Intelligent customer service system based on AI big model

Abstract

The present invention discloses an intelligent customer service system based on an AI big model, and belongs to the technical field of intelligent customer service systems. The present invention improves the flexibility of the system and user experience through multimodal interaction technology, intelligently identifies data types and performs corresponding processing, automatically adjusts the language style and emotional intonation of the response content according to the user's tone and intention, and the knowledge base integrates static data and dynamic interaction data to ensure the real-time update and comprehensiveness of the knowledge content. It performs rapid retrieval based on the knowledge base and makes accurate replies in combination with a large language model. It generates high-quality answers and solutions through rule reasoning and generative models, and records and manages the status information of each round of user interaction by building a dialogue state tracker, so as to achieve accurate tracking and updating of the context. For complex tasks, the system breaks down user requests into multiple subtasks, and executes them in sequence according to priority and dependency relationships to ensure the efficiency and consistency of task processing.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent customer service systems, and in particular to an intelligent customer service system based on an AI big model. Background Art

[0002] Traditional customer service systems mainly rely on manual customer service or simple rule engines. When faced with a wide variety of complex user questions, it is often difficult to provide fast, accurate and personalized services. In addition, when processing multimodal data, existing customer service systems can usually only process a single input and lack the ability to integrate cross-modal data, resulting in an imperfect user experience.

[0003] At present, intelligent customer service systems based on natural language processing (NLP) and big data technology have been able to achieve automated customer service to a certain extent, but there are still some problems:

[0004] First, most systems are still text-based, with weak capabilities for processing multimodal data such as voice and images, and the integration and response generation of multimodal inputs often lack intelligent adjustments.

[0005] Secondly, many intelligent customer service systems lack a deep understanding and accurate identification of user intentions, which results in the system often failing to give ideal responses when dealing with complex issues. Especially in multi-round dialogue scenarios, the system is unable to efficiently track and update the dialogue status, and often suffers from insufficient understanding of the context.

[0006] Furthermore, existing knowledge base management methods are usually static and lack dynamic updating and automatic optimization capabilities, resulting in delayed updates of knowledge base content and an inability to cope with ever-changing user needs and new business scenarios. Summary of the invention

[0007] The purpose of the present invention is to provide an intelligent customer service system based on an AI big model to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: an intelligent customer service system based on an AI big model, comprising:

[0009] User interaction module, used to:

[0010] Receive multimodal consultation data input by the user, call the service interface to obtain the result according to the user's request type, and convert the response data into one or more of natural language text, voice reply and visual image data;

[0011] Natural language parsing module, used for:

[0012] Perform word segmentation, part-of-speech tagging, and syntactic analysis on the content input by the user, capture context-related semantic information, match the user's request type and intention, extract key entities, time information, and context information elements, and generate structured semantic representation data;

[0013] Knowledge base management module, used to:

[0014] Integrate and normalize static knowledge data and dynamic user interaction data, vectorize knowledge content and build an index structure, and compare the answers generated by the large language model with the knowledge base content;

[0015] AI decision making module for:

[0016] Based on user intent and knowledge base query results, a large language model is called to generate preliminary response content, and rule reasoning is introduced to make step-by-step decisions and execute tasks. The format, language style, and business consistency of the generated content are controlled by fine-tuning the model and inserting prompts.

[0017] Multi-round dialogue management module, used for:

[0018] Build a dialogue state tracker to update the state of each round of interaction input by the user, record the intent, entity and context information, replace the large language model with the dialogue history to generate dynamic replies, analyze user goals based on the dialogue path and predict subsequent operations, and call different models and logic modules for dialogue scenarios of different complexity;

[0019] System monitoring and optimization module for:

[0020] Collect system operation data in real time, detect abnormal situations through data analysis tools, and generate a visual management interface. The visual management interface is used to view the operation status and manually intervene in key nodes. The system operation data includes response time, user satisfaction and error rate. The abnormal situations include model deviation, knowledge base missing and high-frequency user complaints.

[0021] Furthermore, the user interaction module includes:

[0022] The data receiving and identifying unit is used for:

[0023] Receiving consultation requests input by users through text, voice and image channels, and identifying data types, wherein the data types include text data, voice data, image data and combined data;

[0024] For speech data, extract speech features and generate corresponding text content through the text conversion module;

[0025] For image data, key visual information is extracted through image features to generate image description content;

[0026] Data pre-processing unit, used for:

[0027] Clean, format and standardize received user input data;

[0028] Response generation display unit for:

[0029] The best response form is determined based on the output content of the natural language parsing module and the AI ​​decision generation module, and the response form includes text, voice, image and a combination thereof.

[0030] Furthermore, when generating the best response form, the response generation and display unit adjusts the language style according to the tone and intention of the user input and annotates the professional terms and complex information when replying in the form of a text response;

[0031] When replying in the form of voice response, the text content is converted into a voice response, and the voice tone is adjusted according to the user sentiment analysis results;

[0032] When replying in the form of an image response, generate charts, visualizations, and image content based on user needs, and annotate key information points.

[0033] Furthermore, the natural language parsing module includes:

[0034] Semantic analysis unit, used to:

[0035] Segment the user input text, generate word sequences and mark the grammatical structure, assign part-of-speech tags to each word through part-of-speech tagging, analyze the dependency relationship between sentence components, and generate sentence semantic vectors;

[0036] Intent recognition unit, used to:

[0037] Classify semantic vectors and determine the user request type, which includes question and answer, transaction processing, and complaint. Track the intent of continuous input in multiple rounds of dialogue to determine whether the user changes the intent or continues the current task. When the user expresses multiple intents in one input, separate the sub-intents.

[0038] An information extraction unit, for:

[0039] Extract entity information from user input, including names of people, places, organizations, and product names; extract time information from user input, including date, time, and time interval, and standardize the time information; extract quantity information and quantitative description from the input and map them to structured fields;

[0040] Identify the core problem or specific description of the need in the input;

[0041] In a multi-round dialogue scenario, context information is used to supplement and update unclear elements to generate complete structured semantic representation data, where the unclear elements include one or more of omitted subjects and time.

[0042] Furthermore, the knowledge base management module includes:

[0043] Knowledge Integration Unit for:

[0044] Receive and import static knowledge data, including FAQs, corporate policy documents, and technical manuals, and dynamically collect user interaction dynamic data, including historical conversation records, real-time data streams, and user behavior logs;

[0045] Perform format parsing and semantic splitting on static knowledge data documents, divide long static knowledge data contents into independent knowledge items, and update dynamic data in real time;

[0046] Knowledge vectorization unit, used for:

[0047] Generate high-dimensional vector representations of FAQ items and document contents through embedding models, capture the semantic features of data, and introduce time and context information in the dynamic data embedding process;

[0048] Build a vector index structure and add a category tag to each vectorized data;

[0049] Knowledge comparison unit, used for:

[0050] Receive structured semantic representation data and convert it into a query vector;

[0051] Semantically match the query vector with the knowledge vector library to obtain the most relevant knowledge items;

[0052] The matching knowledge items are trimmed and reorganized, and the results with large semantic deviations are eliminated according to the semantic similarity threshold to obtain the retrieval results;

[0053] Mark frequently called and high error rate entries for update and optimization.

[0054] Furthermore, the AI ​​decision making module includes:

[0055] User intent processing unit, used to:

[0056] Compare the intent recognition results of the natural language parsing module with the entry content matched in the knowledge base to determine the coverage of user intent and knowledge content;

[0057] If the user's intention is clear and supported by corresponding knowledge items, the content generation optimization process will be started;

[0058] If the user's intention is unclear or the match with the knowledge base content is low, the generative large model is called to supplement the content and record the unmatched knowledge points;

[0059] Rule reasoning unit, used to:

[0060] The key elements in the user's intent are parsed and matched with the business rule library, wherein the key elements include entity, time, and operation type.

[0061] Build task execution process;

[0062] For transaction processing tasks, user requests are decomposed into multiple subtasks, and execution sequences are generated according to priorities and dependencies. For complex complaint tasks, reasoning logic is called to determine the best processing solution and additional information is added, and intermediate decision data is generated based on the reasoning results.

[0063] Generate content optimization units for:

[0064] Input user intent and knowledge base comparison results, and generate preliminary response content based on intermediate decision data;

[0065] Optimize the generated content, control the format, logical level and semantic integrity of the generated content through preset templates and task prompts, and verify the numbers, time and calculation content;

[0066] Eliminate potential banned words and sensitive content through business rules and semantic review.

[0067] Furthermore, the multi-round dialogue management module includes:

[0068] Dialog state tracking unit, used to:

[0069] At the beginning of a conversation, a blank conversation state container is constructed, which contains the following information fields: intent, entity, time, context, and unfinished task list, identifies the intent and key entity information input by the user in the first round, and fills in the initial state data;

[0070] For each round of user input, the natural language parsing module is called to extract the intent and information elements, and then compared and merged with the initial state data in the current dialog state container;

[0071] After the user interrupts the conversation, the current conversation state data is stored and the key nodes of the conversation are recorded. When resuming the conversation, the stored conversation state data is read to prompt the user to resume to the interrupted position;

[0072] Dialogue path prediction unit, used to:

[0073] Based on the current dialogue state and user input, combined with the intention recognition results of the natural language parsing module, determine the user's goal type, which includes short-term goals and long-term goals. The short-term goals include single-round question-answering and transaction processing, and the long-term goals include multi-step tasks and complex problem solving;

[0074] Use the conversation history to build a path model, associate the user's past operation sequence with the knowledge base content, and generate a list of possible follow-up operation recommendations;

[0075] The predicted subsequent path is presented to the user in the form of visualization and text suggestions;

[0076] Monitor the user's response to route suggestions in real time. When the user deviates from the recommended route, re-evaluate the dialogue goal and generate new route suggestions.

[0077] Furthermore, in the system monitoring and optimization module, the system operation data includes response time, user satisfaction and error rate, and the abnormal conditions include model deviation, knowledge base missing and high-frequency user complaints.

[0078] Furthermore, the multi-round dialogue management module further includes:

[0079] A data collection unit, used to collect each round of interaction data input by the user based on a preset frequency, and to organize the collected each round of interaction data to obtain interaction data to be analyzed;

[0080] Data analysis unit for:

[0081] Performing word segmentation on the interaction data to be analyzed, and performing part-of-speech tagging on the word segmentation results based on the natural language system, and determining the interaction objects and speech logic corresponding to the interaction data to be analyzed based on the part-of-speech tagging results;

[0082] Conduct analogical analysis on the speech logic of the interactive object in different interactive scenarios, determine the user's interactive characteristics based on the analogical analysis, and obtain the user's interactive behavior habits based on the interactive characteristics;

[0083] Smart reply configuration unit, used to:

[0084] Generate personalized reply strategies for users based on interactive behavior habits, integrate and deploy the personalized reply strategies in the replacement large language model, and dynamically reply to users based on the integrated deployment results;

[0085] Smart reply optimization unit, used to:

[0086] Preprocess each round of user interaction data based on the dynamic reply results, and determine the text structure and text semantics of the interaction text based on the preprocessing results;

[0087] Extract emotional keywords based on text structure and text semantics, and match the emotional keywords with the pre-built emotional dictionary to determine the emotional tendency of the emotional keywords;

[0088] Extract tone symbols from interactive texts based on text structure, and determine the user's tone features based on tone symbols. At the same time, determine the user's current interaction scenario based on text semantics, and obtain the user's current interaction emotions based on tone features, interaction scenarios, and emotional tendencies.

[0089] Determine the response tone to the user's dynamic reply content based on the interaction emotion, match available tone words from a preset vocabulary library based on the response tone, and logically associate the available tone words with the user's dynamic reply content;

[0090] Dynamically respond to users based on logical key results.

[0091] Furthermore, the system monitoring and optimization module also includes evaluating and monitoring the system:

[0092] A first calculation unit, used for reading a data response parameter of system operation data, and calculating the operation efficiency of the system according to the response parameter of the system operation data;

[0093] ;

[0094] in, Indicates the operating efficiency of the system; represents the best response time; is the actual response time; Indicates system response delay, and ; represents the best response rate; represents the actual response rate, and ; Indicates system operation data; Indicates that in the system operation data, the correct system operation data is responded; Indicates the system operation data that responds to errors in the system operation data;

[0095] The second calculation unit is used to randomly collect the number of customer service visits within a number of preset time periods, and at the same time, obtain the satisfaction score of each customer for the visit interaction results within each preset time period;

[0096] Calculate the customer's comprehensive satisfaction rating of the system based on the rating of each customer's access interaction results in each preset time period;

[0097] ;

[0098] in, Indicates the customer's comprehensive satisfaction rating of the system; Indicates the total number of preset time periods; Indicates the serial number value of the preset time period; Indicated in The total number of customers within a preset time period; Indicated in The customer serial number value within a preset time period; Indicated in The first Customer satisfaction rating of the system;

[0099] System evaluation monitoring report generation unit, used to:

[0100] Generate a first evaluation monitoring report for evaluating the system according to the operating efficiency of the system;

[0101] Generate a second evaluation monitoring report for evaluating the system based on the customer's comprehensive satisfaction rating of the system;

[0102] The report transmission unit is used to transmit the first evaluation monitoring report and the second evaluation monitoring report to the monitoring terminal.

[0103] Compared with the prior art, the present invention has the following beneficial effects:

[0104] 1. The present invention can process user input in various forms such as text, voice and image through multimodal interaction, thereby improving the flexibility and usability of the system and meeting the needs of different users. The different needs of users can be accurately responded to through different response methods - whether answering in text form, responding in voice form, or visually displayed through images and charts, the system can provide the response that best meets the user's needs, automatically adjust the language style and tone according to the user's emotions and tone to optimize the experience, and the personalized and flexible response mechanism greatly improves the user experience and increases customer satisfaction and trust in the intelligent customer service system.

[0105] 2. The present invention can intelligently manage and use a large amount of static knowledge data and dynamic user interaction data through the knowledge base management and AI decision generation modules. The system can integrate static knowledge and real-time interaction data such as FAQs, corporate policies, and technical manuals to ensure the comprehensiveness and real-time nature of the knowledge base. When a user asks a question, the system can quickly match relevant information in the knowledge base and compare it with the answer of the large language model to ensure that the generated reply is both accurate and in context. Through rule reasoning, the system can not only generate correct response content based on user intent and related business rules, but also supplement and generate reasonable content through the large language model when the user intent is unclear. This allows the system to not only cope with standard question-and-answer scenarios, but also efficiently handle complex transactions and personalized needs, significantly improving processing efficiency and service quality.

[0106] 3. The present invention establishes a dialogue state tracker to dynamically record and manage the state information of each round of dialogue, ensuring that in multiple rounds of dialogue, the system can accurately track the user's needs and provide continuous, context-consistent answers. In complex multi-step tasks or multi-round dialogues, the system can intelligently analyze the user's goals and make predictions. When the user deviates from the recommended path, the system can adjust the dialogue strategy in real time and re-predict the path, which not only makes the user's operation more efficient, but also avoids stagnation and confusion that may occur during long-term dialogues. The system decomposes the user's request into multiple subtasks, and ensures that each subtask is executed in sequence according to priority and dependencies through intelligent reasoning and decision engines, ensuring efficient processing of complex tasks, improving service quality and user satisfaction, and especially when dealing with complex complaints and problem solving, it can provide the best solution through a reasonable reasoning mechanism.

[0107] 4. By organizing and analyzing the interaction data of each round of users, the interaction behavior habits of users during interaction can be effectively determined, so that personalized reply strategies can be formulated for users based on their interaction behavior habits, thereby improving the effect of intelligent replies. Secondly, the obtained personalized reply strategies are integrated and deployed in the replacement large language model, and finally reliable intelligent replies are achieved, which improves the user's interactive experience. Finally, the user's current emotional state is monitored in real time during the dynamic reply process, and corresponding response measures are taken according to the user's current emotional state to provide effective intelligent replies to users, ensuring that the user's emotional needs are met without affecting the reliability of the reply content, improving the accuracy and reliability of intelligent replies, and ensuring the effect of intelligent replies.

[0108] 5. By calculating the system's operating efficiency and the customer's comprehensive satisfaction score for the system, the first evaluation monitoring report and the second evaluation monitoring report can be effectively obtained, thereby effectively realizing the evaluation results of the system's internal performance and external satisfaction, thereby improving the comprehensiveness of the system evaluation and facilitating timely understanding of the current system's operating status. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 Schematic diagram of the modules of the intelligent customer service system of the present invention. DETAILED DESCRIPTION

[0110] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0111] See also Figure 1 , the present invention provides the following technical solutions:

[0112] Intelligent customer service system based on AI big model, including:

[0113] User interaction module, used to:

[0114] Receive multimodal consultation data input by the user, call the service interface to obtain the result according to the user's request type, and convert the response data into one or more of natural language text, voice reply and visual image data;

[0115] Natural language parsing module, used for:

[0116] Perform word segmentation, part-of-speech tagging, and syntactic analysis on the content input by the user, capture context-related semantic information, match the user's request type and intention, extract key entities, time information, and context information elements, and generate structured semantic representation data;

[0117] Knowledge base management module, used to:

[0118] Integrate and normalize static knowledge data and dynamic user interaction data, vectorize knowledge content and build an index structure, and compare the answers generated by the large language model with the knowledge base content;

[0119] AI decision making module for:

[0120] Based on user intent and knowledge base query results, a large language model is called to generate preliminary response content, and rule reasoning is introduced to make step-by-step decisions and execute tasks. The format, language style, and business consistency of the generated content are controlled by fine-tuning the model and inserting prompts.

[0121] Multi-round dialogue management module, used for:

[0122] Build a dialogue state tracker to update the state of each round of interaction input by the user, record the intent, entity and context information, replace the large language model with the dialogue history to generate dynamic replies, analyze user goals based on the dialogue path and predict subsequent operations, and call different models and logic modules for dialogue scenarios of different complexity;

[0123] System monitoring and optimization module for:

[0124] Collect system operation data in real time, detect abnormal situations through data analysis tools, and generate a visual management interface for viewing the operation status and manually intervening in key nodes.

[0125] In the above embodiment, through the natural language parsing module, image recognition module and voice processing module, the system can intelligently process different types of data, thereby improving the convenience and flexibility of user interaction. Whether it is voice assistant, image query, or text interaction, users can get a smooth and accurate service experience. Through the AI ​​decision generation module combined with knowledge base query and natural language processing, the optimal response content is automatically generated. It is not only efficient and accurate in providing answers to routine questions, but also can generate answers with high logic, emotional adaptation and business consistency when facing complex business problems, thereby improving service quality, and making adaptive adjustments based on user feedback and sentiment analysis to ensure maximum user experience.

[0126] In the above embodiment, the multi-round dialogue management module provides powerful dialogue state tracking and path prediction functions. By tracking the user's dialogue state and predicting possible subsequent operations in real time, the system can intelligently adjust the dialogue path to ensure that user questions are responded to and processed in a timely manner. At the same time, in complex dialogues, the system can dynamically generate follow-up operation suggestions based on user goals, greatly improving the continuity and efficiency of the dialogue. Through the knowledge base management module, static and dynamic knowledge are efficiently integrated and managed, and rapid access and update of corporate FAQs, technical documents and user behavior data are achieved. Through vectorized representation and knowledge comparison, the system can accurately match and extract the most relevant knowledge items, greatly improving the accuracy and efficiency of problem solving.

[0127] User interaction module, including:

[0128] The data receiving and identifying unit is used for:

[0129] Receiving consultation requests input by users through text, voice and image channels, and identifying data types, wherein the data types include text data, voice data, image data and combined data;

[0130] For speech data, extract speech features and generate corresponding text content through the text conversion module;

[0131] For image data, key visual information is extracted through image features to generate image description content;

[0132] Data pre-processing unit, used for:

[0133] Clean, format and standardize received user input data;

[0134] Response generation display unit for:

[0135] Determine the best response form based on the output content of the natural language parsing module and the AI ​​decision generation module, and the response form includes text, voice, image and combination;

[0136] When responding in text form, adjust the language style based on the tone and intent of the user's input, and annotate professional terms and complex information;

[0137] When replying in the form of voice response, the text content is converted into a voice response, and the voice tone is adjusted according to the user sentiment analysis results;

[0138] When replying in the form of an image response, generate charts, visualizations, and image content based on user needs, and annotate key information points.

[0139] In the above embodiment, the module supports multiple input methods such as text, voice and image, so that users can flexibly interact with the system in any scenario. When voice input is used, the system can extract voice features and convert them into text data for subsequent processing; for image input, the system can extract key information from the image and provide image description. This multimodal interaction capability greatly improves the user experience, especially in complex situations such as the interaction of image recognition and voice commands, which enhances the intelligence and diversity of the system. After the various data (text, voice, image) input by the user is cleaned, standardized and formatted, the system can effectively parse it, ensuring the accuracy and uniformity of the data, avoiding erroneous parsing caused by input data problems, and improving the accuracy of system processing. According to the user's needs and context, the system can intelligently select the best response form: text, voice, image or a combination thereof. If the content input by the user is more complex and requires precise explanation, the system will give priority to text response and annotate professional terms; if it is a voice response, the system adjusts the voice tone through sentiment analysis to ensure that the response is more friendly and personalized.

[0140] Natural language parsing module, including:

[0141] Semantic analysis unit, used to:

[0142] Segment the user input text, generate word sequences and mark the grammatical structure, assign part-of-speech tags to each word through part-of-speech tagging, analyze the dependency relationship between sentence components, and generate sentence semantic vectors;

[0143] Intent recognition unit, used to:

[0144] Classify semantic vectors and determine the user request type, which includes question and answer, transaction processing, and complaint. Track the intent of continuous input in multiple rounds of dialogue to determine whether the user changes the intent or continues the current task. When the user expresses multiple intents in one input, separate the sub-intents.

[0145] An information extraction unit, for:

[0146] Extract entity information from user input, including names of people, places, organizations, and product names; extract time information from user input, including date, time, and time interval, and standardize the time information; extract quantity information and quantitative description from the input and map them to structured fields;

[0147] Identify the core problem or specific description of the need in the input;

[0148] In a multi-round dialogue scenario, context information is used to supplement and update unclear elements to generate complete structured semantic representation data, where the unclear elements include one or more of omitted subjects and time.

[0149] In the above embodiment, a high-precision semantic vector is generated by word segmentation, part-of-speech tagging, and syntactic analysis of the user input, thereby effectively capturing the user's intentions and needs and accurately parsing complex user expressions. Even context changes in multiple rounds of conversations can be efficiently tracked. If the user changes the intention of the question or introduces new entity information in a continuous conversation, the system can update the conversation status in a timely manner to ensure the accuracy of the answer.

[0150] In the above embodiment, by extracting key information from the user's input, such as entity information such as name, location, date, time interval, and quantitative information such as quantity and product, through standardization processing, this information can be quickly converted into structured data for subsequent processing and decision-making. In multi-round dialogue scenarios, the module uses contextual information to supplement and update missing elements in the user input, such as omitted subjects, time, etc., to ensure that the generated semantic representation data is complete and accurate, so that the system can still maintain good accuracy and consistency in complex dialogue scenarios. Through efficient semantic understanding and intent recognition, the natural language parsing module provides the intelligent customer service system with powerful language understanding and decision support capabilities, ensuring that the system can accurately respond to user needs in multi-round dialogues.

[0151] Knowledge base management module, including:

[0152] Knowledge Integration Unit for:

[0153] Receive and import static knowledge data, including FAQs, corporate policy documents, and technical manuals, and dynamically collect user interaction dynamic data, including historical conversation records, real-time data streams, and user behavior logs;

[0154] Perform format parsing and semantic splitting on static knowledge data documents, divide long static knowledge data contents into independent knowledge items, and update dynamic data in real time;

[0155] Knowledge vectorization unit, used for:

[0156] Generate high-dimensional vector representations of FAQ items and document contents through embedding models, capture the semantic features of data, and introduce time and context information in the dynamic data embedding process;

[0157] Build a vector index structure and add a category tag to each vectorized data;

[0158] Knowledge comparison unit, used to:

[0159] Receive structured semantic representation data and convert it into a query vector;

[0160] Semantically match the query vector with the knowledge vector library to obtain the most relevant knowledge items;

[0161] The matching knowledge items are trimmed and reorganized, and the results with large semantic deviations are eliminated according to the semantic similarity threshold to obtain the retrieval results;

[0162] Mark frequently called and high error rate entries for update and optimization.

[0163] In the above embodiment, by accessing static knowledge resources such as FAQs, corporate documents, technical manuals, etc., and combining with data dynamically generated during user interaction for real-time updates, the system will automatically record common problems and solutions in user feedback, and convert them into knowledge items, and update the knowledge base in real time to ensure that the corporate knowledge base is always up-to-date and efficient. By vectorizing the documents and FAQ items in the knowledge base in high dimensions, the system can efficiently match user queries, extract the most relevant information, and perform semantic matching and tailoring to ensure the accuracy and efficiency of knowledge retrieval. By constructing a vector index structure, the system greatly improves the response speed of knowledge base queries, automatically marks and updates knowledge items according to their call frequency and error rate, optimizes the knowledge base content, and continuously optimizes the knowledge base to ensure that the answers provided by the system are always accurate and adapt to the ever-changing needs of users.

[0164] AI decision generation module, including:

[0165] User intent processing unit, used to:

[0166] Compare the intent recognition results of the natural language parsing module with the entry content matched in the knowledge base to determine the coverage of user intent and knowledge content;

[0167] If the user's intention is clear and supported by corresponding knowledge items, the content generation optimization process will be started;

[0168] If the user's intention is unclear or the match with the knowledge base content is low, the generative large model is called to supplement the content and record the unmatched knowledge points;

[0169] Rule reasoning unit, used to:

[0170] The key elements in the user's intent are parsed and matched with the business rule library, wherein the key elements include entity, time, and operation type.

[0171] Build task execution process;

[0172] For transaction processing tasks, user requests are decomposed into multiple subtasks, and execution sequences are generated according to priorities and dependencies. For complex complaint tasks, reasoning logic is called to determine the best processing solution and additional information is added, and intermediate decision data is generated based on the reasoning results.

[0173] Generate content optimization units for:

[0174] Input user intent and knowledge base comparison results, and generate preliminary response content based on intermediate decision data;

[0175] Optimize the generated content, control the format, logical level and semantic integrity of the generated content through preset templates and task prompts, and verify the numbers, time and calculation content;

[0176] Eliminate potential banned words and sensitive content through business rules and semantic review.

[0177] In the above embodiment, according to the intention recognition results of the natural language parsing module, combined with the knowledge base content, the coverage of the user's intention and the knowledge base is judged, and different processing strategies are adopted. When the user's intention is clear and the knowledge base supports it, the system directly generates an optimized answer; if the user's intention is unclear, it is supplemented and optimized through the generative large model to ensure that each response meets the user's needs.

[0178] In the above embodiment, the user's request is parsed and the task is decomposed according to the matching of the key elements in the user's intention with the business rule library. For example, for transaction processing tasks, the system can decompose the user's request into multiple subtasks and prioritize them; for complex complaint tasks, the system can use reasoning logic to give the most reasonable processing solution.

[0179] In the above embodiment, by generating a content optimization unit, the format, logical level and semantic integrity of the response content are optimized to ensure that the answer is not only accurate but also easy to understand and accept. In addition, through semantic review and rule review, potential banned words or sensitive content are eliminated to ensure that the system's response always complies with the specification.

[0180] Multi-round dialogue management module, including:

[0181] Dialog state tracking unit, used to:

[0182] At the beginning of a conversation, a blank conversation state container is constructed, which contains the following information fields: intent, entity, time, context, and unfinished task list, identifies the intent and key entity information input by the user in the first round, and fills in the initial state data;

[0183] For each round of user input, the natural language parsing module is called to extract the intent and information elements, and then compared and merged with the initial state data in the current dialog state container;

[0184] After the user interrupts the conversation, the current conversation state data is stored and the key nodes of the conversation are recorded. When resuming the conversation, the stored conversation state data is read to prompt the user to resume to the interrupted position;

[0185] Dialogue path prediction unit, used to:

[0186] Based on the current dialogue state and user input, combined with the intention recognition results of the natural language parsing module, determine the user's goal type, which includes short-term goals and long-term goals. The short-term goals include single-round question-answering and transaction processing, and the long-term goals include multi-step tasks and complex problem solving;

[0187] Use the conversation history to build a path model, associate the user's past operation sequence with the knowledge base content, and generate a list of possible follow-up operation recommendations;

[0188] The predicted subsequent path is presented to the user in the form of visualization and text suggestions;

[0189] Monitor the user's response to route suggestions in real time. When the user deviates from the recommended route, re-evaluate the dialogue goal and generate new route suggestions.

[0190] In the above embodiment, the module dynamically tracks and updates the user's intention, entity, time and other information in each round of dialogue, and effectively manages the user's needs and tasks. By saving the dialogue state, the system can handle the user's interruption and recovery, avoid repeated user input when processing complex tasks, and improve service efficiency and experience. According to the current state of user input, combined with historical dialogue information, possible subsequent operations and user goals are predicted, thereby generating a recommended dialogue path to help users quickly complete tasks and avoid unnecessary detours or information loss. According to the dialogue history and user preferences, a personalized service experience is provided. The dialogue path prediction not only improves service efficiency, but also adjusts the dialogue strategy in time according to changes in user needs, optimizing the entire interaction process.

[0191] In one embodiment, an intelligent customer service system based on an AI big model is provided, a multi-round dialogue management module, and further includes:

[0192] The data collection unit is used to collect each round of interaction data input by the user based on a preset frequency, and to organize the collected each round of interaction data to obtain interaction data to be analyzed.

[0193] Data analysis unit for:

[0194] Performing vocabulary segmentation on the interaction data to be analyzed, and performing part-of-speech tagging on the vocabulary segmentation results based on the natural language system, and determining the interaction objects and speech logic corresponding to the interaction data to be analyzed based on the part-of-speech tagging results;

[0195] Conduct analogical analysis on the speech logic of the interactive object in different interactive scenarios, determine the user's interactive characteristics based on the analogical analysis, and obtain the user's interactive behavior habits based on the interactive characteristics;

[0196] Smart reply configuration unit, used to:

[0197] Generate a personalized reply strategy for users based on their interactive behavior habits, integrate and deploy the personalized reply strategy in the replacement large language model, and dynamically reply to users based on the integrated deployment results;

[0198] Smart reply optimization unit, used to:

[0199] Preprocess each round of user interaction data based on the dynamic reply results, and determine the text structure and text semantics of the interaction text based on the preprocessing results;

[0200] Extract emotional keywords based on text structure and text semantics, and match the emotional keywords with the pre-built emotional dictionary to determine the emotional tendency of the emotional keywords;

[0201] Extract tone symbols from interactive texts based on text structure, and determine the user's tone characteristics based on tone symbols. At the same time, determine the user's current interaction scenario based on text semantics, and obtain the user's current interaction emotion based on tone characteristics, interaction scenario, and emotional tendency.

[0202] Determine the response tone to the user's dynamic reply content based on the interaction emotion, match available tone words from a preset vocabulary library based on the response tone, and logically associate the available tone words with the user's dynamic reply content;

[0203] Dynamically respond to users based on logical key results.

[0204] In this embodiment, the preset frequency is set in advance, for example, it may be 50 times or 100 times.

[0205] In this embodiment, the interaction data to be analyzed refers to the results obtained after classification and cleaning of each round of collected interaction data, which provides reliable data support for data analysis.

[0206] In this embodiment, part-of-speech tagging refers to annotating the part of speech of each word after the word is split, for example, it can be noun, verb, and adjective annotation.

[0207] In this embodiment, the interaction object refers to the interaction entity represented by the interaction data to be analyzed, for example, it may be the interaction content developed for a certain object or topic.

[0208] In this embodiment, speech logic refers to the thinking logic and speaking style represented by the data information input by the user during the interaction process.

[0209] In this embodiment, analogy analysis refers to analyzing the speech logic of the same interactive object in different interactive scenarios, in order to determine the user's interactive behavior habits (including speaking style and word usage habits, etc.).

[0210] In this embodiment, the interaction features refer to the methods and ways in which users interact with the intelligent customer service, as well as their speaking style and habits.

[0211] In this embodiment, the personalized reply strategy refers to a specific method or expression suitable for interacting with the user, which is formulated according to the user's interactive behavior habits.

[0212] In this embodiment, preprocessing refers to performing semantic extraction and vocabulary information extraction on each round of interaction data.

[0213] In this embodiment, the emotion keywords refer to text words related to emotions.

[0214] In this embodiment, the sentiment dictionary is constructed in advance and includes sentiment states corresponding to different words.

[0215] In this embodiment, the emotional tendency refers to the type of emotion represented by the emotional keyword, for example, it can be negative or positive.

[0216] In this embodiment, the tone symbol refers to a special symbol that can represent the current speaking tone of the user, such as an exclamation mark or an ellipsis mark.

[0217] In this embodiment, the tone feature refers to a specific tone situation represented by a tone symbol, such as emphasis, happiness, anger, etc.

[0218] In this embodiment, the response tone refers to the tone that needs to be adopted to deal with the user's interactive emotions. For example, if the user is currently in a negative state and speaks angrily, the response tone needs to be smooth in order to alleviate the user's angry state.

[0219] The working principle and beneficial effects of the above technical solution are: by organizing and analyzing the interaction data of each round of users, the interaction behavior habits of users during the interaction can be effectively determined, so as to facilitate the formulation of personalized reply strategies for users based on the interaction behavior habits, thereby improving the effect of intelligent replies. Secondly, the obtained personalized reply strategies are integrated and deployed in the replacement large language model, and finally reliable intelligent replies are achieved, which improves the user's interactive experience. Finally, the user's current emotional state is monitored in real time during the dynamic reply process, and corresponding response measures are taken according to the user's current emotional state to provide effective intelligent replies to the user, ensuring that the user's emotional needs are met without affecting the reliability of the reply content, thereby improving the accuracy and reliability of intelligent replies and ensuring the effect of intelligent replies.

[0220] In one embodiment, an intelligent customer service system based on an AI big model is provided, wherein the system monitoring and optimization module further includes evaluating and monitoring the system:

[0221] A first calculation unit, used for reading a data response parameter of system operation data, and calculating the operation efficiency of the system according to the response parameter of the system operation data;

[0222] ;

[0223] in, Indicates the operating efficiency of the system; represents the best response time; is the actual response time; Indicates system response delay, and ; represents the best response rate; represents the actual response rate, and ; Indicates system operation data; Indicates that in the system operation data, the correct system operation data is responded; Indicates the system operation data that responds to errors in the system operation data;

[0224] The second calculation unit is used to randomly collect the number of customer service visits within a number of preset time periods, and at the same time, obtain the satisfaction score of each customer for the visit interaction results within each preset time period;

[0225] Calculate the customer's comprehensive satisfaction rating of the system based on the rating of each customer's access interaction results in each preset time period;

[0226] ;

[0227] in, Indicates the customer's comprehensive satisfaction rating of the system; Indicates the total number of preset time periods; Indicates the serial number value of the preset time period; Indicated in The total number of customers within a preset time period; Indicated in The customer serial number value within a preset time period; Indicated in The first Customer satisfaction rating of the system;

[0228] System evaluation monitoring report generation unit, used to:

[0229] Generate a first evaluation monitoring report for evaluating the system according to the operating efficiency of the system;

[0230] Generate a second evaluation monitoring report for evaluating the system based on the customer's comprehensive satisfaction rating of the system;

[0231] The report transmission unit is used to transmit the first evaluation monitoring report and the second evaluation monitoring report to the monitoring terminal.

[0232] In this embodiment, the optimal response time is set in advance and is the fastest system response time obtained through multiple experiments and experience.

[0233] In this embodiment, the optimal response rate is set in advance and is the fastest system response speed obtained through multiple experiments and experience.

[0234] In this embodiment, the first evaluation monitoring report is a result of the system's operating efficiency during operation response obtained based on the evaluation of the system operation data, which is helpful for timely grasping the dynamics of the current system operation efficiency and is an evaluation of the system's internal performance.

[0235] In this embodiment, the second evaluation monitoring report is a result of the customer's comprehensive satisfaction score for the system obtained after evaluating the customer's satisfaction score, which is conducive to timely grasping the user's evaluation of the external satisfaction with the system.

[0236] In this embodiment, the data response parameters of the system operation data include: response time, response rate, data of accurate operation data response, and data of incorrect operation data response.

[0237] The working principle and beneficial effects of the above technical solution are: by calculating the system's operating efficiency and the customer's comprehensive satisfaction score of the system, the first evaluation monitoring report and the second evaluation monitoring report are effectively obtained, and then the internal performance and external satisfaction evaluation results of the system are effectively realized, thereby improving the comprehensiveness of the system evaluation and facilitating timely grasp of the current system operation status.

[0238] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. Intelligent customer service system based on AI big model, characterized by: include: User interaction module, used to: Receive multimodal consultation data input by the user, call the service interface to obtain the result according to the user's request type, and convert the response data into one or more of natural language text, voice reply and visual image data; Natural language parsing module, used for: Perform word segmentation, part-of-speech tagging, and syntactic analysis on the content input by the user, capture context-related semantic information, match the user's request type and intention, extract key entities, time information, and context information elements, and generate structured semantic representation data; Knowledge base management module, used to: Integrate and normalize static knowledge data and dynamic user interaction data, vectorize knowledge content and build an index structure, and compare the answers generated by the large language model with the knowledge base content; AI decision making module for: Based on user intent and knowledge base query results, a large language model is called to generate preliminary response content, and rule reasoning is introduced to make step-by-step decisions and execute tasks. The format, language style, and business consistency of the generated content are controlled by fine-tuning the model and inserting prompts. Multi-round dialogue management module, used for: Build a dialogue state tracker to update the state of each round of interaction input by the user, record the intent, entity and context information, replace the large language model with the dialogue history to generate dynamic replies, analyze user goals based on the dialogue path and predict subsequent operations, and call different models and logic modules for dialogue scenarios of different complexity; System monitoring and optimization module for: Collect system operation data in real time, detect abnormal situations through data analysis tools, and generate a visual management interface for viewing the operation status and manually intervening in key nodes; Wherein, the AI ​​decision generation module includes: User intent processing unit, used to: Compare the intent recognition results of the natural language parsing module with the entry content matched in the knowledge base to determine the coverage of user intent and knowledge content; If the user's intention is clear and supported by corresponding knowledge items, the content generation optimization process will be started; If the user's intention is unclear or the match with the knowledge base content is low, the generative large model is called to supplement the content and record the unmatched knowledge points; Among them, in the system monitoring and optimization module, the system operation data includes response time, user satisfaction and error rate, and the abnormal conditions include model deviation, knowledge base missing and high-frequency user complaints; The system monitoring and optimization module also includes evaluating and monitoring the system: A first calculation unit, used for reading a data response parameter of system operation data, and calculating the operation efficiency of the system according to the response parameter of the system operation data; ； in, Indicates the operating efficiency of the system; represents the best response time; is the actual response time; Indicates system response delay, and ; represents the best response rate; represents the actual response rate, and ; Indicates system operation data; Indicates that in the system operation data, the correct system operation data is responded; Indicates the system operation data that responds to errors in the system operation data; The second calculation unit is used to randomly collect the number of customer service visits within a number of preset time periods, and at the same time, obtain the satisfaction score of each customer for the visit interaction results within each preset time period; Calculate the customer's comprehensive satisfaction rating of the system based on the rating of each customer's access interaction results in each preset time period; ； in, Indicates the customer's comprehensive satisfaction rating of the system; Indicates the total number of preset time periods; Indicates the serial number value of the preset time period; Indicated in The total number of customers within a preset time period; Indicated in The customer serial number value within a preset time period; Indicated in The first The customer satisfaction rating of the system.

2. The intelligent customer service system based on the AI ​​big model as claimed in claim 1, characterized in that: The user interaction module comprises: The data receiving and identifying unit is used for: Receiving consultation requests input by users through text, voice and image channels, and identifying data types, wherein the data types include text data, voice data, image data and combined data; For speech data, extract speech features and generate corresponding text content through the text conversion module; For image data, key visual information is extracted through image features to generate image description content; Data pre-processing unit, used for: Clean, format and standardize received user input data; Response generation display unit for: The best response form is determined based on the output content of the natural language parsing module and the AI ​​decision generation module, and the response form includes text, voice, image and a combination thereof.

3. The intelligent customer service system based on the AI ​​big model as claimed in claim 2, characterized in that: When generating the best response form, the response generation and display unit adjusts the language style according to the tone and intention of the user input and annotates the professional terms and complex information when replying in the form of a text response; When replying in the form of voice response, the text content is converted into a voice response, and the voice tone is adjusted according to the results of the user's sentiment analysis; When replying in the form of an image response, generate charts, visualizations, and image content based on user needs, and annotate key information points.

4. The intelligent customer service system based on the AI ​​big model as claimed in claim 1, characterized in that: The natural language parsing module comprises: Semantic analysis unit, used to: Segment the user input text, generate word sequences and mark the grammatical structure, assign part-of-speech tags to each word through part-of-speech tagging, analyze the dependencies between sentence components, and generate the semantic vector of the sentence; Intent recognition unit, used to: Classify semantic vectors to determine the user's request type, which includes question and answer, transaction processing, and complaint. Track the intent of continuous input in multiple rounds of dialogue to determine whether the user changes the intent or continues the current task. When the user expresses multiple intents in one input, separate the sub-intents. An information extraction unit, for: Extract entity information from user input, including names of people, places, organizations, and product names; extract time information from user input, including date, time, and time interval, and standardize the time information; extract quantity information and quantitative description from the input and map them to structured fields; Identify the core problem or specific description of the need in the input; In a multi-round dialogue scenario, context information is used to supplement and update unclear elements to generate complete structured semantic representation data, where the unclear elements include one or more of omitted subjects and time.

5. The intelligent customer service system based on the AI ​​big model as claimed in claim 1, characterized in that: The knowledge base management module includes: Knowledge Integration Unit for: Receive and import static knowledge data, including FAQs, corporate policy documents, and technical manuals, and dynamically collect user interaction dynamic data, including historical conversation records, real-time data streams, and user behavior logs; Perform format parsing and semantic splitting on static knowledge data documents, divide long static knowledge data contents into independent knowledge items, and update dynamic data in real time; Knowledge vectorization unit, used for: Generate high-dimensional vector representations of FAQ items and document contents through embedding models, capture the semantic features of data, and introduce time and context information in the dynamic data embedding process; Build a vector index structure and add a category tag to each vectorized data; Knowledge comparison unit, used to: Receive structured semantic representation data and convert it into a query vector; Semantically match the query vector with the knowledge vector library to obtain the most relevant knowledge items; The matching knowledge items are trimmed and reorganized, and the results with large semantic deviations are eliminated according to the semantic similarity threshold to obtain the retrieval results; Mark frequently called and high error rate entries for update and optimization.

6. The intelligent customer service system based on the AI ​​big model as claimed in claim 1, characterized in that: The AI ​​decision making module further includes: Rule reasoning unit, used to: Parse key elements in user intent and match them with the business rule library, including entity, time, and operation type; Build task execution process; For transaction processing tasks, user requests are decomposed into multiple subtasks, and execution sequences are generated according to priorities and dependencies. For complex complaint tasks, reasoning logic is called to determine the best processing solution and additional information is added, and intermediate decision data is generated based on the reasoning results. Generate content optimization units for: Input user intent and knowledge base comparison results, and generate preliminary response content based on intermediate decision data; Optimize the generated content, control the format, logical level and semantic integrity of the generated content through preset templates and task prompts, and verify the numbers, time and calculation content; Eliminate potential banned words and sensitive content through business rules and semantic review.

7. The intelligent customer service system based on the AI ​​big model as claimed in claim 1, characterized in that: The multi-round dialogue management module includes: Dialog state tracking unit, used to: At the beginning of a conversation, a blank conversation state container is constructed, which contains the following information fields: intent, entity, time, context, and unfinished task list, identifies the intent and key entity information input by the user in the first round, and fills in the initial state data; For each round of user input, the natural language parsing module is called to extract the intent and information elements, and then compared and merged with the initial state data in the current dialog state container; After the user interrupts the conversation, the current conversation state data is stored and the key nodes of the conversation are recorded. When resuming the conversation, the stored conversation state data is read to prompt the user to resume to the interrupted position; Dialogue path prediction unit, used to: Based on the current dialogue state and user input, combined with the intention recognition results of the natural language parsing module, determine the user's goal type, which includes short-term goals and long-term goals. The short-term goals include single-round question-answering and transaction processing, and the long-term goals include multi-step tasks and complex problem solving; Use the conversation history to build a path model, associate the user's past operation sequence with the knowledge base content, and generate a list of possible follow-up operation recommendations; The predicted subsequent path is presented to the user in the form of visualization and text suggestions; Monitor the user's response to route suggestions in real time. When the user deviates from the recommended route, re-evaluate the dialogue goal and generate new route suggestions.

8. The intelligent customer service system based on the AI ​​big model as claimed in claim 7, characterized in that: The multi-round dialogue management module further includes: A data collection unit, used to collect each round of interaction data input by the user based on a preset frequency, and to organize the collected each round of interaction data to obtain interaction data to be analyzed; Data analysis unit for: Performing word segmentation on the interaction data to be analyzed, and performing part-of-speech tagging on the word segmentation results based on the natural language system, and determining the interaction objects and speech logic corresponding to the interaction data to be analyzed based on the part-of-speech tagging results; Conduct analogical analysis on the speech logic of the interactive object in different interactive scenarios, determine the user's interactive characteristics based on the analogical analysis, and obtain the user's interactive behavior habits based on the interactive characteristics; Smart reply configuration unit, used to: Generate personalized reply strategies for users based on interactive behavior habits, integrate and deploy the personalized reply strategies in the replacement large language model, and dynamically reply to users based on the integrated deployment results; Smart reply optimization unit, used to: Preprocess each round of user interaction data based on the dynamic reply results, and determine the text structure and text semantics of the interaction text based on the preprocessing results; Extract emotional keywords based on text structure and text semantics, and match the emotional keywords with the pre-built emotional dictionary to determine the emotional tendency of the emotional keywords; Extract tone symbols from interactive texts based on text structure, and determine the user's tone features based on tone symbols. At the same time, determine the user's current interaction scenario based on text semantics, and obtain the user's current interaction emotions based on tone features, interaction scenarios, and emotional tendencies. Determine the response tone to the user's dynamic reply content based on the interaction emotion, match available tone words from a preset vocabulary library based on the response tone, and logically associate the available tone words with the user's dynamic reply content; Dynamically respond to users based on logical key results.

9. The intelligent customer service system based on the AI ​​big model as claimed in claim 1, characterized in that: The system monitoring and optimization module further includes: System evaluation monitoring report generation unit, used to: Generate a first evaluation monitoring report for evaluating the system according to the operating efficiency of the system; Generate a second evaluation monitoring report for evaluating the system based on the customer's comprehensive satisfaction rating of the system; The report transmission unit is used to transmit the first evaluation monitoring report and the second evaluation monitoring report to the monitoring terminal.

Images