Customer service data quality inspection method and device based on dynamic reasoning, equipment and medium

Abstract

The invention discloses a customer service data quality inspection method and device based on dynamic reasoning, equipment and a medium, and the method comprises the steps: analyzing customer service dialogue data, and generating structured dialogue data containing an intention label and a key problem node; according to an intention label in the structured dialogue data, a framework is generated through retrieval enhancement, and matched domain knowledge is recalled in real time from a standard knowledge base so as to construct a dynamic context knowledge graph; taking the dynamic context knowledge graph as input, utilizing a thinking chain prompt template to guide a large model to carry out step-by-step logical reasoning, and outputting a preliminary quality inspection conclusion chain; performing retrieval verification on assertion nodes in the initial quality inspection conclusion chain to generate a traceable final quality inspection conclusion; and based on the final quality inspection conclusion, calculating a quantitative score of the customer service dialogue data and generating a visual thinking chain report. And through dynamic retrieval of the knowledge base and logical reasoning, the accuracy, interpretability and adaptability of customer service data quality inspection are improved.

Description

Technical Field

[0001] This application relates to the technical fields of artificial intelligence and natural language processing, and particularly to a customer service data quality inspection method, device, equipment, and medium based on dynamic reasoning. Background Art

[0002] In recent years, in the field of customer service quality inspection, the existing technologies mainly rely on rule engines, keyword matching, or methods based on traditional machine learning models. These methods show certain effects when dealing with structured data and simple problems, but there are obvious limitations when facing complex semantics, multi-round conversations, and dynamically changing business requirements. For example, rule engines are difficult to handle diverse conversation contents, keyword matching methods are prone to ignoring context semantics, and traditional machine learning models require a large amount of labeled data and are difficult to adapt to new scenarios. In addition, the existing technologies also have deficiencies in knowledge update, reasoning transparency, and result interpretability, and are difficult to meet the requirements of high-quality customer service quality inspection. Summary of the Invention

[0003] The purpose of this application is to provide a customer service data quality inspection method, device, equipment, and medium based on dynamic reasoning, so as to at least solve the problems that the current customer service quality inspection system cannot meet the requirements in terms of knowledge update, reasoning transparency, and result interpretability.

[0004] To solve the above technical problems, this application provides a customer service data quality inspection method based on dynamic reasoning, including: Parsing customer service conversation data based on a pre-configured large model to generate structured conversation data containing intent tags and key question nodes; According to the intent tags in the structured conversation data, through a retrieval-augmented generation framework, retrieving and recalling matching domain knowledge from a standard knowledge base in real time to construct a dynamic context knowledge graph; Using the dynamic context knowledge graph as input, guiding the large model to perform step-by-step logical reasoning on the customer service conversation data by means of a chain of thought prompt template, and outputting a preliminary quality inspection conclusion chain containing the reasoning path; Performing multi-modal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; Based on the final quality inspection conclusion, calculating a quantitative score for the customer service conversation data and generating a visual chain of thought report. The quantitative score includes quantitative scores for at least one dimension of compliance, accuracy, and tone and emotion. The report contains error node annotations, correction paths, and service optimization suggestions for any one of the dimensions.

[0005] Optionally, the parsing includes word segmentation, entity recognition, and context semantic parsing; parsing the customer service dialogue data based on a pre-configured large model to generate structured dialogue data containing intent tags and key question nodes, including: Based on the zero-shot learning ability of the large model, pre-classify the intent of the customer service dialogue data to generate preliminary intent tags; Through context semantic reinforcement configuration, disambiguate and correct the preliminary intent tags to output multi-dimensional intent tags; Combined with the results of the entity recognition, associate the multi-dimensional intent tags with key question nodes to construct the structured dialogue data.

[0006] Optionally, the domain knowledge includes at least one of canonical texts, historical cases, and industry policies; constructing the dynamic context knowledge graph includes: Based on the semantic vectors of the intent tags, recall canonical texts in the standard knowledge base using cosine similarity matching; In the historical case library, screen the historical cases with the highest similarity to the dialogue scenario of the customer service dialogue data based on the time weight retrieval strategy; Fuse the canonical texts and historical cases to generate a dynamic context knowledge graph.

[0007] Optionally, through the retrieval-enhanced generation framework, recall matching domain knowledge from the standard knowledge base in real time to construct a dynamic context knowledge graph, including: Determine the timeliness requirements according to the dialogue scenario of the customer service dialogue data; Based on the timeliness requirements, perform time decay weighting on the retrieval results recalled in real time from the standard knowledge base, and preferentially recall the domain knowledge updated recently to construct a dynamic context knowledge graph.

[0008] Optionally, the step-by-step logical reasoning includes compliance verification, knowledge accuracy verification, and sentiment polarity analysis; using the thought chain prompt template to guide the large model to perform step-by-step logical reasoning on the customer service dialogue data, and output a preliminary quality inspection conclusion chain containing the reasoning path, including: Through the compliance rule chain in the prompt template, layer by layer verify whether the customer service responses in the customer service dialogue data comply with industry norms; Based on the historical case similarity, evaluate the accuracy score of the customer service responses in the customer service dialogue data; Use the sentiment analysis model to quantify the sentiment polarity of the customer service responses in the customer service dialogue data, and compare it with the preset threshold to generate a sentiment score; Output the preliminary quality inspection conclusion chain containing the reasoning path of compliance verification, knowledge accuracy verification, and sentiment polarity analysis.

[0009] Optionally, the multimodal retrieval verification includes verification of at least one of text, a rule base, a historical case base, or manually annotated data. The multimodal retrieval verification of the assertion nodes in the preliminary quality inspection conclusion chain includes: Based on the assertion nodes in the preliminary quality inspection conclusion chain, extract the assertion content, perform alignment verification with the structured text specifications in the rule base through a semantic matching algorithm, calculate the semantic similarity score. If the semantic similarity score is lower than the preset threshold, mark the assertion node as a potential logical conflict node; According to the context scenario of the assertion node, recall similar dialogue cases from the historical case base, and compare the consistency between the current assertion conclusion and the historical processing result through a time-weighted strategy and scenario similarity. If the consistency meets the condition of significant difference, mark the assertion node as a conflict warning node.

[0010] Optionally, the method further includes: Based on the multi-dimensional scoring results, through a preset business impact factor model, calculate the optimization priority labels for each dimension, and use a policy mapping algorithm to associate the scoring defects with a preset policy library to generate an initial optimization policy matrix; According to the priority labels in the optimization policy matrix, combined with the real-time dialogue scenario features, recall the adapted optimization rule set from the policy library, and dynamically adjust the policy weights through a rule engine to output a scenario-based optimization policy execution list; Feed back the optimization policy execution list to the corresponding customer service system.

[0011] To solve the above technical problems, the present application also provides a customer service data quality inspection device based on dynamic reasoning, including: A data parsing module, configured to parse customer service dialogue data based on a pre-configured large model to generate structured dialogue data including intent labels and key problem nodes; A graph construction module, configured to, according to the intent labels in the structured dialogue data, recall matching domain knowledge from a standard knowledge base in real time through a retrieval-enhanced generation framework to construct a dynamic context knowledge graph; A logical reasoning module, configured to use the dynamic context knowledge graph as an input, and use a thought chain prompt template to guide the large model to perform step-by-step logical reasoning on the customer service dialogue data, and output a preliminary quality inspection conclusion chain including an inference path; A retrieval verification module, configured to perform multimodal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; A quantization output module, configured to calculate a quantization score of customer service dialogue data based on the final quality inspection conclusion and generate a visualized thought chain report. The quantization score includes quantization scores of at least one dimension of compliance, accuracy, and tone and emotion. The report includes error node annotations, correction paths, and service optimization suggestions for any one of the dimensions.

[0012] To solve the above technical problems, the present application also provides a computer device, including a memory and a processor. Computer-readable instructions are stored in the memory. When the computer-readable instructions are executed by the processor, the processor executes the steps of the above-mentioned customer service data quality inspection method based on dynamic reasoning.

[0013] To solve the above technical problems, the present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the above-mentioned customer service data quality inspection method based on dynamic reasoning.

[0014] The beneficial effects of the embodiments of the present application are as follows: By parsing customer service dialogue data, structured dialogue data including intent tags and key problem nodes is generated; According to the intent tags in the structured dialogue data, through the Retrieval-Augmented Generation (RAG) framework, matching domain knowledge is recalled in real time from the standard knowledge base to construct a dynamic context knowledge graph; Taking the dynamic context knowledge graph as the input, using the thought chain (CoT) prompt template to guide the large model to perform step-by-step logical reasoning, and output a preliminary quality inspection conclusion chain; Retrieving and verifying the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; Based on the final quality inspection conclusion, calculate the quantization score of the customer service dialogue data and generate a visualized thought chain report. Based on the deep semantic understanding and dynamic knowledge retrieval of the large model, accurate parsing and quality inspection of customer service conversations are realized, and the problem of insufficient understanding of complex semantics by traditional methods is solved. And by combining thought chain reasoning and multi-modal verification, an interpretable quality inspection conclusion is generated, improving the transparency and accuracy of quality inspection. At the same time, multi-dimensional scoring and visualization reports provide clear improvement directions for the customer service team, support the self-optimization of the system, and significantly improve the quality of customer service and management efficiency. Description of the Drawings

[0015] The above and / or additional aspects and advantages of the present application will become apparent and easy to understand from the following description of the embodiments in conjunction with the drawings, where: Figure 1 It is a schematic diagram of the basic process of the customer service data quality inspection method based on dynamic reasoning for a specific embodiment of the present application; Figure 2 It is a schematic diagram of the basic structure of the customer service data quality inspection device based on dynamic reasoning for a specific embodiment of the present application; Figure 3 It is the basic structural block diagram of a computer device according to a specific embodiment of the present application. Specific Embodiments

[0016] The embodiments of the present application will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application and should not be construed as a limitation to the present application.

[0017] Those skilled in the art of the present technology can understand that unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present application means the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups.

[0018] Those skilled in the art of the present technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the field to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless specifically defined as here.

[0019] Those skilled in the art can understand that the "terminal" used herein includes both a device with a wireless signal receiver that only has the ability to receive without transmitting, and a device with receiving and transmitting hardware that has the receiving and transmitting hardware capable of performing two-way communication on a two-way communication link. Such devices may include: cellular or other communication devices, which may have a single-line display or a multi-line display or a cellular or other communication device without a multi-line display; PCS (Personal Communications Service), which may combine voice, data processing, fax, and / or data communication capabilities; PDA (Personal Digital Assistant), which may include a radio frequency receiver, a pager, Internet / intranet access, a web browser, a notepad, a calendar, and / or a GPS (Global Positioning System) receiver; conventional laptop and / or palm computers or other devices, which are conventional laptop and / or palm computers or other devices with and / or including a radio frequency receiver. The "terminal" used herein may be portable, transportable, installed in a vehicle (air, sea, and / or land), or suitable for and / or configured to operate locally, and / or operate in a distributed manner at any other location on the earth and / or in space. The "terminal" used herein may also be a communication terminal, an Internet access terminal, a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback functions, or may also be a smart TV, a set-top box, and other devices.

[0020] The hardware referred to by names such as "server", "client", and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer, which is a hardware device with the necessary components disclosed by the von Neumann principle, including a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device. The computer program is stored in its memory, and the central processing unit loads the program stored in the external memory into the internal memory for execution, executes the instructions in the program, and interacts with the input / output devices to complete specific functions.

[0021] It should be noted that the concept of "server" in this application can similarly be extended to the case of server clusters. According to the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can either be independent of each other but can be called through interfaces, or integrated into a single physical computer or a set of computer clusters. Those skilled in the art should understand this flexibility and should not be restricted by this when implementing the network deployment method of this application.

[0022] One or several technical features of this application, unless explicitly specified, can either be deployed on the server and accessed by the client remotely invoking the online service interface provided by the server, or directly deployed and run on the client for access.

[0023] The AI models cited or possibly cited in this application, unless explicitly specified, can either be deployed on a remote server and remotely invoked on the client, or deployed on a client capable of handling the device capabilities for direct invocation. In some embodiments, when it runs on the client, its corresponding intelligence can be obtained through transfer learning to reduce the requirements for the client's hardware operating resources and avoid over-occupying the client's hardware operating resources.

[0024] All kinds of data involved in this application, unless explicitly specified, can either be remotely stored on the server or stored on the local terminal device, as long as it is suitable for being invoked by the technical solution of this application.

[0025] Those skilled in the art should be aware of this: Although the various methods of this application are described based on the same concept and thus show commonality with each other, unless otherwise specified, these methods can all be executed independently. Similarly, for each of the embodiments disclosed in this application, they are all proposed based on the same inventive concept. Therefore, for concepts with the same expression, as well as concepts that are only appropriately transformed for convenience although the concept expressions are different, they should be equally understood.

[0026] For each of the embodiments to be disclosed in this application, unless explicitly stated that there is a mutually exclusive relationship between them, otherwise, the relevant technical features involved in each embodiment can be cross-combined to flexibly construct new embodiments, as long as this combination does not deviate from the creative spirit of this application and can meet the needs in the prior art or solve certain deficiencies in the prior art. Those skilled in the art should be aware of this flexibility.

[0027] Please refer to Figure 1 , Figure 1 which is the basic process schematic diagram of the customer service data quality inspection method based on dynamic inference for this embodiment.

[0028] As Figure 1 shown, it includes: S1100. Analyze the customer service conversation data based on a pre-configured large model to generate structured conversation data containing intent tags and key question nodes; This embodiment is applied to the quality inspection scenarios of customer service call data and intelligent customer service systems in various fields such as finance, e-commerce, insurance, education, medical care, and law. In this embodiment, a quality inspection system equipped with a basic model (such as a large language model) is configured to dynamically detect the quality of various customer service scenarios. The quality inspection system can be connected to the customer service system to obtain the data of the customer service system, and define this data as customer service conversation data. The quality inspection system can analyze the customer service conversation data based on a pre-configured large model to generate structured conversation data containing intent tags and key question nodes. Specifically, the quality inspection system first receives the original text data of the customer service conversation, which may include the content of multiple rounds of conversations between the customer and the customer service. Then, it analyzes the text through a pre-configured large model (large language model), including processes such as word segmentation, entity recognition, and context semantic analysis of the text. Among them, word segmentation is performed first, splitting the sentences of the original text data into words or phrases. For example, the sentence "I want to consult the return policy" will be split into "I / want / to / consult / return / policy". Subsequently, entity recognition is performed to identify the key entities in the conversation. For example, "return policy" is recognized as an entity related to after-sales service. Furthermore, the large model uses its context understanding ability to perform semantic analysis on the entire conversation to judge the intent of the conversation. For example, the above sentence is classified as an "inquiry" intent, and the key question node is marked as "return policy". Finally, structured conversation data containing the intent tag "inquiry" and the key question node "return policy" is generated, providing the basic input for the subsequent quality inspection process.

[0029] It should be noted that the large model in this embodiment is configured to not only process Chinese conversations, but also perform word segmentation, entity recognition, and context semantic analysis on conversations in other languages such as English and Japanese. For example, for the English conversation "I want to return this product", the large model segments it into "I / want / to / return / this / product", recognizes "return" as the key entity, and classifies its intent as "return inquiry". In addition, the large model can also translate conversations in other languages into a preset language (such as Chinese) and then perform subsequent analysis, so as to achieve unified quality inspection of multi-language customer service data.

[0030] It should be noted that the large model in this embodiment is configured to not only analyze the conversation content of the current turn, but also combine the context information of the previous few turns of conversations for semantic understanding. For example, in a conversation, the customer first mentions "I want to know about the product functions", and then asks "How about the after-sales guarantee?". The large model will link the two turns of conversations, identify that the overall intention of the customer is "product consultation + after-sales consultation", and mark the key question nodes as "product functions" and "after-sales guarantee". In this way, the large model can more accurately capture the coherence and complexity of the conversation, and provide more comprehensive structured data for subsequent quality inspection.

[0031] S1200. According to the intention labels in the structured conversation data, through the Retrieval-Augmented Generation (RAG) framework, recall the matching domain knowledge from the standard knowledge base in real time to construct a dynamic context knowledge graph; After parsing the customer service conversation data based on a pre-configured large model to generate structured conversation data containing intention labels and key question nodes, according to the intention labels in the structured conversation data, through the Retrieval-Augmented Generation (RAG) framework, recall the matching domain knowledge from the standard knowledge base in real time to construct a dynamic context knowledge graph. Specifically, the quality inspection system extracts the intention labels (such as "consultation") and key question nodes (such as "return policy") among them. Based on this information, the Retrieval-Augmented Generation (RAG) framework recalls the matching domain knowledge from the standard knowledge base in real time. The retriever in the RAG framework will retrieve the specification texts, historical high-quality conversation cases, industry policies, etc. related to the "return policy" from the standard knowledge base. For example, the retriever may find the "Return Policy Explanation" document within the enterprise, as well as the conversation records of previous customer service representatives successfully handling return problems. These retrieved knowledge are integrated into a dynamic context knowledge graph, which contains knowledge nodes and association relationships related to the current conversation. For example, the "return policy" node may be connected to sub-nodes such as "refund process" and "requirements for product integrity". Ultimately, this dynamic context knowledge graph provides rich real-time knowledge support for the subsequent chain-of-thought quality inspection reasoning.

[0032] It should be noted that in this embodiment, when the quality inspection system receives the intention labels and key question nodes, the RAG framework not only performs keyword matching, but also utilizes the semantic understanding ability of the large model to calculate the semantic similarity between the conversation content and the knowledge in the knowledge base. For example, even if the customer consults about the "return and exchange process", and the "return policy" is stored in the knowledge base, the system can identify the association between the two through semantic similarity calculation and recall the relevant knowledge. In addition, the system can also sort the retrieval results according to the similarity score, and preferentially select the knowledge entries with the closest semantics, thereby improving the accuracy and efficiency of the retrieval.

[0033] It should be noted that the standard knowledge base of this embodiment not only includes static specification texts and historical cases, but also accesses external data sources in real time, such as industry news, policy updates, etc. For example, when new consumer protection regulations are introduced, the knowledge base can update relevant content in a timely manner. During the retrieval process, the RAG framework will give priority to retrieving the latest knowledge entries to ensure that the system can obtain domain knowledge with strong timeliness. For example, when processing a customer's consultation on the "latest privacy policy", the system can recall the recently updated privacy policy content in real time instead of the outdated version, thereby improving the accuracy and reliability of quality inspection.

[0034] It should be noted that the standard knowledge base of this embodiment is configured into multiple levels, including a general knowledge base (such as industry general policies), an enterprise knowledge base (such as enterprise internal service specifications), and a personalized knowledge base (such as special policies for specific customer groups). When the quality inspection system receives the intent label and the key question node, the RAG framework will first retrieve basic knowledge from the general knowledge base, then retrieve more specific specifications from the enterprise knowledge base, and finally retrieve knowledge related to specific customers from the personalized knowledge base. For example, for a VIP customer's consultation on the "priority service policy", the system will first find the "customer service policy" from the general knowledge base, then find the "VIP customer rights and interests" from the enterprise knowledge base, and finally find the exclusive service terms for this VIP customer from the personalized knowledge base. Through this multi-level retrieval mechanism, the system can provide more accurate personalized quality inspection support for different customer groups.

[0035] S1300. Using the dynamic context knowledge graph as the input, guiding the large model to perform step-by-step logical reasoning on the customer service conversation data by using the thought chain (CoT) prompt template, and outputting a preliminary quality inspection conclusion chain including the reasoning path. After retrieving and enhancing the generated (RAG) framework based on the intent tags in the structured conversation data to recall the matching domain knowledge from the standard knowledge base in real time and construct a dynamic context knowledge graph, using the dynamic context knowledge graph as input, the large model is guided by the chain of thought (CoT) prompt template to perform step-by-step logical reasoning on the customer service conversation data, and an initial quality inspection conclusion chain containing the reasoning path is output. Specifically, the quality inspection system uses the constructed dynamic context knowledge graph as input and guides the large model to reason using a preset CoT prompt template (such as "Please analyze step by step whether the customer service reply meets the following specifications: {retrieved content}"). For example, for a customer's inquiry about the "return policy", the large model will analyze step by step whether the customer service reply mentions key points such as "return conditions" and "refund process" according to the prompt template. The reasoning process is decomposed into multiple logical steps, and each step corresponds to a specific specification or knowledge node. For example, in the first step, it is verified whether the customer service clearly states the "requirements for product integrity", and in the second step, it is checked whether the "refund time range" is mentioned. Finally, the large model generates an initial quality inspection conclusion chain containing the reasoning path, such as "The customer service clearly states the requirements for product integrity (meets the specifications), but does not mention the refund time range (does not meet the specifications)". This step-by-step reasoning method not only improves the quality inspection accuracy but also provides a clear logical chain for subsequent correction and review.

[0036] It should be noted that this embodiment also configures a hierarchical reasoning mechanism to adjust the reasoning priority according to the importance and complexity of the problem. For example, for key issues involving customer complaints (such as "product quality problems"), the large model will give priority to detailed reasoning to ensure that each key point meets the enterprise specifications and knowledge base requirements. For general inquiries (such as "product price"), the reasoning process will be relatively simplified, and the focus will be on verifying the accuracy of the information. In addition, the system can also dynamically adjust the depth and breadth of reasoning according to the customer type (such as VIP customers) or the conversation scenario (such as high-value orders). For example, for the inquiries of VIP customers, the system will increase the reasoning steps to ensure the integrity and professionalism of the customer service reply. This hierarchical reasoning and priority adjustment mechanism can optimize resource allocation and improve the quality inspection efficiency and accuracy of the system in different scenarios.

[0037] S1400. Perform multimodal retrieval verification on the assertion nodes in the initial quality inspection conclusion chain to generate a traceable final quality inspection conclusion; After outputting the preliminary quality inspection conclusion chain containing the reasoning path, perform multimodal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion. Specifically, the quality inspection system first extracts the key assertion nodes in the preliminary quality inspection conclusion chain. An assertion node (Assertion Node) refers to a key judgment point or conclusive statement generated in the step-by-step logical reasoning process. Each node corresponds to a quality inspection logic verification result for a specific dimension (such as compliance, accuracy, sentiment, etc.), and the assertion determination is based on a dynamically retrieved knowledge base or preset rules. For example, "The customer service did not mention the refund time range in the return policy". For this assertion, the system uses a multimodal retrieval engine to verify the accuracy of the assertion from the knowledge base by combining various forms of knowledge resources such as text, tables, and pictures. For example, the retrieval engine will search for the normative text, historical conversation cases, and relevant charts (such as flowcharts) about the "refund time range" in the knowledge base. If the retrieval results show that there is a clear specification in the knowledge base and the customer service did not mention this content in the conversation, the assertion is confirmed as correct; otherwise, if the retrieval results show that the customer service has mentioned the relevant content, the system corrects the assertion. Finally, the system generates a traceable final quality inspection conclusion containing the verification process and results, ensuring that each assertion has a clear basis and correction path.

[0038] It should be noted that the assertion node in this embodiment is the smallest logical unit of the reasoning chain. For example: "Whether the customer service reply mentions the return policy" (compliance node); "Whether the data provided by the customer service is consistent with the knowledge base" (accuracy node); "Whether the customer service tone meets the positive sentiment threshold" (sentiment node). Each node contains the assertion content (judgment criterion) and the assertion result (whether it meets the requirements). By means of step-by-step assertion nodes, the "black box reasoning" of the large model is transformed into a transparent logical chain, which is convenient for tracing the source of errors. And the determination result of each node directly affects the final score. The assertion node is the core unit connecting dynamic knowledge, reasoning logic, and quality inspection conclusions, and improves the accuracy and interpretability of the system through fine-grained verification.

[0039] S1500. Based on the final quality inspection conclusion, calculate the quantitative score of the customer service conversation data and generate a visual thinking chain report. The quantitative score includes the quantitative scores of at least one dimension of compliance, accuracy, and tone sentiment. The report contains the annotation of error nodes, correction paths, and service optimization suggestions for any one of the dimensions.

[0040] After performing multimodal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain and generating a traceable final quality inspection conclusion, a quantitative score of the customer service conversation data is calculated based on the final quality inspection conclusion, and a visual thinking chain report is generated. Specifically, the quality inspection system calculates quantitative scores respectively according to dimensions such as compliance, accuracy, and emotional attitude in the final quality inspection conclusion. For example, the compliance score may be based on whether the customer service completely follows the enterprise service specifications, the accuracy score is based on whether the information provided by the customer service is consistent with the knowledge base, and the emotional attitude score is based on whether the tone and diction of the customer service are positive and friendly. Suppose a certain customer service conversation has a compliance score of 90% (most of the specifications are correctly implemented), an accuracy score of 85% (some information is missing), and an emotional attitude score of 95% (the tone is friendly). The system will integrate these scores into a comprehensive quality score. Subsequently, the system generates a visual thinking chain report, which displays the quality inspection results in the form of charts and text. Each dimension's score is clearly marked in the report, and problem nodes are highlighted. For example, "The customer service did not mention the refund time range" is marked in red, and at the same time, a correction path and improvement suggestions are provided, such as "It is recommended to supplement the description of the refund time range in subsequent conversations". This visual report not only helps managers quickly understand the overall quality of customer service conversations but also provides clear improvement directions for customer service staff, supporting the continuous optimization of the system.

[0041] It should be noted that the quality inspection system in this embodiment is also configured with a multi-dimensional scoring mechanism, which dynamically adjusts the weights of each dimension to adapt to different business requirements and conversation scenarios. For example, when handling customer complaints, the weight of the compliance dimension can be increased because the standardization of complaint handling directly affects customer satisfaction; while when answering technical questions, the weight of the accuracy dimension can be increased because the correctness of information is crucial. The system can dynamically allocate weights according to preset business rules or real-time analysis of conversation content. For example, for conversations involving high-value orders, the weights of the compliance and accuracy dimensions can each be increased by 10%, while the weight of the emotional attitude dimension is appropriately reduced. In this way, the system can more accurately reflect the quality of customer service conversations in different scenarios, generating more targeted quantitative scores and improvement suggestions.

[0042] Furthermore, the quality inspection system of this embodiment is also configured with personalized feedback and intelligent recommendation functions to further enhance the value of the quality inspection report. The system generates personalized improvement suggestions for each customer service staff according to the final quality inspection conclusion. For example, if a certain customer service staff scores low in terms of emotional attitude, the system will recommend relevant communication skills training resources, such as online courses or simulated dialogue exercises. In addition, the system can also intelligently recommend optimized dialogue templates according to the type of problem nodes. For example, if the customer service misses key information when explaining the product function, the system will provide a verified dialogue template containing all necessary information for the customer service to refer to. This personalized feedback and intelligent recommendation function not only helps the customer service staff to improve quickly, but also improves the overall intelligence level of the system, supporting the self-optimization and growth of the customer service team.

[0043] In the above implementation manner, by parsing the customer service dialogue data, structured dialogue data including intent tags and key problem nodes is generated; according to the intent tags in the structured dialogue data, through the Retrieval-Augmented Generation (RAG) framework, the matching domain knowledge is retrieved in real time from the standard knowledge base to construct a dynamic context knowledge graph; taking the dynamic context knowledge graph as the input, using the Chain of Thought (CoT) prompt template to guide the large model to perform step-by-step logical reasoning, and outputting a preliminary quality inspection conclusion chain; retrieving and verifying the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; based on the final quality inspection conclusion, calculating the quantitative score of the customer service dialogue data and generating a visual Chain of Thought report. Based on the deep semantic understanding and dynamic knowledge retrieval of the large model, the accurate parsing and quality inspection of the customer service dialogue are realized, solving the problem of insufficient understanding of complex semantics by traditional methods. And by combining Chain of Thought reasoning and multi-modal verification, an interpretable quality inspection conclusion is generated, improving the transparency and accuracy of the quality inspection. At the same time, the multi-dimensional scoring and visual report provide a clear improvement direction for the customer service team, supporting the self-optimization of the system, and significantly improving the quality of customer service and management efficiency.

[0044] In some implementation manners, the parsing includes word segmentation, entity recognition, and context semantic parsing; step S1100 includes: S1111. Based on the zero-shot learning ability of the large model, perform intent pre-classification on the customer service dialogue data to generate preliminary intent tags; In this embodiment, parsing the customer service dialogue data includes word segmentation, entity recognition, and context semantic parsing. Among them, the zero-shot learning ability of the large model is used to pre-classify the intent of the customer service dialogue data to generate preliminary intent labels. Specifically, the system receives the original text data of the customer service dialogue. For example, the customer asks: "I want to know about the product warranty policy." Through its pre-trained language understanding and pattern recognition capabilities, the large model directly analyzes the dialogue content without additional labeled samples. It identifies the keyword "warranty policy" and, combined with the context semantics, determines that the intent of this dialogue belongs to the "consultation" category and generates a preliminary intent label of "consultation - warranty policy". This zero-shot learning method enables the system to quickly adapt to new dialogue scenarios and question types, and complete preliminary intent classification without a large amount of labeled data, providing a basis for subsequent disambiguation and correction.

[0045] It should be noted that in this embodiment, the system can also be based on zero-shot learning of industry templates. When the system performs intent pre-classification, it combines predefined industry templates (such as common question templates in industries such as e-commerce, finance, and healthcare) to help the large model better understand the dialogue content; combines zero-shot learning of user portraits, and when receiving customer service dialogue data, simultaneously obtains the customer's background information (such as purchase history, membership level, etc.); and zero-shot learning of multi-language dialogues.

[0046] S1112. Disambiguate and correct the preliminary intent label through context semantic reinforcement configuration, and output a multi-dimensional intent label; Furthermore, the system also disambiguates and corrects the preliminary intent label through context semantic reinforcement configuration, and outputs a multi-dimensional intent label. Specifically, after generating the preliminary intent label, the system further analyzes the context semantics of the dialogue. For example, the preliminary intent label is "consultation - product information", but the dialogue content is "I want to know if the product is suitable for the elderly". Through the context semantic reinforcement configuration, the system identifies the key context information of "used by the elderly" and corrects the preliminary intent label to "consultation - product suitability (for the elderly)". In addition, the system can also further enrich the intent label by combining multi-modal information such as intonation and emojis. For example, if the customer uses a questioning tone or adds a "?" emoji in the dialogue, the system will further refine the intent label to "consultation (questioning tone) - product suitability (for the elderly)". This context semantic reinforcement configuration can effectively eliminate the ambiguity of the intent label, generate more accurate and richer multi-dimensional intent labels, and provide a more accurate basis for subsequent dialogue processing.

[0047] It should be noted that the quality inspection system in this embodiment also eliminates the ambiguity of the intent label based on the dialogue history and in combination with the domain knowledge base.

[0048] S1113. Combine the results of the entity recognition, associate the multi-dimensional intent tags with the key question nodes, and construct the structured dialogue data; Further, the system combines the results of the entity recognition, associates the multi-dimensional intent tags with the key question nodes, and constructs the structured dialogue data. Specifically, after the system completes the correction of the intent tags, it further extracts the entity information in the dialogue. For example, the dialogue content is "I want to know about the repair policy of iPhone 14". After entity recognition, the system recognizes "iPhone 14" as the product entity and "repair policy" as the question entity. Subsequently, the system associates the multi-dimensional intent tag "Consultation - Product Repair Policy (for iPhone 14)" with these key entities to generate structured dialogue data. The final output structured data includes the intent tag, key question nodes (such as product model, question type), and the corresponding entity information, for example: {Intent tag: Consultation, Question node: Repair policy, Product entity: iPhone 14}. This structured data not only clearly shows the core content of the dialogue but also provides a standardized input for subsequent quality inspection and analysis.

[0049] It should be noted that the system in this embodiment also optimizes the association relationship between entities and intents by introducing a domain knowledge base. For example, the dialogue content is "I want to ask about the lens configuration of this camera". The initial intent tag is "Consultation - Product Information", and the entity recognition results are "camera" and "lens configuration". The system retrieves the domain knowledge base and recognizes that "lens configuration" is a key feature of the camera product and belongs to the category of technical parameters. Therefore, the system refines the intent tag to "Consultation - Camera Lens Configuration (Technical Parameter)" and closely associates the entity with the intent. The final generated structured data is: {Intent tag: Consultation, Question node: Lens configuration, Product entity: Camera}. By introducing the domain knowledge base, the system can more accurately identify and associate entities with intents, improving the professionalism and accuracy of the structured data.

[0050] This embodiment quickly generates preliminary intent tags through zero-shot learning ability, performs disambiguation and correction by combining context semantic reinforcement configuration, and constructs the structured dialogue data by associating the entity recognition results. This process significantly improves the accuracy and efficiency of intent recognition, and at the same time provides a more accurate and richer data basis for subsequent customer service dialogue processing and quality inspection, enhancing the intelligent level and adaptability of the system.

[0051] In some embodiments, the domain knowledge includes at least one of normative texts, historical cases, and industry policies; S1200 constructs a dynamic context knowledge graph, including: S1211. Based on the semantic vector of the intent tag, recall the normative text in the standard knowledge base using cosine similarity matching; When constructing the dynamic context knowledge graph in this embodiment, first, based on the semantic vector of the intent label, canonical texts are recalled by using cosine similarity matching in the standard knowledge base. Specifically, the system first converts the intent label (such as "Consultation - Return Policy") into a semantic vector, and uses a pre-trained language model (such as BERT) to extract its semantic features. Then, the canonical texts stored in the standard knowledge base (such as the "Return Policy Description" within the enterprise) are vectorized, and the cosine similarity between the semantic vector of each canonical text and the intent label vector is calculated. For example, if the semantic vector of the intent label "Consultation - Return Policy" has the highest similarity with the vector of the "Return Policy Description", the system recalls this canonical text. Through cosine similarity matching, the system can efficiently find the knowledge most relevant to the current conversation intent from a large number of canonical texts, providing an accurate canonical basis for subsequent quality inspection reasoning.

[0052] It should be noted that when performing cosine similarity matching based on the semantic vector of the intent label, a keyword enhancement mechanism is combined. In addition to using the semantic vector of the intent label, the system also extracts key entities in the conversation (such as "return" and "refund") as keywords. When recalling canonical texts, the system not only calculates the cosine similarity of the semantic vectors but also weights the matching degree of the keywords. For example, texts containing keywords such as "return" and "refund" in the canonical text will be given higher weights, thereby improving the accuracy of recall. This method of cosine similarity matching combined with keyword enhancement can better handle the situation where the semantics of the intent label are not completely clear, further enhancing the precision of recall.

[0053] S1212. In the historical case library, based on the time-weighted retrieval strategy, screen the historical case with the highest similarity to the conversation scenario of the customer service conversation data; After or synchronously with the matching and recall of the canonical text, the quality inspection system of this embodiment screens the historical case with the highest similarity to the conversation scenario of the customer service conversation data in the historical case library. Specifically, the system first extracts the key features of the customer service conversation, including intent labels, key problem nodes, and conversation scenarios (such as consultation, complaint, after-sales, etc.). Then, the system calculates the similarity for each case in the historical case library, and at the same time introduces a time-weight factor. The time-weight factor is adjusted according to the generation time of the case, and more recent cases have higher weights because they are more likely to reflect the current business processes and service specifications. For example, if the current conversation scenario is "Customer complains about product failure", the system will first retrieve similar complaint cases within the last three months and calculate their similarity to the current conversation. Through this time-weighted retrieval strategy, the system can quickly find the historical case that best matches the current conversation scenario and has strong timeliness, providing strong support for subsequent knowledge fusion.

[0054] It should be noted that the system in this embodiment is also configured to dynamically adjust the time weight factor to optimize the retrieval effect of historical cases. The system dynamically adjusts the calculation method of the time weight factor according to business requirements and the timeliness of historical data. For example, in scenarios where the business is updated frequently (such as during e-commerce promotion activities), the system will increase the weight of recent cases to ensure that the retrieval results reflect the latest business specifications; while in scenarios where the business is relatively stable (such as regular after-sales service), the system will appropriately relax the time weight and allow the retrieval of earlier cases. This dynamic adjustment mechanism enables the system to better adapt to different business environments, ensuring that the retrieval results of historical cases are both timely and can make full use of the rich resources in the knowledge base.

[0055] S1213. Perform knowledge fusion on the specification text and historical cases to generate a dynamic context knowledge graph.

[0056] After recalling the specification text and historical cases, perform knowledge fusion on the specification text and historical cases to generate a dynamic context knowledge graph. Specifically, the system first extracts the key information in the specification text (such as policy terms, operation processes) and the key nodes in the historical cases (such as problem descriptions, solutions, customer service scripts). For example, the specification text mentions that "when a customer complains about a product failure, a response should be made within 48 hours", and the historical case records that "a customer complained about a product failure, and the customer service responded within 2 hours and provided a solution". The system fuses this information through semantic analysis and association rules. For example, it associates "response within 48 hours" in the specification text with "response within 2 hours" in the historical case to form a knowledge node "response time". At the same time, the system also combines the solutions in the historical cases with the operation processes in the specification text to form a complete knowledge chain. Finally, the generated dynamic context knowledge graph not only contains the authoritative information of the specification text but also combines the practical operation experience of historical cases, providing rich context support for subsequent quality inspection reasoning. This knowledge fusion method can effectively improve the practicality and accuracy of the knowledge graph, enabling it to better serve the quality inspection work of customer service conversations.

[0057] This embodiment accurately recalls the specification text through semantic vector matching, combines the time weight strategy to screen highly similar historical cases, and fuses the two to generate a dynamic context knowledge graph. This process significantly improves the efficiency and accuracy of knowledge retrieval, provides richer and more timely knowledge support for customer service quality inspection, enhances the intelligence level and adaptability of the system, and ensures the reliability and practicality of the quality inspection results.

[0058] In some embodiments, S1200 constructs a dynamic context knowledge graph by real-time recalling matching domain knowledge from a standard knowledge base through a Retrieval-Augmented Generation (RAG) framework, including: S1221. Determine the timeliness requirements according to the dialogue scenario of the customer service dialogue data; In this embodiment, when constructing the dynamic context knowledge graph, the timeliness requirements are determined according to the dialogue scenario of the customer service dialogue data. Specifically, the system first analyzes the key features of the customer service dialogue, such as intent tags, question nodes, and dialogue topics. For example, if the dialogue scenario is "customer inquires about the latest promotion activities", the system recognizes that this scenario has a high timeliness requirement because the content of the promotion activities may be updated at any time. The system will set the timeliness requirement to "high" and mark that knowledge updated within the last month needs to be recalled preferentially. On the contrary, if the dialogue scenario is "customer inquires about the basic functions of the product", the system recognizes that this scenario has a relatively low timeliness requirement because the basic functions of the product are relatively stable. The timeliness requirement can be set to "medium", allowing knowledge updated within the last three months to be recalled. In this way, the system can flexibly adjust the knowledge recall strategy according to the timeliness requirements of different dialogue scenarios to ensure that the recalled knowledge matches the timeliness of the current dialogue.

[0059] It should be noted that the system in this embodiment is also configured to consider not only the characteristics of the dialogue scenario itself but also factors such as industry policy changes and market dynamics. For example, in the financial industry, regulatory policies may be updated at any time. Therefore, for dialogue scenarios related to financial regulations (such as "customer inquires about the compliance of financial products"), the system will automatically set the timeliness requirement to "highest" and preferentially recall knowledge updated within the last week. By introducing industry dynamic information, the system can more accurately judge the timeliness requirements of the dialogue scenario to ensure that the recalled knowledge is always the latest in the case of frequent industry changes.

[0060] It should be noted that the system in this embodiment is also configured to adjust the timeliness requirements according to the urgency of the dialogue. The system judges the urgency of the dialogue by analyzing the keywords in the dialogue (such as "urgent", "right away", "immediately") or the semantic features of the dialogue. For example, if the customer mentions in the dialogue "I urgently need to know the latest price of this product", the system will recognize that the urgency of the dialogue is relatively high, set the timeliness requirement to "high", and preferentially recall the price information updated within the last week. On the contrary, if the dialogue does not show urgency, the system can appropriately relax the timeliness requirements and recall more extensive knowledge. This adjustment mechanism based on the urgency of the dialogue can ensure that the system provides the latest knowledge support quickly when processing urgent dialogues.

[0061] S1222. Based on the timeliness requirements, perform time decay weighting on the retrieval results retrieved in real time from the standard knowledge base, and preferentially recall the domain knowledge updated recently to construct a dynamic context knowledge graph.

[0062] After determining the timeliness requirements based on the dialogue scenario of the customer service dialogue data, based on the timeliness requirements, time decay weighting is performed on the retrieval results retrieved in real time from the standard knowledge base, and domain knowledge updated recently is preferentially retrieved to construct a dynamic context knowledge graph. Specifically, the system sorts and weights the retrieval results in the knowledge base according to the timeliness requirements (such as "high" or "medium") determined in step S1211. For example, if the timeliness requirement is "high", the system assigns a higher weight to the knowledge updated within the last month, and the weight gradually decays over time. Suppose there are two records about "product price" in the knowledge base: one was updated last week with a weight of 0.9; the other was updated three months ago with a weight of 0.3. The system will preferentially retrieve the knowledge with a higher weight (i.e., the record updated last week) and incorporate it into the dynamic context knowledge graph. Through time decay weighting, the system can ensure that the latest knowledge is preferentially used when constructing the knowledge graph, thereby improving the timeliness and accuracy of the knowledge graph.

[0063] It should be noted that the quality inspection system of this embodiment is also configured with a segmented time decay strategy to weight the retrieval results. The system divides the time range into multiple segments according to the timeliness requirements, such as "last week", "last month", "last three months", etc., and sets different weights for each segment. Taking "high timeliness requirement" as an example, the weight of the knowledge within the last week is 0.9, the weight of the knowledge within the last month is 0.7, and the weight of the knowledge within the last three months is 0.5. This segmented time decay strategy can more finely control the newness and oldness of knowledge, ensure that the knowledge retrieved within different time ranges has appropriate weights, and thus improve the dynamics and adaptability of the knowledge graph.

[0064] This embodiment constructs a more timely and dynamic context knowledge graph by accurately determining the timeliness requirements of the dialogue scenario and performing time decay weighting on the retrieved knowledge. This process significantly improves the accuracy and practicality of knowledge retrieval, ensures that the quality inspection system can reason based on the latest knowledge, thereby improving the accuracy and reliability of the quality inspection results, and enhancing the adaptability and intelligence level of the system.

[0065] In some embodiments, the step-by-step logical reasoning includes compliance verification, knowledge accuracy verification, and sentiment polarity analysis; S1300 uses a chain of thought (CoT) prompt template to guide the large model to perform step-by-step logical reasoning on the customer service dialogue data, and outputs a preliminary quality inspection conclusion chain including the reasoning path, including: S1311: Through the compliance rule chain in the CoT prompt template, layer by layer verify whether the customer service responses in the customer service dialogue data comply with industry norms; In this embodiment, when using the Chain of Thought (CoT) prompt template to guide the large model to perform step-by-step logical reasoning on the customer service dialogue data and output a preliminary quality inspection conclusion chain containing the reasoning path, the compliance rule chain in the CoT prompt template is used to verify layer by layer whether the customer service responses in the customer service dialogue data comply with industry norms. Specifically, the system first calls a preset CoT prompt template according to the intent label and key question nodes of the customer service dialogue. For example, for a dialogue where the customer consults about the "product return and exchange policy", the prompt template will include a series of compliance rule chains, such as "whether the return and exchange conditions are clearly mentioned" and "whether the refund process and time limit are stated". The system uses the reasoning ability of the large model to verify the content of the customer service response layer by layer according to the rule chain. For example, the first layer of the rule chain verifies whether the customer service clearly states the "requirements for the integrity of the goods", and the second layer verifies whether the "refund time range" is mentioned. If the customer service response fully complies with all the terms in the rule chain, the compliance verification passes; if there are omissions or errors, it is marked as non-compliant, and the specific non-compliant terms are recorded. Finally, the system generates a preliminary quality inspection conclusion chain containing the compliance verification path, such as: "The customer service clearly stated the requirements for the integrity of the goods (complies with the specification), but did not mention the refund time range (does not comply with the specification)". This layer-by-layer verification method not only ensures the compliance of the customer service response but also provides a clear logical path for subsequent quality inspections.

[0066] S1312: Based on the historical case similarity, evaluate the accuracy score of the customer service response in the customer service dialogue data; In this embodiment, the system also evaluates the accuracy score of the customer service replies in the customer service conversation data based on the similarity of historical cases. Specifically, the system first extracts the key information in the customer service conversation, including the problem description, the content of the customer service's reply, and the conversation scenario. For example, when a customer consults about "how to repair a product failure", the customer service replies with a detailed repair process and contact information. The system will compare this conversation with similar cases in the historical case library. The historical case library stores a large number of labeled customer service conversation samples, each of which contains a problem description, a customer service reply, and an accuracy score labeled by an expert. The system calculates the similarity between the current conversation and the historical cases to find the case most similar to the current conversation. The similarity calculation can be based on multiple dimensions such as the semantic similarity of the text content, the problem type, and the conversation scenario. For example, the system may find a historical case where the customer consulted a similar product failure problem, the content of the customer service's reply is highly similar to the current conversation, and the accuracy score of this historical case is 90%. The system generates an accuracy score for the reply of the current customer service conversation based on the accuracy score of the historical case. If the similarity between the current conversation and the historical case is very high (e.g., more than 80%), the system can directly use the accuracy score of the historical case as the accuracy score of the current conversation. If the similarity is low, the system will adjust the score according to the similarity ratio. For example, if the similarity is 60%, the accuracy score of the current conversation may be 60% of the historical case score. Finally, the system outputs the accuracy score of the current customer service conversation as part of the preliminary quality inspection conclusion chain, providing a basis for subsequent comprehensive evaluation.

[0067] S1313: Quantify the sentiment polarity of the customer service reply in the customer service conversation data using a sentiment analysis model, and generate a sentiment score by comparing it with a preset threshold.

[0068] In this embodiment, the system also uses a sentiment analysis model to quantify the sentiment polarity of the customer service responses in the customer service conversation data, and compares it with a preset threshold to generate a sentiment score. Specifically, the system first extracts the text content in the customer service conversation, including the customer's questions and the customer service responses. For example, when the customer consults "The product has a malfunction", the customer service responds "We are very sorry for the inconvenience caused to you. We will arrange maintenance for you as soon as possible." The system inputs this conversation into a pre-trained sentiment analysis model. The sentiment analysis model is based on natural language processing technology and can identify the sentiment tendency in the text, such as positive, negative or neutral. The model will analyze the keywords, tone words and overall semantics in the customer service response. For example, words such as "very sorry" and "arrange as soon as possible" will show a positive service attitude. The model will output a sentiment polarity score, usually ranging from -1 to 1, where -1 indicates very negative, 1 indicates very positive, and 0 indicates neutral. The system compares the sentiment polarity score with the preset threshold to generate a sentiment score. For example, the preset positive sentiment threshold is 0.5, the neutral sentiment threshold is 0, and the negative sentiment threshold is -0.5. If the sentiment polarity score is 0.7, it means that the customer service response has a strong positive sentiment, and the system will give a higher sentiment score, such as 90 points; if the sentiment polarity score is -0.3, it means that the customer service response has a slight negative sentiment, and the system will give a lower sentiment score, such as 60 points. Finally, the system incorporates the sentiment score into the preliminary quality inspection conclusion chain to provide a quantitative basis for the overall quality assessment of the customer service conversation in terms of sentiment dimension. This quantification method based on the sentiment analysis model can effectively evaluate the service attitude of the customer service, help managers quickly identify potential service problems, and improve customer satisfaction.

[0069] S1314: Output the preliminary quality inspection conclusion chain including the inference paths of compliance verification, knowledge accuracy verification and sentiment polarity analysis.

[0070] In this embodiment, the quality inspection system integrates the results of compliance verification, knowledge accuracy verification and sentiment polarity analysis, and outputs a preliminary quality inspection conclusion chain including the inference path. Specifically, the system summarizes the evaluation results of the first three steps. For example, the compliance verification result shows that the customer service response fully complies with industry norms; the knowledge accuracy verification score is 85 points, indicating that the response content is relatively accurate but there are a few details missing; the sentiment polarity analysis score is 90 points, indicating that the customer service attitude is positive. The system integrates this information into the preliminary quality inspection conclusion chain: "The customer service response has good compliance, the accuracy score is 85 points (some details need to be supplemented), and the sentiment polarity is positive (score 90 points)". At the same time, the system marks the details that need to be improved, such as "It is recommended to supplement the specific time range for product malfunction maintenance". This output method of the comprehensive inference path provides a clear basis for the subsequent quality inspection and review, helps to quickly locate problems and guide the customer service to improve.

[0071] In some embodiments, the multimodal retrieval verification includes verification of at least one of text, a rule base, a historical case base, or manually annotated data; S1400 performs multimodal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain, including: S1411: Based on the assertion nodes in the preliminary quality inspection conclusion chain, extract the assertion content, perform alignment verification with the structured text specifications in the rule base through a semantic matching algorithm, calculate the semantic similarity score. If the semantic similarity score is lower than a preset threshold, mark this assertion node as a potential logical conflict node; In this embodiment, the multimodal retrieval verification includes verification of text and a rule base. Specifically, based on the assertion nodes in the preliminary quality inspection conclusion chain, extract the assertion content and perform alignment verification with the structured text specifications in the rule base through a semantic matching algorithm. For example, an assertion node in the preliminary quality inspection conclusion chain is "The customer service did not mention the detailed process of product return and exchange". The system first extracts the core content of this assertion, which is "the detailed process of product return and exchange". Subsequently, the system invokes the semantic matching algorithm to perform alignment verification between this assertion content and the structured text specifications regarding "return and exchange process" in the rule base. The structured text specifications in the rule base are preprocessed and contain clear terms and descriptions. For example, there may be a specification in the rule base: "The customer service must clearly state the conditions for applying for return and exchange, the submission method, and the processing time". The system calculates the semantic similarity score between the assertion content and the rule base specification through the semantic matching algorithm. The semantic matching algorithm is usually based on techniques such as word embeddings and cosine similarity, which can quantify the semantic relevance between texts. Assume that the preset semantic similarity threshold is 0.7, and the semantic similarity score calculated by the system is 0.65, which is lower than the preset threshold. This indicates that there is a certain semantic deviation between the assertion content and the specification in the rule base, and the system marks this assertion node as a potential logical conflict node. This marking mechanism provides clues for subsequent further verification, helps the system identify possible errors or omissions, and ensures the accuracy and reliability of the preliminary quality inspection conclusion.

[0072] S1412: According to the context scenario of the assertion node, recall similar dialogue cases from the historical case base, and compare the consistency between the current assertion conclusion and the historical processing result through a time-weighted strategy and scenario similarity. If the consistency meets the condition of significant difference, mark this assertion node as a conflict warning node.

[0073] In this embodiment, the multimodal retrieval verification includes the verification of historical cases. Specifically, according to the context scenario of the assertion node, similar dialogue cases are recalled from the historical case library, and the consistency between the current assertion conclusion and the historical processing result is compared through a time-weighting strategy and scenario similarity. For example, an assertion node in the preliminary quality inspection conclusion chain is "The customer service did not mention the detailed process of product return and exchange". The system first extracts the context scenario of this assertion node, including the dialogue topic (return and exchange), the customer's intention (consulting the process), and the key information in the dialogue (such as the product model). Subsequently, the system recalls dialogue cases similar to the current scenario from the historical case library. The historical case library stores a large number of annotated customer service dialogue samples, each of which contains the dialogue content, the processing result, and the accuracy evaluation annotated by experts. The system calculates the similarity between the current dialogue scenario and the historical cases to find the case that best matches the current dialogue. The similarity calculation can be based on multiple dimensions such as the dialogue topic, the customer's intention, and the key information. For example, the system may find a historical case where the customer consulted a similar product return and exchange problem and the customer service detailed the return and exchange process in the reply. To further verify the accuracy of the current assertion conclusion, the system adopts a time-weighting strategy, assigning higher weights to recent cases. For example, cases within the last month have a weight of 0.9, cases within the last three months have a weight of 0.7, and cases over three months old have a weight of 0.5. The system compares the current assertion conclusion with the processing results of historical cases and calculates a consistency score. If the consistency score is lower than a preset significant difference threshold (e.g., 0.6), it indicates that there is a significant difference between the current assertion conclusion and the historical processing result, and the system marks this assertion node as a conflict warning node.

[0074] In this way, the system not only utilizes the empirical knowledge of historical cases but also combines the time-weighting strategy, ensuring that the verification process takes into account both the richness of historical data and highlights the timeliness and relevance of recent cases. This comprehensive verification mechanism can effectively identify potential errors or anomalies and improve the accuracy and reliability of the quality inspection conclusion.

[0075] In some embodiments, the method further includes: S1511. Based on the multi-dimensional scoring results, through a preset business impact factor model, calculate the optimization priority labels for each dimension, and use a policy mapping algorithm to associate the scoring defects with a preset policy library to generate an initial optimization policy matrix.

[0076] In this embodiment, when providing an optimization strategy, based on the multi-dimensional scoring results, through a preset business impact factor model, the optimization priority labels for each dimension are calculated, and a strategy mapping algorithm is used to associate the scoring defects with a preset strategy library to generate an initial optimization strategy matrix. Specifically, the system first receives the multi-dimensional scoring results. For example, the compliance score of a certain customer service conversation is 85 points, the accuracy score is 70 points, and the sentiment polarity score is 90 points. The system evaluates the impact of each dimension on the overall business through a preset business impact factor model. For example, compliance has a greater impact on customer satisfaction and corporate reputation, and the business impact factor may be set to 0.4; accuracy has a greater impact on problem-solving efficiency, and the business impact factor is 0.3; sentiment polarity has a greater impact on the customer experience, and the business impact factor is 0.3. The system calculates the optimization priority labels for each dimension according to the scoring results and business impact factors. For example, the accuracy score is relatively low and the business impact factor is relatively high, so it is marked as "high priority"; the compliance and sentiment polarity scores are relatively high, and they are marked as "medium priority". Subsequently, the system uses a strategy mapping algorithm to associate the scoring defects with a preset strategy library. For example, for the problem of relatively low accuracy scores, the strategy library may have optimization strategies such as "supplementing key information" and "optimizing knowledge base retrieval". The system associates these strategies with the corresponding priority labels to generate an initial optimization strategy matrix.

[0077] S1512. According to the priority labels in the optimization strategy matrix, combined with the real-time conversation scenario features, recall an adapted set of optimization rules from the strategy library, and dynamically adjust the strategy weights through a rule engine, and output a scenario-based optimization strategy execution list.

[0078] Furthermore, the system also recalls an adapted set of optimization rules from the policy library according to the priority tags in the optimization policy matrix, combines the real-time conversation scenario features, dynamically adjusts the policy weights through a rule engine, and outputs a scenario-based optimization policy execution list. Specifically, the system first filters out high-priority optimization policies according to the priority tags in the optimization policy matrix. For example, for the problem of low accuracy scores, the high-priority policies in the optimization policy matrix are "supplement key information" and "optimize knowledge base retrieval". The system then further refines the set of optimization rules in combination with the real-time conversation scenario features. For example, when the current conversation scenario is "customer consulting product failure", the system will recall optimization rules related to "product failure consultation" from the policy library, such as "supplement failure handling process" and "provide repair contact information". To ensure the adaptability and effectiveness of the optimization policy, the system dynamically adjusts the policy weights through a rule engine. The rule engine will adjust the weights of the optimization policies according to the complexity and urgency of the real-time conversation scenario. For example, if the current conversation scenario involves high-value customers or urgent issues, the system will increase the weight of the policy of "providing repair contact information" to ensure a quick response to customer needs. The rule engine will also consider historical optimization effect data to dynamically adjust the policy weights to optimize the overall execution effect. Finally, the system outputs a scenario-based optimization policy execution list.

[0079] S1513. Feed back the optimization policy execution list to the corresponding customer service system.

[0080] In this embodiment, after the scenario-based optimization policy execution list is output, the optimization policy execution list is fed back to the corresponding customer service system. Specifically, the system transmits the generated scenario-based optimization policy execution list to the customer service system through a preset interface. For example, if the optimization policy execution list includes two policies, namely "supplement failure handling process" and "provide repair contact information", the system will push the specific content and priority information of these policies to the operation interface of the customer service system. After receiving the optimization policies, the customer service system will automatically embed these policies into the conversation template of the customer service staff. For example, when the customer service staff is handling "product failure consultation", the system will automatically prompt the specific conversation phrases of "supplement failure handling process" and "provide repair contact information" in the dialog box. At the same time, the system will also adjust the priority of the prompt according to the policy weights to ensure that the customer service staff gives priority to executing high-weight optimization policies.

[0081] The system of this embodiment can directly feedback the optimization policy to the actual operation of the customer service system, helping the customer service staff quickly adjust the service process and improve the service quality. This real-time feedback mechanism not only improves the execution efficiency of the optimization policy but also ensures that the customer service system can dynamically adjust the service policy according to the real-time conversation scenario, thus better meeting the customer needs.

[0082] For details, please refer toFigure 2 , Figure 2 This is a schematic diagram of the basic structure of the customer service data quality inspection device based on dynamic reasoning in this embodiment.

[0083] As Figure 2 shown, a customer service data quality inspection device based on dynamic reasoning includes: a data parsing module 1100, configured to parse customer service dialogue data based on a pre-configured large model to generate structured dialogue data including intent tags and key question nodes; a graph construction module 1200, configured to, according to the intent tags in the structured dialogue data, recall matching domain knowledge from a standard knowledge base in real time through a retrieval-augmented generation (RAG) framework to construct a dynamic context knowledge graph; a logical reasoning module 1300, configured to use the dynamic context knowledge graph as an input, and use a chain of thought (CoT) prompt template to guide the large model to perform step-by-step logical reasoning on the customer service dialogue data, and output a preliminary quality inspection conclusion chain including an inference path; a retrieval verification module 1400, configured to perform multi-modal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; a quantization output module 1500, configured to calculate a quantization score of the customer service dialogue data based on the final quality inspection conclusion, and generate a visual chain of thought report, where the quantization score includes quantization scores of at least one dimension of compliance, accuracy, and tone emotion, and the report includes error node annotation, correction path, and service optimization suggestions for any one of the dimensions.

[0084] The above-mentioned customer service data quality inspection device based on dynamic reasoning realizes accurate parsing and quality inspection of customer service conversations based on the deep semantic understanding and dynamic knowledge retrieval of the large model, and solves the problem of insufficient understanding of complex semantics by traditional methods. And by combining chain of thought reasoning and multi-modal verification, an interpretable quality inspection conclusion is generated, improving the transparency and accuracy of quality inspection. At the same time, the multi-dimensional scoring and visual report provide a clear improvement direction for the customer service team, support the self-optimization of the system, and significantly improve the customer service quality and management efficiency.

[0085] To solve the above technical problems, an embodiment of the present application also provides a computer device. Specifically, please refer to Figure 3 , Figure 3 This is a basic structure block diagram of the computer device in this embodiment.

[0086] As Figure 3As shown, it is a schematic internal structure diagram of a computer device. The computer device includes a processor, a non-volatile storage medium, a memory, and a network interface connected through a system bus. Among them, the non-volatile storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The database can store a control information sequence. When the computer-readable instructions are executed by the processor. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device can store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute a customer service data quality inspection method based on dynamic reasoning. The network interface of the computer device is used to connect and communicate with a terminal. Those skilled in the art can understand, Figure 3 The structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0087] In this embodiment, the processor is used to execute Figure 2 In the data parsing module 1100, the graph construction module 1200, the logical reasoning module 1300, the retrieval verification module 1400, and the quantization output module 1500, by parsing the customer service conversation data, generate structured conversation data containing intent tags and key question nodes; according to the intent tags in the structured conversation data, through the retrieval-augmented generation (RAG) framework, recall matching domain knowledge from the standard knowledge base in real time to construct a dynamic context knowledge graph; using the dynamic context knowledge graph as input, use the chain of thought (CoT) prompt template to guide the large model to perform step-by-step logical reasoning and output a preliminary quality inspection conclusion chain; retrieve and verify the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; based on the final quality inspection conclusion, calculate the quantization score of the customer service conversation data and generate a visual chain of thought report. Based on the deep semantic understanding and dynamic knowledge retrieval of the large model, the accurate parsing and quality inspection of customer service conversations are realized, and the problem of insufficient understanding of complex semantics by traditional methods is solved. And combined with chain of thought reasoning and multi-modal verification, an interpretable quality inspection conclusion is generated, which improves the transparency and accuracy of quality inspection. At the same time, the multi-dimensional scoring and visual report provide a clear improvement direction for the customer service team, support the self-optimization of the system, and significantly improve the quality of customer service and management efficiency.

[0088] This application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, one or more processors are caused to execute the steps of the customer service data quality inspection method based on dynamic reasoning described in any of the above embodiments.

[0089] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0090] Those skilled in the art of the present technology can understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in the present application can be alternated, changed, combined, or deleted. Further, other steps, measures, and solutions in the various operations, methods, and processes discussed in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted. Further, the steps, measures, and solutions in the prior art that are the same as those disclosed in the various operations, methods, and processes in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0091] The above are only some embodiments of the present application. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A customer service data quality inspection method based on dynamic reasoning, characterized in that Including: Parsing customer service dialogue data based on a pre-configured large model to generate structured dialogue data containing intent tags and key question nodes; According to the intent tags in the structured dialogue data, through a retrieval-augmented generation framework, retrieving matching domain knowledge from a standard knowledge base in real time to construct a dynamic context knowledge graph; Using the dynamic context knowledge graph as input, guiding the large model to perform step-by-step logical reasoning on the customer service dialogue data by means of a chain-of-thought prompting template, and outputting a preliminary quality inspection conclusion chain containing the reasoning path; Performing multi-modal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; Based on the final quality inspection conclusion, calculating a quantitative score for the customer service dialogue data and generating a visual chain-of-thought report. The quantitative score includes quantitative scores for at least one dimension of compliance, accuracy, and tone sentiment. The report contains error node annotations, correction paths, and service optimization suggestions for any one of the dimensions.

2. The customer service data quality inspection method based on dynamic reasoning according to claim 1, wherein The parsing includes word segmentation, entity recognition, and context semantic parsing. The parsing of the customer service dialogue data based on the pre-configured large model to generate structured dialogue data containing intent tags and key question nodes includes: Based on the zero-shot learning ability of the large model, pre-classifying the intent of the customer service dialogue data to generate preliminary intent tags; Disambiguating and correcting the preliminary intent tags through context semantic reinforcement configuration to output multi-dimensional intent tags; Combining the results of the entity recognition, associating the multi-dimensional intent tags with key question nodes to construct the structured dialogue data.

3. The customer service data quality inspection method based on dynamic reasoning according to claim 1, wherein The domain knowledge includes at least one of normative texts, historical cases, and industry policies. The construction of the dynamic context knowledge graph includes: Based on the semantic vectors of the intent tags, retrieving normative texts in the standard knowledge base by using cosine similarity matching; In the historical case library, screening the historical case with the highest similarity to the dialogue scenario of the customer service dialogue data based on a time-weighted retrieval strategy; Fusing the normative texts and historical cases to generate a dynamic context knowledge graph.

4. The customer service data quality inspection method based on dynamic reasoning according to claim 1, wherein, The retrieving matching domain knowledge from the standard knowledge base in real time through a retrieval-augmented generation framework to construct a dynamic context knowledge graph includes: Determining the timeliness requirements according to the dialogue scenario of the customer service dialogue data; Based on the timeliness requirements, performing time decay weighting on the retrieval results retrieved in real time from the standard knowledge base, and preferentially retrieving recently updated domain knowledge to construct a dynamic context knowledge graph.

5. The customer service data quality inspection method based on dynamic reasoning according to claim 1, characterized in that The step-by-step logical reasoning includes compliance verification, knowledge accuracy verification, and sentiment polarity analysis. The guiding of the large model to perform step-by-step logical reasoning on the customer service dialogue data by means of a chain-of-thought prompting template and outputting a preliminary quality inspection conclusion chain containing the reasoning path includes: Through the compliance rule chain in the prompting template, layer-by-layer verifying whether the customer service replies in the customer service dialogue data comply with industry norms; Evaluating the accuracy score of the customer service replies in the customer service dialogue data based on historical case similarity; Quantify the sentiment polarity of the customer service replies in the customer service conversation data using a sentiment analysis model, and generate a sentiment score by comparing it with a preset threshold. Output the preliminary quality inspection conclusion chain including the inference paths of compliance verification, knowledge accuracy verification, and sentiment polarity analysis.

6. The customer service data quality inspection method based on dynamic reasoning according to claim 1, wherein The multimodal retrieval verification includes verification of at least one of text, rule base, historical case base, or manually annotated data. The multimodal retrieval verification of the assertion nodes in the preliminary quality inspection conclusion chain includes: Based on the assertion nodes in the preliminary quality inspection conclusion chain, extract the assertion content, perform alignment verification with the structured text specifications in the rule base through a semantic matching algorithm, calculate the semantic similarity score. If the semantic similarity score is lower than the preset threshold, mark this assertion node as a potential logical conflict node. According to the context scenario of the assertion node, recall similar conversation cases from the historical case base, and compare the consistency between the current assertion conclusion and the historical processing result through a time-weighted strategy and scenario similarity. If the consistency meets the condition of significant difference, mark this assertion node as a conflict warning node.

7. The customer service data quality inspection method based on dynamic reasoning according to claim 1, characterized in that The method further includes: Based on the multi-dimensional scoring results, through a preset business impact factor model, calculate the optimization priority labels for each dimension, and use a strategy mapping algorithm to associate the scoring defects with a preset strategy library to generate an initial optimization strategy matrix. According to the priority labels in the optimization strategy matrix, combined with the real-time conversation scenario features, recall the adapted optimization rule set from the strategy library, and dynamically adjust the strategy weights through a rule engine to output a scenario-based optimization strategy execution list. Feed back the optimization strategy execution list to the corresponding customer service system.

8. A customer service data quality inspection device based on dynamic reasoning, characterized in that, Includes: A data parsing module for parsing customer service conversation data based on a pre-configured large model to generate structured conversation data containing intent labels and key question nodes. A graph construction module for, according to the intent labels in the structured conversation data, using a retrieval-augmented generation framework to recall matching domain knowledge from a standard knowledge base in real time to construct a dynamic context knowledge graph. A logical reasoning module for using the dynamic context knowledge graph as input, guiding the large model to perform step-by-step logical reasoning on the customer service conversation data using a chain-of-thought prompting template, and outputting a preliminary quality inspection conclusion chain including the inference path. A retrieval verification module for performing multimodal retrieval verification on the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion. A quantification output module for calculating the quantification score of the customer service conversation data based on the final quality inspection conclusion and generating a visual chain-of-thought report. The quantification score includes quantification scores for at least one dimension of compliance, accuracy, and tone sentiment. The report includes error node annotation, correction path, and service optimization suggestions for any one of the dimensions.

9. A computer device, characterized in that, Includes a memory and a processor. When the computer-readable instructions stored in the memory are executed by the processor, the processor executes the steps of the customer service data quality inspection method based on dynamic reasoning according to any one of claims 1 to 7.

10. A storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the method for dynamically inferring customer service data quality inspection according to any one of claims 1 to 7.

Images