, It is the natural logarithm function. For context vectors, In time step The true target output; Through the above implementation methods, the memory summary generation subunit solves the problem of information overload in long dialogue history processing, provides the system with a compact and interpretable memory foundation, supports the accuracy and real-time performance of empathic responses, and the entire design reflects the advantages of the sequence-to-sequence model in dialogue compression, which is in line with the overall goals of the system.

[0023] The emotion type classification unit of the emotion judgment module adopts a multi-label classification neural network model. The multi-label classification neural network model extracts the emotional keyword features in the dialogue text based on the attention mechanism and outputs a multi-dimensional emotion probability distribution. The emotion intensity assessment unit calculates the continuous value of emotion intensity through a regression model and discretizes it into a level from 1 to 5. The implementation of regression models typically employs a supervised learning framework. Model types can include linear regression, support vector regression, and lightweight neural network regression. The training phase uses labeled datasets, where each sample contains an input feature vector and a true sentiment intensity value. The model is optimized by minimizing the error between the predicted and true values, and the loss function often uses mean squared error. ; in For true strength, These are the model's predicted values. For model parameters, Represents the total number of samples. Indicates the end index of the summation. Indicates the starting index for summation. It is a loss function; During the inference phase, given new dialogue input, the model outputs continuous sentiment intensity values. This value reflects the relative strength of emotion. This continuous value calculation enhances the ability of fine-grained analysis and avoids the information loss caused by binarization.

[0024] The empathy decision-making unit of the empathy judgment and attitude determination module integrates a rule engine and a deep learning model. The rule engine triggers empathy decisions based on the emotional intensity threshold and historical empathy frequency. The deep learning model generates an empathy necessity score by fusing multimodal features and combining emotional type, intensity, and dialogue context. The attitude label selection unit selects the optimal label from a predefined attitude set based on the empathy necessity score.

[0025] The character retrieval submodule of the knowledge retrieval module uses the BM25 algorithm for keyword matching, the semantic retrieval submodule uses a pre-trained embedding vector model to calculate the semantic similarity between the query and the knowledge base, and the deduplication and reordering unit integrates the character retrieval score and the semantic retrieval score through a weighted score fusion strategy, and sets an upper limit on the fragment length to control the output scale. The core of the BM25 algorithm is the calculation of relevance scores, given a query. and knowledge base documents The algorithm calculates a score for each document. A higher score indicates that the document is more relevant to the query. The scoring formula is based on a combination of the following factors: Keyword frequency: the frequency of each keyword in a document The frequency of occurrence of words indicates that high-frequency words contribute more to the relevance, but overweighting is avoided by using a saturation function; Inverse Document Frequency (IDF): Measures the global importance of keywords. Rare words have higher IDF values, which improves their distinguishability. Document length normalization: via parameters To mitigate the impact of document length, prevent long documents from gaining an advantage due to word frequency accumulation, and ensure fairness for short documents, the standard BM25 scoring formula can be expressed as: ; in, It is the result of the entire function. For query, For knowledge base documents, Keywords In the document word frequency in It is a document Length, It is the average length of all documents in the knowledge base. This is a parameter for adjusting word frequency saturation, usually... , For length normalization parameters, usually , It is a query Total number of words after word segmentation It is an index. Inverse document frequency; Similarity calculation: Calculate With all document vectors in the knowledge base The semantic similarity between two objects is commonly measured by cosine similarity, and its formula is: ; in, Represents similarity, For query, For knowledge base documents, For index variables, The denominator represents the encoding function, with the numerator being the vector dot product and the denominator being the L2 norm product of the vectors. The cosine similarity range is [-1, 1], and the closer the value is to 1, the more semantically related it is.

[0026] The response generation module's conditional generation unit is based on the Transformer architecture. This unit uses attitude labels as prefix conditions and retrieved knowledge fragments as contextual constraints to generate response text through an autoregressive approach. The policy control unit incorporates a reinforcement learning strategy to dynamically adjust repetition penalties and diversity parameters during the generation process. The design of the reward function is crucial to its implementation; it integrates multiple quality dimensions of empathetic responses. The formula combines immediate rewards and sparse rewards, and is as follows: ; in, This is the current state. For optimal action, For consistent rewards, For diverse rewards, For repeated penalty items, To align with empathy and reward, to These are weighting coefficients; During training, the policy network updates its parameters using a proximal policy optimization algorithm, with the optimization objective being to maximize the expected cumulative reward. ; in, Represents the objective function. It is the objective function The parameters, As a discount factor, For state distribution, It is the expected value operator. The final time step for summation, For time steps, At time step state, In time step The action performed; Through the above implementation methods, the strategy control unit achieves dynamic optimization of the generation process, balancing the accuracy and diversity of empathetic responses and improving the user experience. This method not only solves the rigidity problem of traditional rule parameter adjustment but also adapts to diverse emotional interaction scenarios through data-driven learning.

[0027] The feedback collection unit of the log and data closed-loop module includes a user satisfaction rating interface and an error feedback labeling tool. The user satisfaction rating interface collects real-time rating data of users' responses, and the error feedback labeling tool allows administrators to manually correct and relabel the system output. The feedback collection unit sends the collected data to the data construction pipeline for incremental model training.

[0028] The system also includes a data generation and filtering module. This module synthesizes empathic dialogue data by simulating an environment through multi-agent interaction and introduces forward-looking information for the next round of questions during the response generation phase to enhance cross-round coherence. The data generation and filtering module integrates a reward model scoring mechanism to perform quality filtering on the synthesized data and retain highly consistent samples.

[0029] The system also includes an interpretable knowledge distillation module, which organizes the output of the teacher model into structured thinking fragments, including emotional evidence, reasons for empathy judgment, basis for attitude selection, and knowledge fragment identifiers. The interpretable knowledge distillation module trains a lightweight student model through efficient parameter fine-tuning technology, enabling the student model to internalize the "think first, then respond" decision-making paradigm.

[0030] The parameter fine-tuning technique of the interpretable knowledge distillation module adopts the LoRA method. The configurable parameters of the LoRA method include rank dimension, scaling factor, learning rate and training epochs, which are dynamically optimized according to the target hardware resources. The student model only needs to load the adapter weights during the inference stage, reducing memory usage and inference latency. The core idea of ​​LoRA is to assume that the weight updates during the fine-tuning process of the model have low-rank properties, for the pre-trained weight matrix Its update volume It can be decomposed into the product of two low-rank matrices: ; in, For the updated weights, This is the original weight matrix. For weight update amount, and It is a low-rank matrix. , and rank , It is the original weight matrix Dimensions, rank It is much smaller and Integers; In implementation, each linear transformation of the student model can be plugged into a LoRA adapter, with adapter parameters... Initialize as a Gaussian distribution. Initialize to zero to ensure the adapter output is zero during the initial training phase, without affecting the original pre-trained weights, rank The key hyperparameter controls the balance between the expressive power of the adapter and the number of parameters. In summary, the LoRA method achieves efficient parameter fine-tuning through a low-rank adapter, and works in conjunction with the goal of interpretable knowledge distillation to provide a scalable and low-cost model optimization solution for empathic customer service systems.

[0031] The operation steps of the empathic intelligent customer service system based on multi-agent and interpretable distillation are as follows: Step 1: Input and Memory Processing The system first captures the user's current input and the history of the last N rounds of dialogue through the input and memory modules. The dialogue history acquisition unit uses a sliding window mechanism to dynamically update the data to avoid information overload. The memory generation unit uses a sequence-to-sequence model to compress the historical dialogue and generate summary-style structured memory data. This step transforms the lengthy dialogue into a compact contextual representation, providing a foundation for subsequent sentiment analysis. The sliding window mechanism is implemented through a queue data structure to ensure that only the most recent N rounds of dialogue are retained. The memory summary model generates context vectors through an encoder-decoder architecture, retaining key information.

[0032] Step Two: Emotion Assessment The sentiment determination module receives structured memory data, identifies the user's sentiment category through a multi-label classification neural network model, and evaluates the sentiment intensity using a regression model. An attention mechanism is used to extract sentiment keywords to ensure the accuracy of the multi-dimensional sentiment probability distribution. The sentiment intensity calculation adopts continuous value regression followed by discretization to avoid binarization bias. The sentiment type classification unit outputs the probability distribution, and the sentiment intensity evaluation unit is optimized based on mean squared error loss to achieve fine-grained analysis.

[0033] Step 3: Empathy Judgment and Attitude Determination: The empathy assessment and attitude determination module integrates a rule engine and a deep learning model. It combines the emotion assessment results with the dialogue context to calculate the empathy necessity score. The rule engine triggers decisions based on the emotion intensity threshold and historical empathy frequency, while the deep learning model generates scores through multimodal feature fusion. The attitude label selection unit selects the optimal label from a predefined set based on the score to ensure that the response strategy matches the user's emotional state. The empathy decision-making unit dynamically adjusts the threshold to avoid excessive or insufficient empathy. The attitude selection is based on score ranking to achieve real-time adaptation.

[0034] Step 4: Knowledge Retrieval The knowledge retrieval module implements a dual-channel recall strategy: the character retrieval submodule uses the BM25 algorithm for keyword matching, and the semantic retrieval submodule calculates semantic similarity through a pre-trained embedding model. After each channel recalls Top-K candidate knowledge fragments, the deduplication and reordering unit outputs the final retrieval fragment through a weighted score fusion strategy, and controls the fragment length to optimize the scale. The BM25 algorithm considers word frequency, inverse document frequency, and document length normalization, and semantic retrieval uses vector similarity calculation to ensure that the retrieved content is both relevant and diverse.

[0035] Step 5: Response Generation The response generation module uses attitude tags as prefix conditions and retrieves knowledge fragments as contextual constraints. It autoregressively synthesizes response content through a Transformer-based conditional generation unit. The policy control unit introduces a reinforcement learning strategy to dynamically adjust the repetition penalty and diversity parameters in the generation process. The reward function integrates coherence, diversity, and empathy alignment indicators to ensure that the response is natural and emotionally appropriate. The generation process aims to maximize the cumulative reward. The conditional generation unit uses attitude tags as hard constraints to enhance the relevance of the response.

[0036] Step Six: Log and Data Closure: The log and data closed-loop module integrates data from the entire chain. The log recording unit stores interactive information, while the feedback collection unit integrates user satisfaction rating interfaces and manual annotation tools to collect real-time feedback data for incremental model training and system optimization, forming a data-driven closed loop. Feedback data is piped to the training process to achieve continuous learning, and error annotations allow administrators to correct them, improving data quality.

[0037] Step 7: Data Generation and Filtering Empathic dialogue data is synthesized by simulating an environment through multi-agent interaction, and "forward information for the next round of questions" is introduced into the response generation to enhance cross-round coherence. The reward model scoring mechanism filters the synthesized data, retains highly consistent samples, and solves the problem of scarcity of real data. Multi-agent interaction simulates real-world scenarios, and the reward model scores based on consistency indicators to ensure data quality.

[0038] Step 8: Distillation of Explainable Knowledge: The interpretable knowledge distillation module organizes the output of the teacher model into structured thought fragments and trains a lightweight student model through efficient parameter fine-tuning technology. The student model internalizes the "think first, then respond" decision paradigm, reducing the consumption of inference resources. The LoRA method realizes weight updates through low-rank matrix factorization, and the rank dimension is dynamically optimized according to hardware resources. The student model only loads adapter weights, improving efficiency.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An empathetic intelligent customer service system based on multi-agent and explainable distillation, characterized in that, include: The input and memory module uses the dialogue history acquisition unit to collect the user's current input and the history of the last N rounds of dialogue, and outputs structured dialogue memory data through the memory generation unit; The emotion determination module receives the structured dialogue memory data, identifies the user's emotion category through the emotion type classification unit, and outputs the emotion determination result using the emotion intensity assessment unit. The emotion categories include natural, happy, sad, angry, surprised, fearful, disgusted, frustrated, and excited, and the emotion intensity is a level from 1 to 5. The empathy judgment and attitude determination module receives the emotion judgment result and structured dialogue memory data, determines whether an empathy response is needed through the empathy decision unit, and outputs attitude tags using the attitude tag selection unit. The attitude tags include appeasement, apology, clarification, explanation, delay explanation, and price explanation. The knowledge retrieval module receives user input and the structured dialogue memory data, performs dual-channel retrieval through the character retrieval submodule and the semantic retrieval submodule, retrieves the Top-K candidate knowledge fragments from each, and performs weighted fusion using the deduplication and rearrangement unit to output the retrieved knowledge fragment; The response generation module receives the attitude tags and retrieved knowledge fragments, synthesizes the response content through the condition generation unit, and outputs the empathetic response through the strategy control unit; The log and data closed-loop module integrates end-to-end data, including user input, sentiment assessment results, attitude tags, retrieved knowledge fragments, and response content. Interaction information is stored through the log recording unit and used by the feedback collection unit for system optimization and data construction.

2. The empathic intelligent customer service system based on multi-agent and interpretable distillation as described in claim 1, characterized in that, The dialogue history acquisition unit of the input and memory module specifically includes a dialogue history caching subunit and a memory summary generation subunit. The dialogue history caching subunit is used to dynamically store the dialogue text of the last N rounds and update the historical data through a sliding window mechanism. The memory summary generation subunit uses a sequence-to-sequence model to compress the dialogue history, generate a summary memory representation, and output it to the sentiment determination module.

3. The empathic intelligent customer service system based on multi-agent and interpretable distillation as described in claim 1, characterized in that, The emotion type classification unit of the emotion determination module adopts a multi-label classification neural network model. The multi-label classification neural network model extracts the emotional keyword features in the dialogue text based on the attention mechanism and outputs a multi-dimensional emotion probability distribution. The emotion intensity assessment unit calculates the continuous value of emotion intensity through a regression model and discretizes it into a level from 1 to 5.

4. The empathic intelligent customer service system based on multi-agent and interpretable distillation as described in claim 1, characterized in that, The empathy judgment and attitude determination module integrates a rule engine and a deep learning model in its empathy decision-making unit. The rule engine triggers empathy decisions based on the emotional intensity threshold and historical empathy frequency. The deep learning model generates an empathy necessity score by fusing multimodal features and combining emotional type, intensity, and dialogue context. The attitude label selection unit selects the optimal label from a predefined attitude set based on the empathy necessity score.

5. The empathic intelligent customer service system based on multi-agent and interpretable distillation according to claim 1, characterized in that, The character retrieval submodule of the knowledge retrieval module uses the BM25 algorithm for keyword matching, the semantic retrieval submodule uses a pre-trained embedding vector model to calculate the semantic similarity between the query and the knowledge base, and the deduplication and reordering unit integrates the character retrieval score and the semantic retrieval score through a weighted score fusion strategy, and sets an upper limit on the fragment length to control the output scale.

6. The empathic intelligent customer service system based on multi-agent and interpretable distillation according to claim 1, characterized in that, The condition generation unit of the response generation module is based on the Transformer architecture. The condition generation unit uses attitude labels as prefix conditions and retrieves knowledge fragments as context constraints to generate response text through an autoregressive approach. The policy control unit introduces a reinforcement learning strategy to dynamically adjust the repetition penalty and diversity parameters in the generation process.

7. The empathic intelligent customer service system based on multi-agent and interpretable distillation according to claim 1, characterized in that, The feedback collection unit of the log and data closed-loop module includes a user satisfaction rating interface and an error feedback labeling tool. The user satisfaction rating interface collects real-time rating data of users' responses, and the error feedback labeling tool allows administrators to manually correct and relabel the system output. The feedback collection unit sends the collected data to the data construction pipeline for incremental model training.

8. The empathic intelligent customer service system based on multi-agent and interpretable distillation according to claim 1, characterized in that, The system also includes a data generation and filtering module, which synthesizes empathic dialogue data through a multi-agent simulated environment and introduces forward-looking information for the next round of questions during the response generation stage to enhance cross-round coherence. The data generation and filtering module integrates a reward model scoring mechanism to perform quality filtering on the synthesized data and retain highly consistent samples.

9. The empathic intelligent customer service system based on multi-agent and interpretable distillation according to claim 1, characterized in that, The system also includes an interpretable knowledge distillation module, which organizes the output of the teacher model into structured thinking fragments, including emotional evidence, reasons for empathy judgment, basis for attitude selection, and knowledge fragment identifiers. The interpretable knowledge distillation module trains a lightweight student model through efficient parameter fine-tuning technology, enabling the student model to internalize the "think first, then respond" decision-making paradigm.

10. The empathic intelligent customer service system based on multi-agent and interpretable distillation according to claim 9, characterized in that, The parameter fine-tuning technique of the interpretable knowledge distillation module adopts the LoRA method. The configurable parameters of the LoRA method include rank dimension, scaling factor, learning rate and training epochs, which are dynamically optimized according to the target hardware resources. The student model only needs to load adapter weights during the inference stage, reducing memory usage and inference latency.

Images