User complaint processing method and device, electronic equipment and storage medium

Abstract

The invention provides a user complaint processing method and apparatus, an electronic device and a storage medium. The method comprises the steps of obtaining user complaint information; based on the first cue word, guiding the knowledge analysis large language model to determine key information, guiding the knowledge analysis large language model to retrieve the key information in the knowledge graph based on a path sorting algorithm, and generating a plurality of candidate solutions based on a retrieval result; performing sentiment analysis on the user complaint information based on a sentiment analysis large language model, and determining the sentiment level of the user complaint information; and based on a second cue word, guiding a knowledge analysis large language model, and based on the emotion level, determining a final solution from the plurality of candidate solutions. According to the method, an automatic complaint processing process from information extraction, knowledge retrieval to scheme generation and emotion adaptation is constructed through division and cooperation of the knowledge analysis large language model and the emotion analysis large language model, a judgment process needing manual intervention is avoided, and complaint processing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a user complaint handling method, device, electronic device, and storage medium. Background Art

[0002] With the widespread adoption and development of wireless networks, users are increasingly demanding higher quality of service. However, in actual use, users may encounter various network issues, such as weak signals, network outages, and slow speeds. These issues require prompt and effective resolution. Traditional complaint handling methods suffer from low efficiency and poor user experience, failing to meet user needs.

[0003] Existing methods for handling customer complaints rely on manual customer service. These processes are plagued by issues such as response delays, high labor costs, and poor emotional perception, resulting in low processing efficiency. Summary of the Invention

[0004] The present invention provides a user complaint processing method, device, electronic device and storage medium, which are used to improve the processing efficiency of customer complaints.

[0005] The present invention provides a method for handling user complaints, comprising the following steps: Obtain user complaint information; Based on the first prompt word, guiding the knowledge analysis large language model to determine key information of the user complaint information, and guiding the knowledge analysis large language model to retrieve the key information in the knowledge graph based on the path sorting algorithm, and generating multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from the historical complaint processing information; Based on the sentiment analysis large language model, sentiment analysis is performed on the user complaint information to determine the sentiment level of the user complaint information; Based on the second prompt word, the knowledge analysis large language model is guided to determine a final solution from the multiple candidate solutions based on the sentiment level.

[0006] According to a user complaint handling method provided by the present invention, the user complaint information includes text complaint information, image complaint information, and voice complaint information, and obtaining the user complaint information includes: Extracting features of the user complaint information to obtain text complaint features, image complaint features, and voice complaint features; Performing multimodal feature fusion on the text complaint feature, the image complaint feature, and the voice complaint feature to obtain a fused feature; The user complaint information is determined based on the fusion feature.

[0007] According to a user complaint processing method provided by the present invention, performing multimodal feature fusion on the text complaint feature, the image complaint feature, and the voice complaint feature to obtain a fused feature includes: A multimodal feature fusion framework based on a multi-head attention mechanism performs multimodal fusion on the text complaint features, the image complaint features, and the voice complaint features to obtain fused features.

[0008] According to a user complaint handling method provided by the present invention, after guiding the knowledge analysis large language model based on the second prompt word and determining a final solution from the multiple candidate solutions based on the sentiment level, the method further includes: Based on the reinforcement learning algorithm, the dialogue strategy model is trained to obtain the trained dialogue strategy model; After the final solution is sent to the target user corresponding to the user complaint information, a dialogue interaction is performed with the target user based on the trained dialogue strategy model, and feedback information of the target user regarding the final solution is determined from the dialogue interaction process; The reward function of the reinforcement learning is: ; in, is the reward, N is the number of solution steps, and S is the user score.

[0009] According to a user complaint handling method provided by the present invention, after determining the target user's feedback information regarding the final solution from the dialogue interaction process, the method further includes: Based on the feedback information, the knowledge analysis large language model is trained, and based on the training result, the knowledge analysis large language model is updated.

[0010] According to a user complaint handling method provided by the present invention, the process of constructing the knowledge graph includes: Based on keyword matching, entity information in the historical complaint handling information is identified to obtain the entity information in the historical complaint handling information; Determining associations between entities in the historical complaint handling information based on relationship rules in the historical complaint handling information, and determining relationship information based on the associations between the entities; The knowledge graph is constructed based on the entity information and the relationship information.

[0011] According to a user complaint handling method provided by the present invention, the key information includes intention information, complaint problem information and demand information.

[0012] According to a user complaint handling method provided by the present invention, the first prompt word is used to guide the knowledge analysis large language model to perform intent recognition on the user complaint information, obtain the intent information of the user complaint information, and guide the knowledge analysis large language model to perform entity recognition on the user complaint information, determine the complaint problem information of the user complaint information, and guide the knowledge analysis large language model to perform demand mining on the user complaint information, and determine the demand information in the user complaint information.

[0013] The present invention also provides a user complaint handling device, comprising the following modules: Information acquisition module, used to obtain user complaint information; a key information extraction module, configured to guide the knowledge analysis large language model to determine key information of the user complaint information based on the first prompt word, guide the knowledge analysis large language model to retrieve the key information in the knowledge graph based on a path sorting algorithm, and generate multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from historical complaint processing information; A sentiment analysis module, configured to perform sentiment analysis on the user complaint information based on a sentiment analysis language model to determine the sentiment level of the user complaint information; A solution determination module is used to guide the knowledge analysis large language model based on the second prompt word, and determine a final solution from the multiple candidate solutions based on the sentiment level.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the program, any of the above-mentioned user complaint handling methods is implemented.

[0015] The user complaint handling method, device, electronic device, and storage medium provided by this invention utilize a large language model for knowledge analysis and a large language model for sentiment analysis to collaborate, establishing an automated complaint handling process from information extraction and knowledge retrieval to solution generation and sentiment adaptation. This eliminates the need for manual judgment and improves complaint handling efficiency. The large language model for knowledge analysis generates candidate solutions, while the large language model for sentiment analysis provides decision-making support based on subjective sentiment. This multi-model collaborative mechanism not only further improves efficiency but also enhances decision-making accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 It is a flowchart of the user complaint handling method provided by the present invention.

[0018] Figure 2 It is a schematic diagram of the retrieval analysis architecture provided by the present invention.

[0019] Figure 3 It is a structural diagram of the user complaint handling device provided by the present invention.

[0020] Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0022] With the widespread adoption and development of wireless networks, users are increasingly demanding higher quality of service. However, in practice, users may encounter various network issues, such as weak signals, network outages, and slow speeds. These issues require prompt and effective resolution. Traditional complaint handling methods suffer from low efficiency and poor user experience, failing to meet user needs. Therefore, an intelligent customer service system is needed that can quickly and accurately handle user complaints.

[0023] The processing process of related methods for handling customer complaints relies on manual customer service, which has problems such as response delay, high labor cost, and weak emotional perception ability.

[0024] The shortcomings of related methods include: Rule-based systems: rely on preset templates and cannot handle complex semantics (such as metaphors and multi-intent complaints).

[0025] Unimodal processing: Only supports text interaction and has difficulty analyzing emotional fluctuations in voice or evidence information in images / videos.

[0026] Static knowledge base: The solution is fixed and cannot be dynamically linked to changes in corporate policies or external data (such as logistics information and third-party credit reports).

[0027] Emotional blind spot: Lack of real-time feedback on customers’ emotional state can easily lead to secondary complaints.

[0028] In view of the defects in the related methods, the present invention provides a method for handling user complaints. Figure 1 FIG. 1 is a flow chart of the user complaint handling method provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 110: Obtain user complaint information; Step 120: Based on the first prompt word, guide the knowledge analysis large language model to determine key information of the user complaint information, guide the knowledge analysis large language model to search for the key information in the knowledge graph based on the path sorting algorithm, and generate multiple candidate solutions based on the search results. The knowledge graph is constructed based on entity information and relationship information extracted from the historical complaint processing information; Step 130: Perform sentiment analysis on the user complaint information based on a sentiment analysis language model to determine the sentiment level of the user complaint information. Step 140 : Based on the second prompt word, guide the knowledge analysis large language model to determine a final solution from the multiple candidate solutions based on the sentiment level.

[0029] The user complaint handling method provided by the present invention may be executed by an electronic device, a component of an electronic device, an integrated circuit, or a chip. The electronic device may be a mobile electronic device or a non-mobile electronic device. For example, the mobile electronic device may be a mobile phone, tablet computer, laptop computer, PDA, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), while the non-mobile electronic device may be a server, network attached storage (NAS), or personal computer (PC), etc., although this invention does not impose any specific limitations thereon.

[0030] The following takes the user complaint handling method provided by the present invention executed by a computer as an example to describe the technical solution of the present invention in detail.

[0031] In step 110 , user complaint information sent by a user may be received based on a user interface module.

[0032] User complaint information can include text, voice, image, or video input. After receiving input from multiple modalities, cross-modal feature alignment can be performed based on the Transformer architecture, and user complaint information can be fused.

[0033] In step 120, based on the first prompt word, the knowledge analysis big language model is guided to determine the key information of the user complaint information, and the knowledge analysis big language model is guided to retrieve the key information in the knowledge graph based on the path sorting algorithm, and multiple candidate solutions are generated based on the retrieval results. The knowledge graph is constructed based on entity information and relationship information extracted from historical complaint processing information.

[0034] Pre-build a knowledge graph. Identify core entities such as complainants, products, issue types, and resolution status from historical complaint data. Construct complaint relationships (e.g., "user-complaint-product"), causal relationships (e.g., "fault-cause-service interruption"), and solution relationships (e.g., "issue-matching-solution"). Use a graph database (e.g., Neo4j) to store triples, facilitating complex queries and reasoning.

[0035] The knowledge graph constructed can include three levels: the enterprise level (product information, after-sales service policies), the industry level (relevant laws and regulations), and the user level (historical complaint records, user preferences).

[0036] The constructed knowledge graph can be used as the knowledge base of a large language model. Combined with RAG (Retrieval-Augmented Generation) knowledge base retrieval technology, it can achieve efficient retrieval and application of knowledge.

[0037] When processing user complaints, we first need to deeply analyze the complaint text using a knowledge analysis large language model (which can be an advanced AI system with natural language understanding and information extraction capabilities) to accurately extract key information. This process involves technologies such as semantic recognition, entity extraction, and relationship mapping. It aims to extract structured data such as the core issue, involved parties, time points, and scope of impact from unstructured text, laying the foundation for subsequent processing.

[0038] The extracted key information is then converted into query nodes or path constraints recognizable by the knowledge graph. As a structured knowledge base, the knowledge graph stores domain knowledge in the form of entity-relationship-entity triples. Its powerful semantic association capabilities support complex query scenarios. Based on a path sorting algorithm, multi-hop reasoning is performed within the graph, dynamically evaluating the strength of each node's association with the complaint topic and prioritizing highly relevant paths to accurately locate knowledge nodes or solution templates directly related to the complaint.

[0039] Ultimately, the search results generate multiple candidate solutions. These solutions may include direct fixes for specific issues (such as technical troubleshooting guides), analog solutions based on historical cases, suggestions for business process optimization, or coordinated solutions requiring cross-departmental collaboration. Each solution is accompanied by a confidence score and feasibility statement to help decision makers quickly prioritize.

[0040] By deeply integrating natural language understanding, knowledge graph reasoning and path optimization algorithms, the intelligence level of complaint handling can be significantly improved, and an end-to-end automated closed loop from problem identification to solution generation can be achieved.

[0041] Optionally, the process of determining key information in user complaint information based on knowledge analysis large language model can be as follows: Figure 2 The retrieval analysis architecture provided by the present invention is shown in the schematic diagram.

[0042] Specifically, it includes the large model layer, conversation flow data layer, index layer and Chunk layer.

[0043] The large model layer is the underlying foundation of the entire system, responsible for understanding and generating natural language, executing user intent, and generating corresponding responses.

[0044] Dialogue link layer: The dialogue link layer includes query history, query rewriting, prompt word encapsulation, prompt encapsulation and security filtering.

[0045] Query history: The system first checks whether the user's query is related to previous queries to ensure the continuity of the conversation. Query rewriting: Based on the query history, the user's query is rewritten to make it more consistent with the format that the system understands and processes. Prompt word encapsulation: The query vector and knowledge graph information are encapsulated into appropriate prompt words for use by the upper-level large model. Prompt encapsulation: The prompt words are encapsulated into prompts to guide the large model to generate appropriate responses. Security filtering: The responses generated by the large model are securely filtered to ensure that the response content complies with the regulations and does not trigger sensitive or inappropriate reactions.

[0046] The indexing layer is used to build indexes and perform retrieval processes in the knowledge graph based on the built indexes.

[0047] In step 130, sentiment analysis is performed on the user complaint information based on the sentiment analysis language model to determine the sentiment level of the user complaint information.

[0048] In the user complaint handling process, the complaint text is first deeply decoded through the sentiment analysis large language model. This model can identify the intensity of negative emotions contained in the user's text and quantitatively assess its emotional level (such as mild dissatisfaction, moderate complaint, high anger, etc.). This process relies on a comprehensive analysis of multiple dimensions such as lexical sentiment polarity, syntactic structure, and contextual relevance to ensure the accuracy of the sentiment grading results.

[0049] The emotion level can be determined based on the emotion analysis language model and the emotion quantification formula. The emotion quantification level E_total is: E_total=α×E_text+β×E_voice+γ×E_image; Among them, α+β+γ=1 is the weight, and the weight is dynamically adjusted according to the input modality. E_text is the sentiment analysis result corresponding to the text complaint feature, E_voice is the sentiment analysis result corresponding to the voice complaint feature, and E_image is the corresponding sentiment analysis result corresponding to the image complaint feature.

[0050] In step 140 , based on the second prompt word, the knowledge analysis large language model is guided to determine a final solution from the multiple candidate solutions based on the sentiment level.

[0051] Specifically, a decision-making framework can be constructed in conjunction with the second prompt word, with sentiment rating as the core decision variable input into the knowledge analysis large language model. Based on the generated multiple candidate solutions, this model integrates association rules between sentiment rating and solution attributes (such as response timeliness, resource commitment, and user satisfaction predictions). Using weighted scoring, rule engines, or reinforcement learning mechanisms, it dynamically adjusts the priority ranking of each solution. For example, for highly angry complaints, the system will prioritize resource-intensive solutions that can quickly calm emotions; whereas for mild dissatisfaction, it will tend to recommend more cost-effective improvement suggestions (such as process optimization and service remediation).

[0052] The knowledge analysis large language model outputs the final solution that best matches the sentiment level, along with an explanation of the decision-making basis (such as sentiment-solution mapping and expected impact predictions). This process achieves the coordinated optimization of sentiment perception and knowledge reasoning, avoiding the emotional neglect caused by mechanical processing while ensuring the professionalism of the solution through the semantic association capabilities of the knowledge graph. This balances emotional response and problem solving, significantly improving user satisfaction and efficiency in complaint handling.

[0053] For issues that can't be resolved directly, a complaint ticket is automatically generated and assigned to the appropriate technician based on factors such as the problem type and location. By combining large and small model collaboration technology, intelligent classification and dispatching are achieved, improving processing efficiency and accuracy.

[0054] The user complaint handling method provided by this invention, through the collaboration of a large language model for knowledge analysis and a large language model for sentiment analysis, establishes an automated complaint handling process from information extraction and knowledge retrieval to solution generation and sentiment adaptation. This eliminates the need for manual judgment and improves complaint handling efficiency. The large language model for knowledge analysis generates candidate solutions, while the large language model for sentiment analysis provides decision-making based on the subjective emotional dimension. This multi-model collaborative mechanism not only further improves efficiency but also enhances decision-making accuracy.

[0055] In one embodiment, user complaint information includes text complaint information, image complaint information, and voice complaint information. Obtaining the user complaint information includes: extracting features of the user complaint information to obtain text complaint features, image complaint features, and voice complaint features; performing multimodal feature fusion on the text complaint features, the image complaint features, and the voice complaint features to obtain fused features; and determining the user complaint information based on the fused features.

[0056] Specifically, the acquired user complaint information includes text complaint information, image complaint information, and voice complaint information.

[0057] After obtaining user complaint information, feature extraction is performed on the user complaint information. Feature extraction is performed on text complaint information to obtain text complaint features. For image complaint information, optical character recognition (OCR) technology can be used to extract text information from the image. Feature extraction is then performed based on the recognized text information to obtain image complaint features. For voice complaint information, voice recognition can be used to first determine the text information corresponding to the voice complaint information. Feature extraction is then performed on the text information corresponding to the determined voice complaint information to obtain voice complaint features.

[0058] Perform multimodal feature fusion on the text complaint feature, the image complaint feature, and the voice complaint feature to obtain a fusion feature; and determine the user complaint information based on the fusion feature.

[0059] The core goal of multimodal feature fusion is to eliminate the semantic gap between different modalities and construct a unified feature that can comprehensively reflect complaint information.

[0060] Independent models (such as text classifiers, image classifiers, and speech classifiers) can be trained for each modality, and the outputs of each model can be integrated through methods such as weighted voting and attention mechanisms to obtain fused features. For example, in a complaint classification task, the text model, image model, and speech model each output classification probabilities, which are ultimately aggregated into a comprehensive probability using attention weights.

[0061] After obtaining the fused features, user complaint information is generated for input into the large language model for knowledge analysis.

[0062] In one embodiment, multimodal feature fusion is performed on the text complaint features, the image complaint features, and the voice complaint features to obtain fused features, including: a multimodal feature fusion framework based on a multi-head attention mechanism, multimodal feature fusion is performed on the text complaint features, the image complaint features, and the voice complaint features to obtain fused features.

[0063] Specifically, text, image, and speech features are input into the multi-head attention module as query, key, and value, and the association weights between modalities are calculated through the self-attention mechanism or the cross-modal attention mechanism.

[0064] The features are mapped to multiple subspaces, each head independently calculates the attention weight, and finally the results are fused through linear transformation to enhance the model's modeling ability for different semantic dimensions.

[0065] According to the weight matrix output by the multi-head attention, the features of each modality are weighted and summed to generate fusion features.

[0066] Specifically, we can use the multi-head attention mechanism formula: ; The calculation of each attention head is: ; Through the multi-head attention mechanism, multiple attention heads can be calculated in parallel to capture different features, thereby gaining a more comprehensive understanding of user complaint information.

[0067] In one embodiment, after guiding the knowledge analysis large language model based on the second prompt word and determining a final solution from the multiple candidate solutions based on the sentiment level, the method further includes: Based on the reinforcement learning algorithm, the dialogue strategy model is trained to obtain the trained dialogue strategy model; After the final solution is sent to the target user corresponding to the user complaint information, a dialogue interaction is performed with the target user based on the trained dialogue strategy model, and feedback information of the target user regarding the final solution is determined from the dialogue interaction process; The reward function of the reinforcement learning is: ; in, is the reward, N is the number of solution steps, and S is the user score.

[0068] Reinforcement learning uses the interaction between an agent and its environment to learn the optimal strategy to maximize cumulative rewards. In a dialogue policy model, the environment consists of user input, contextual information, and historical system conversations. The agent must choose the optimal action based on its current state.

[0069] The reward function is the core of reinforcement learning, used to measure the quality of a strategy. In complaint handling scenarios, the reward function must consider both resolution efficiency (number of steps) and user satisfaction (rating): Penalty for number of steps: With each additional round of interaction, the reward decreases by a fixed value (e.g., -1) to encourage the model to solve problems in the fewest steps. User rating incentive: The user's rating of the solution (e.g., 1-5) is directly converted into a reward (e.g., 5 points corresponds to +10, 1 point corresponds to -5).

[0070] After obtaining the trained dialogue strategy model, the final solution is sent to the target user corresponding to the user complaint information. Interaction with the target user is carried out based on the dialogue strategy model. It should be noted that the core function of the dialogue strategy model is to optimize the user complaint handling process through intelligent interaction, improving user experience and service efficiency.

[0071] After the interaction is completed, feedback information from the target user regarding the final solution during the dialogue interaction is obtained. The feedback information may be based on guiding the target user to rate the final solution.

[0072] Based on the feedback information, corresponding optimization measures can be taken. For example, if the target user repeats the question more than three times based on the feedback information, the channel will be automatically upgraded to a higher level.

[0073] In one embodiment, after determining the target user's feedback information regarding the final solution from the dialogue interaction process, it also includes: training the knowledge analysis large language model based on the feedback information, and updating the knowledge analysis large language model based on the training results.

[0074] Continuous training and iterative optimization of large language models for knowledge analysis based on user feedback is a key path to improving the practicality and accuracy of the model.

[0075] Specifically, in the user complaint handling scenario, after completing the interaction with the target user through the reinforcement learning dialogue strategy model, feedback data can be collected, such as the user's rating of the solution or satisfaction evaluation.

[0076] After being cleaned and annotated, this feedback data is fed into the training framework of the knowledge analysis large language model, serving as a supervisory signal to guide model learning. During training, incremental learning techniques are used to update only model parameters relevant to the complaint scenario, avoiding interference with the general knowledge base. Furthermore, comparative learning methods enable the model to distinguish between the representations of "effective solutions" and "inefficient solutions."

[0077] Deploying the optimized model parameters to the production environment forms a closed loop of "feedback collection - model training - strategy update," achieving a spiral improvement in service quality. This process enhances the adaptability of the knowledge analysis large language model to complex complaint scenarios.

[0078] In one embodiment, the process of constructing a knowledge graph includes: identifying entity information in the historical complaint handling information based on keyword matching to obtain the entity information in the historical complaint handling information; determining the association between entities in the historical complaint handling information based on the relationship rules in the historical complaint handling information, and determining the relationship information based on the association between the entities; and constructing the knowledge graph based on the entity information and the relationship information.

[0079] Through keyword matching and entity recognition technology, key entity information is extracted from historical complaint processing information. For example, regular expressions or pre-trained named entity recognition models are used to accurately locate entities such as complaint subjects, problem types, and processing links, and these entities are uniformly mapped to a standardized ontology vocabulary to eliminate expression differences.

[0080] Based on relational rules and semantic analysis, the association logic between entities is established. For example, through template matching or graph neural network (GNN), the causal relationship between "user A complains that product B has quality problems" and "customer service C proposes a refund plan at time T" can be inferred, or the aggregation pattern of "similar problems occur multiple times" can be identified, thereby connecting scattered entities into a structured network.

[0081] By building a knowledge graph based on entity-relationship triples (e.g., <user A, complaint, product B>, <product B, issue type, quality issue>), and using graph databases like Neo4j for storage and visualization, we transform the implicit knowledge in complaint data (e.g., the distribution of high-frequency issues and processing time bottlenecks) into a queryable and reasonable explicit graph, providing data support for subsequent complaint prediction and service optimization. This process transforms unstructured text into structured knowledge, significantly improving the efficiency of complaint data utilization.

[0082] In one embodiment, key information includes intent information, complaint problem information, and demand information. The first prompt word is used to guide the knowledge analysis large language model to perform intent recognition on the user complaint information to obtain the intent information of the user complaint information, guide the knowledge analysis large language model to perform entity recognition on the user complaint information to determine the complaint problem information of the user complaint information, and guide the knowledge analysis large language model to perform demand mining on the user complaint information to determine the demand information in the user complaint information.

[0083] Specifically, the first prompt word is used to implement a three-level data analysis process. First-level intent recognition: Determines the complaint type (refund / compensation / consultation). Second-level entity extraction: Extracts structured fields such as time, order number, and problem object. Third-level implicit demand mining: Captures potential customer demands through comparative learning (e.g., long-term customers seeking additional compensation).

[0084] By designing a hierarchical first prompt word to guide the knowledge analysis large language model, a three-level progressive data analysis is completed: First, in the first-level intent recognition stage, the model is based on a predefined complaint type classification system (such as refund, compensation, consultation, technical support, etc.), and performs global semantic matching on the complaint text through the classification head to obtain the intent information corresponding to the complaint type.

[0085] Secondly, in the secondary entity extraction stage, the model can call the named entity recognition module, use the pre-trained domain vocabulary (such as order number, product model, timestamp) and positional encoding (Positional Encoding) to perform structured analysis of the complaint text, extract entity information such as time, place, and problem object, and map it to the standard knowledge base through entity linking technology.

[0086] In the third-level implicit demand mining stage, the model can use a comparative learning framework to compare and analyze current complaint cases with similar historical cases (such as comparing the differences in compensation plans for "ordinary user complaints" and "star user complaints"), identify potential demand patterns by calculating the semantic similarity matrix, and infer implicit demands (such as long-term users may expect to obtain additional points compensation or priority processing rights) in combination with user portraits (such as consumption frequency and historical complaint records).

[0087] The hierarchical analysis framework not only improves the structuring of complaint information, but also provides a more accurate basis for service optimization through the dual mining of explicit and implicit demands.

[0088] The user complaint handling device provided by the present invention is described below. The user complaint handling device described below and the user complaint handling method described above can be referenced to each other.

[0089] like Figure 3 As shown, the device includes: Information acquisition module 310, used to obtain user complaint information; Key information extraction module 320, configured to guide the knowledge analysis large language model to determine key information of the user complaint information based on the first prompt word, guide the knowledge analysis large language model to search for the key information in the knowledge graph based on a path sorting algorithm, and generate multiple candidate solutions based on the search results. The knowledge graph is constructed based on entity information and relationship information extracted from historical complaint handling information; Sentiment analysis module 330, configured to perform sentiment analysis on the user complaint information based on a sentiment analysis language model to determine a sentiment level of the user complaint information; The solution determination module 340 is configured to guide the knowledge analysis large language model based on the second prompt word and determine a final solution from the multiple candidate solutions based on the sentiment level.

[0090] The user complaint handling device provided by this invention utilizes a large language model for knowledge analysis and a large language model for sentiment analysis to automatically handle complaints, from information extraction and knowledge retrieval to solution generation and sentiment adaptation. This eliminates the need for manual judgment and improves complaint handling efficiency. The large language model for knowledge analysis generates candidate solutions, while the large language model for sentiment analysis provides decision-making based on subjective sentiment. This multi-model collaborative mechanism not only further improves efficiency but also enhances decision-making accuracy.

[0091] In one embodiment, the information acquisition module 310 is specifically configured to: User complaint information includes text complaint information, image complaint information, and voice complaint information. Acquiring user complaint information includes: Extracting features of the user complaint information to obtain text complaint features, image complaint features, and voice complaint features; Performing multimodal feature fusion on the text complaint feature, the image complaint feature, and the voice complaint feature to obtain a fused feature; The user complaint information is determined based on the fusion feature.

[0092] In one embodiment, the information acquisition module 310 is further specifically configured to: Performing multimodal feature fusion on the text complaint feature, the image complaint feature, and the voice complaint feature to obtain fused features, including: A multimodal feature fusion framework based on a multi-head attention mechanism performs multimodal fusion on the text complaint features, the image complaint features, and the voice complaint features to obtain fused features.

[0093] In one embodiment, the solution determination module 340 is specifically configured to: After guiding the knowledge analysis large language model based on the second prompt word and determining a final solution from the multiple candidate solutions based on the sentiment level, the method further includes: Based on the reinforcement learning algorithm, the dialogue strategy model is trained to obtain the trained dialogue strategy model; After the final solution is sent to the target user corresponding to the user complaint information, a dialogue interaction is performed with the target user based on the trained dialogue strategy model, and feedback information of the target user regarding the final solution is determined from the dialogue interaction process; The reward function of the reinforcement learning is: ; in, is the reward, N is the number of solution steps, and S is the user score.

[0094] In one embodiment, the solution determination module 340 is specifically configured to: After determining the target user's feedback information regarding the final solution from the dialogue interaction process, the method further includes: Based on the feedback information, the knowledge analysis large language model is trained, and based on the training result, the knowledge analysis large language model is updated.

[0095] In one embodiment, the key information extraction module 320 is specifically configured to: The construction process of the knowledge graph includes: Based on keyword matching, entity information in the historical complaint handling information is identified to obtain the entity information in the historical complaint handling information; Determining associations between entities in the historical complaint handling information based on relationship rules in the historical complaint handling information, and determining relationship information based on the associations between the entities; The knowledge graph is constructed based on the entity information and the relationship information.

[0096] In one embodiment, the key information extraction module 320 is specifically configured to: Key information includes intention information, complaint information, and demand information.

[0097] In one embodiment, the key information extraction module 320 is further specifically configured to: The first prompt word is used to guide the knowledge analysis large language model to perform intent recognition on the user complaint information, obtain the intent information of the user complaint information, and guide the knowledge analysis large language model to perform entity recognition on the user complaint information, determine the complaint problem information of the user complaint information, and guide the knowledge analysis large language model to perform demand mining on the user complaint information, and determine the demand information in the user complaint information.

[0098] Figure 4 An example of a physical structure diagram of an electronic device is shown below. Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 may call logic instructions in the memory 430 to execute a user complaint handling method, which includes: obtaining user complaint information; Based on the first prompt word, guiding the knowledge analysis large language model to determine key information of the user complaint information, and guiding the knowledge analysis large language model to retrieve the key information in the knowledge graph based on the path sorting algorithm, and generating multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from the historical complaint processing information; Based on the sentiment analysis large language model, sentiment analysis is performed on the user complaint information to determine the sentiment level of the user complaint information; Based on the second prompt word, the knowledge analysis large language model is guided to determine a final solution from the multiple candidate solutions based on the sentiment level.

[0099] Furthermore, the logic instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0100] On the other hand, the present invention further provides a computer program product, the computer program product including a computer program, the computer program being storable on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can execute the user complaint handling method provided by the above methods, the method including: obtaining user complaint information; Based on the first prompt word, guiding the knowledge analysis large language model to determine key information of the user complaint information, and guiding the knowledge analysis large language model to retrieve the key information in the knowledge graph based on the path sorting algorithm, and generating multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from the historical complaint processing information; Based on the sentiment analysis large language model, sentiment analysis is performed on the user complaint information to determine the sentiment level of the user complaint information; Based on the second prompt word, the knowledge analysis large language model is guided to determine a final solution from the multiple candidate solutions based on the sentiment level.

[0101] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for handling user complaints provided by the above methods is implemented, the method comprising: obtaining user complaint information; Based on the first prompt word, guiding the knowledge analysis large language model to determine key information of the user complaint information, and guiding the knowledge analysis large language model to retrieve the key information in the knowledge graph based on the path sorting algorithm, and generating multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from the historical complaint processing information; Based on the sentiment analysis large language model, sentiment analysis is performed on the user complaint information to determine the sentiment level of the user complaint information; Based on the second prompt word, the knowledge analysis large language model is guided to determine a final solution from the multiple candidate solutions based on the sentiment level.

[0102] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0103] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for handling user complaints, characterized in that: include: Obtain user complaint information; Based on the first prompt word, guiding the knowledge analysis large language model to determine key information of the user complaint information, and guiding the knowledge analysis large language model to retrieve the key information in the knowledge graph based on the path sorting algorithm, and generating multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from the historical complaint processing information; Based on the sentiment analysis large language model, sentiment analysis is performed on the user complaint information to determine the sentiment level of the user complaint information; Based on the second prompt word, the knowledge analysis large language model is guided to determine a final solution from the multiple candidate solutions based on the sentiment level.

2. The user complaint handling method according to claim 1, characterized in that: The user complaint information includes text complaint information, image complaint information, and voice complaint information. The obtaining of the user complaint information includes: Extracting features of the user complaint information to obtain text complaint features, image complaint features, and voice complaint features; Performing multimodal feature fusion on the text complaint feature, the image complaint feature, and the voice complaint feature to obtain a fused feature; The user complaint information is determined based on the fusion feature.

3. The user complaint handling method according to claim 2, characterized in that: The multimodal feature fusion of the text complaint feature, the image complaint feature, and the voice complaint feature to obtain the fused feature includes: A multimodal feature fusion framework based on a multi-head attention mechanism performs multimodal fusion on the text complaint features, the image complaint features, and the voice complaint features to obtain fused features.

4. The user complaint handling method according to claim 1, characterized in that: After guiding the knowledge analysis large language model based on the second prompt word and determining a final solution from the multiple candidate solutions based on the sentiment level, the method further includes: Based on the reinforcement learning algorithm, the dialogue strategy model is trained to obtain the trained dialogue strategy model; After the final solution is sent to the target user corresponding to the user complaint information, a dialogue interaction is performed with the target user based on the trained dialogue strategy model, and feedback information of the target user regarding the final solution is determined from the dialogue interaction process; The reward function of the reinforcement learning is: ； in, is the reward, N is the number of solution steps, and S is the user score.

5. The method for handling user complaints according to claim 4, wherein: After determining the target user's feedback information regarding the final solution from the dialogue interaction process, the method further includes: Based on the feedback information, the knowledge analysis large language model is trained, and based on the training result, the knowledge analysis large language model is updated.

6. The user complaint handling method according to claim 1, characterized in that: The construction process of the knowledge graph includes: Based on keyword matching, entity information in the historical complaint handling information is identified to obtain the entity information in the historical complaint handling information; Determining associations between entities in the historical complaint handling information based on relationship rules in the historical complaint handling information, and determining relationship information based on the associations between the entities; The knowledge graph is constructed based on the entity information and the relationship information.

7. The user complaint handling method according to claim 1, characterized in that: The key information includes intention information, complaint information and demand information.

8. The method for handling user complaints according to claim 7, wherein: The first prompt word is used to guide the knowledge analysis large language model to perform intent recognition on the user complaint information, obtain the intent information of the user complaint information, and guide the knowledge analysis large language model to perform entity recognition on the user complaint information, determine the complaint problem information of the user complaint information, and guide the knowledge analysis large language model to perform demand mining on the user complaint information, and determine the demand information in the user complaint information.

9. A user complaint handling device, characterized in that: include: Information acquisition module, used to obtain user complaint information; a key information extraction module, configured to guide the knowledge analysis large language model to determine key information of the user complaint information based on the first prompt word, guide the knowledge analysis large language model to retrieve the key information in the knowledge graph based on a path sorting algorithm, and generate multiple candidate solutions based on the retrieval results, wherein the knowledge graph is constructed based on entity information and relationship information extracted from historical complaint processing information; A sentiment analysis module, configured to perform sentiment analysis on the user complaint information based on a sentiment analysis language model to determine the sentiment level of the user complaint information; A solution determination module is used to guide the knowledge analysis large language model based on the second prompt word, and determine a final solution from the multiple candidate solutions based on the sentiment level.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the user complaint handling method according to any one of claims 1 to 8 is implemented.

Images