Intelligent interactive customer service system based on AI large model

Abstract

The application particularly relates to the technical field of natural language processing, and discloses an intelligent interactive customer service system based on an AI large model, which comprises a multi-modal input analysis module, an intention preliminary detection module, a large model semantic understanding module, a dynamic knowledge base association module, a dialogue strategy generation module, a natural language generation module and a service feedback work order module; through the deep cooperation of the multi-modal input analysis module and the large model semantic understanding module, entity and attribute information can be accurately extracted, and context semantic association information can be mined; through the cooperative action of the intention preliminary detection module, the dynamic knowledge base association module and the dialogue strategy generation module, the matching calculation is carried out based on the semantic analysis result output by the large model and the constructed dynamic knowledge base, rich knowledge support is provided for the solution of complex problems, personalized knowledge content is recommended for different users through the natural language generation module and the service feedback work order module, and the reply matching degree and the complex problem processing capability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to an intelligent interactive customer service system based on a large AI model. Background Technology

[0002] With the advancement of business digitalization and the increasing demands for consumer service experience, customer service, as a core link, urgently needs to transform from traditional manual service to intelligent and automated service. In order to improve customer service efficiency and reduce operating costs, intelligent customer service centers have become a key solution. With the development of natural language processing technology, it has provided strong technical support for the development of intelligent interactive systems.

[0003] Existing intelligent interactive customer service systems include a rules engine module, a knowledge base module, a voice recognition module, a dialogue management module, and a work order system module. They have the advantages of instant response, stable service, data analysis to assist decision-making, and low cost, providing stable service to customers and strong support for enterprise operation optimization.

[0004] However, it still has some shortcomings in actual use. First, its semantic understanding ability is limited. Existing intelligent interactive customer service systems are usually based on rule engines and keyword matching technology, which can only identify pre-set specific patterns and keywords, making it difficult to accurately understand the user's true intentions, resulting in low matching degree of the reply content.

[0005] Second, the ability to handle complex problems is insufficient. The existing intelligent interactive customer service system has relatively fixed dialogue logic, lacks flexibility and contextual understanding ability, and only processes according to preset processes and rules. When the user's problem exceeds the preset range or has many steps, it is difficult to effectively handle customer problems, and it cannot adjust the processing logic according to the real-time changes of the problem.

[0006] Third, the system lacks personalized service capabilities. Existing intelligent interactive customer service systems lack the ability to accumulate and call up user profiles, and cannot distinguish user identities and needs preferences, resulting in a lack of targeted services. Moreover, existing systems mostly use fixed script templates, which cannot adjust service strategies according to the context of the scenario, and cannot provide personalized knowledge recommendations. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide an intelligent interactive customer service system based on an AI large model. Through the deep collaboration between the multimodal input parsing module and the large model semantic understanding module, and through the synergistic effect of the intent preliminary detection module, the dynamic knowledge base association module, and the dialogue strategy generation module, it possesses powerful capabilities for handling complex problems and effectively solves the problems of limited semantic understanding capabilities, insufficient capabilities for handling complex problems, and lack of personalized service capabilities mentioned in the background art.

[0008] To achieve the above object, the present application provides the following technical solutions: an intelligent interactive customer service system based on an AI large model, comprising a multi-modal input analysis module, an intent preliminary detection module, a large model semantic understanding module, a dynamic knowledge base association module, a dialogue strategy generation module, a natural language generation module, and a service feedback work order module:

[0009] The multi-modal input analysis module comprises an input unit and an analysis unit, wherein the input unit is configured to receive multi-modal user interaction demand information uploaded by a user, and the analysis unit is configured to obtain structured multi-modal user interaction demand information.

[0010] The intent preliminary detection module comprises an information screening unit and an intent anchoring unit, wherein the information screening unit is configured to retain valid information in the structured multi-modal user interaction demand information, and the intent anchoring unit is configured to output user preliminary intent data.

[0011] The large model semantic understanding module performs deep semantic analysis on the user preliminary intent data based on a dialogue AI large model, and outputs semantic analysis results, including semantic vector representation, accurate intent classification results, entity and attribute extraction results, and context semantic association information.

[0012] The dynamic knowledge base association module constructs a dynamic knowledge base, calculates user demand matching degrees based on the semantic analysis results and the dynamic knowledge base, and outputs user demand matching results, including a candidate knowledge item set, an optimal knowledge recommendation result, and context supplement information.

[0013] The dialogue strategy generation module dynamically formulates user interaction strategies based on the semantic analysis results and the user demand matching results, and outputs a structured dialogue execution instruction set, including interaction target instructions, dialogue flow instructions, dialogue style content instructions, and feedback optimization instructions.

[0014] The natural language generation module receives the output semantic analysis results, the user demand matching results, and the structured dialogue execution instruction set, converts them into generation prompts, outputs personalized reply texts through the dialogue AI large model, and supports multi-modal reply output.

[0015] The service feedback work order module automatically generates natural language feedback inquiries, receives user feedback information, performs effect judgment based on the user feedback information, outputs work order distribution instructions, and supports synchronous multi-channel pushing of work order progress information.

[0016] The present application has the following technical effects and advantages:

[0017] 1. This invention, through the deep collaboration between the multimodal input parsing module and the large model semantic understanding module, comprehensively receives various forms of interactive request information uploaded by users. The large model semantic understanding module, relying on the powerful capabilities of the conversational AI large model, performs deep semantic analysis on the preliminary user intent data after initial processing, accurately extracts entity and attribute information, and mines contextual semantic association information, greatly improving the response matching degree.

[0018] 2. This invention possesses powerful capabilities for handling complex problems through the synergistic effect of the preliminary intent detection module, the dynamic knowledge base association module, and the dialogue strategy generation module. The preliminary intent detection module filters out effective information from user demand information and outputs preliminary user intent data, helping the system quickly grasp the core direction of the problem. The dynamic knowledge base association module matches and calculates the semantic analysis results output by the large model with the constructed dynamic knowledge base, providing rich knowledge support for solving complex problems and improving the ability to handle complex problems.

[0019] 3. This invention combines the contextual semantic association information mined by the large model semantic understanding module with the user data accumulated by the system. Based on user profiles and semantic analysis results, the dynamic knowledge base association module recommends personalized knowledge content to different users. The natural language generation module combines semantic analysis results, user demand matching results and structured dialogue execution instruction set to adjust personalized text according to user profiles and supports multimodal output, thereby improving personalized service capabilities. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0021] Figure 2 This is a schematic diagram of the method steps of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the steps for obtaining the semantic analysis results of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] As attached Figure 1The AI-based large model intelligent interaction customer service system shown comprises a multi-modal input analysis module, an intent preliminary detection module, a large model semantic understanding module, a dynamic knowledge base association module, a dialogue strategy generation module, a natural language generation module and a service feedback work order module.

[0025] It needs to be specifically pointed out that the output end of the multi-modal input analysis module is connected to the input end of the intent preliminary detection module, the output end of the intent preliminary detection module is connected to the input end of the large model semantic understanding module, the output end of the large model semantic understanding module is connected to the input end of the dynamic knowledge base association module, the output end of the dynamic knowledge base association module is connected to the input end of the dialogue strategy generation module, the output end of the dialogue strategy generation module is connected to the input end of the natural language generation module, and the output end of the natural language generation module is connected to the input end of the service feedback work order module.

[0026] The specific embodiments of the application include the following contents:

[0027] The multi-modal input analysis module comprises an input unit and an analysis unit, wherein the input unit is used to receive multi-modal user interaction demand information uploaded by the user, and the analysis unit is used to obtain structured multi-modal user interaction demand information.

[0028] Further, the multi-modal user interaction demand information comprises text type input information, voice type input information, visual type input information and interaction context type input information.

[0029] In this embodiment, it needs to be specifically pointed out that the text type input information comprises instant dialogue text, form / questionnaire text, document / attachment text and specific format text; the voice type input information comprises natural voice instruction / appeal, auxiliary information in voice and voice-to-text intermediate result; the visual type input information comprises image information containing text, object / scene type image information and image metadata information; the interaction context type input information comprises historical interaction context, user identity / scene context and system preset context, wherein the system preset context refers to the basic information related to the user built in the customer service system, including but not limited to the order list bound to the user account and the region where the user is located.

[0030] It needs to be specifically pointed out that the parsing unit receives the multi-modal user interaction demand information of the input unit, performs speech to text and image text extraction for voice type input information and visual type input information, and outputs text semantic in a unified format; based on the rule engine and regular expression, the structured parameters in the user demand are extracted from the text semantic content after parsing, including but not limited to order number, commodity type, time and amount, and the key parameters in the key-value pair format are output; based on the text semantic, key parameters and original modal identifier, the multi-modal information of the same user is associated according to "session ID + timestamp", and a structured data packet, i.e. structured multi-modal user interaction demand information, is generated.

[0031] The intent preliminary detection module includes an information screening unit and an intent anchoring unit, wherein the information screening unit is used to retain valid information in the structured multi-modal user interaction demand information, and the intent anchoring unit is used to output user preliminary intent data;

[0032] In this embodiment, it needs to be specifically pointed out that the information screening unit identifies invalid information in the structured multi-modal user interaction demand information through rule matching and lightweight semantic judgment, including garbled code / meaningless character, test statement and repeated submission information, retains the content containing demand keywords or key parameters in the structured multi-modal user interaction demand information, and outputs valid information.

[0033] Further, the user preliminary intent data acquisition needs the intent anchoring unit to predefine the core intent classification of the customer service scene, perform keyword matching and semantic similarity calculation, and output the user preliminary intent data, including valid information, preliminary intent label, matching confidence and associated key parameters.

[0034] In this embodiment, it needs to be specifically pointed out that the preset of the core intent classification is divided in two dimensions according to the user core appeal and business processing logic, including primary intent classification and secondary intent classification, the primary intent includes order related, after-sales related, logistics related, commodity related and member related, and the secondary intent includes but is not limited to order inquiry, order modification, refund application, logistics inquiry, commodity consultation and member benefit consultation, each secondary intent is assigned a unique code, associated with intent description, typical keyword library and key business parameters, and forms an intent classification dictionary.

[0035] It needs to be specifically pointed out that the keyword matching needs to be based on the intent classification dictionary combined with the historical user upload data to dynamically supplement a dynamic keyword library, and the accurate matching and fuzzy matching are performed on the valid information and the dynamic keyword library, the intent of the hit keyword is taken as a candidate intent, and the keyword hit times of each candidate intent are counted;

[0036] The semantic similarity calculation specifically inputs the effective information and the standard description of each candidate intent into a lightweight pre-trained semantic model to generate a text semantic vector, calculates the semantic similarity of the effective information and each candidate intent through a cosine similarity formula, selects the candidate intent with the highest semantic similarity as the preliminary intent label, and takes the semantic similarity value of the candidate intent as the matching confidence.

[0037] The acquisition of the associated key parameters needs to use a combination of regular expressions and named entity recognition to extract the business parameters associated with the preliminary intent in the effective information.

[0038] The large model semantic understanding module: based on the dialog AI large model, performing deep semantic analysis on the user preliminary intent data, outputting the semantic analysis result, including semantic vector representation, accurate intent classification result, entity and attribute extraction result, and context semantic association information;

[0039] Further, as shown in Figure 3 The acquisition steps of the semantic analysis result are as follows:

[0040] S1.1: converting the received user preliminary intent data into standardized text, converting the standardized text into a model computable Token sequence and initial word vector, and outputting the initial word vector matrix;

[0041] In this embodiment, it needs to be specifically explained that the effective content in the user preliminary intent data is filtered through regular expressions, and the colloquial expressions are replaced based on the preset synonym dictionary to obtain the standardized text; the text is segmented into subwords Token by the model Tokenizer, each Token is mapped to a unique ID in the model dictionary to form an ID sequence, and the ID sequence is converted into a fixed-dimension initial word vector by the model Embedding layer to obtain the initial word vector matrix.

[0042] S1.2: importing the initial word vector matrix into the dialog AI large model, injecting position coding, identifying the association relationship of the Token inside the text, and outputting a global semantic coding matrix;

[0043] In this embodiment, it needs to be specifically explained that injecting position coding specifically refers to calculating the position coding vector of each Token through the sine cosine formula, adding the position coding vector and the initial word vector matrix element by element to obtain a word vector matrix with time sequence information, obtaining a self-attention feature matrix through linear projection, calculating attention score and multi-head merging, obtaining a global semantic coding matrix by performing nonlinear transformation on the self-attention feature matrix.

[0044] S1.3: generating a semantic vector representation based on the global semantic coding matrix, calculating an accurate intent classification, extracting entities and attributes, and calculating upper and lower semantic associations, and outputting a semantic analysis result.

[0045] In this embodiment, it needs to be specifically pointed out that the semantic vector representation refers to extracting the semantic vector of the whole sentence from the global semantic encoding matrix to obtain the text semantic vector; the calculation of the accurate intent classification refers to judging the probability that the user input belongs to the preset sub-intent, taking the text semantic vector to perform linear mapping through the classification weight matrix to generate an original score, converting the original score into the probability of each intent, selecting the intent with the maximum probability as the accurate intent label, the corresponding probability as the confidence, and outputting the accurate intent classification result; the extraction of the entity and the attribute refers to generating the entity label score of each Token through the NER weight matrix of the global semantic encoding matrix, taking the label with the maximum probability after Softmax, performing entity merging and attribute association, calculating the attention weight of the entity Token and the attribute description Token, and associating the attribute when the weight is greater than 0.7, and outputting the entity and attribute extraction result; the calculation of the upper and lower semantic association refers to generating the historical semantic encoding matrix of the multi-round dialogue history text, calculating the cosine similarity of the current and historical semantic encoding matrix, and outputting the context semantic association information.

[0046] The dynamic knowledge base association module: constructs a dynamic knowledge base, calculates the user demand matching degree based on the semantic analysis result and the dynamic knowledge base, and outputs the user demand matching result, including the candidate knowledge item set, the optimal knowledge recommendation result and the context supplement information;

[0047] Further, the dynamic knowledge base supports multi-source data import, and can realize automatic knowledge update through timed crawling and API interface synchronization. The knowledge in the dynamic knowledge base is stored in the form of “semantic vector + text description”.

[0048] In this embodiment, it needs to be specifically pointed out that the multi-source data import includes but is not limited to enterprise official website policy documents, commodity manual updates and activity rule adjustments. The dynamic knowledge base includes four layers of architecture, namely the data layer, the structure layer, the index layer and the update layer. The dynamic knowledge base establishes the association relationship between the knowledge by constructing the knowledge graph, and constructs multi-dimensional indexes, including semantic vector index, attribute index and association index.

[0049] Further, the calculation steps of the user demand matching degree are as follows:

[0050] S2.1: Based on the semantic analysis result and the dynamic knowledge base, obtain the user demand analysis parameters, including the user demand semantic vector, the knowledge base semantic vector, the user demand entity type set, the knowledge association entity type set, the user demand time and the knowledge effective time;

[0051] In this embodiment, it is necessary to specifically explain that the user demand semantic vector is obtained through the user's initial intent data, the knowledge semantic vector is obtained through the dynamic knowledge base, the user demand entity type set and user demand time are obtained based on the semantic analysis results output by the large model semantic understanding module, and the knowledge associated entity type set and knowledge effective time are obtained through the dynamic knowledge base. The knowledge effective time refers to the "effective start time" and "effective end time" in the metadata of the knowledge entry in the dynamic knowledge base.

[0052] S2.2: Based on user requirement analysis parameters, the semantic vector matching degree S and entity type overlap degree M are calculated. t Entity attribute matching degree M a Timeliness adaptability (F), contextual relevance (R), and knowledge confidence (C);

[0053] In this embodiment, it should be specifically noted that the semantic vector matching degree S is calculated using the cosine similarity formula to determine the spatial correlation between the user's demand semantic vector and the knowledge base semantic vector, retaining knowledge with a semantic vector matching degree ≥ 0.6; the entity type overlap degree M... t It needs to be based on the user's requirement entity type set E u The set of entity types E associated with knowledge d Through the formula:

[0054] ,

[0055] The calculation yields a value where Num represents the number of geometric elements, and M is retained. t Knowledge with a value ≥0.5; Obtaining the entity attribute matching degree F requires extracting the key attributes of the user requirement entity, extracting the usage attribute constraints of the knowledge base, and calculating it through the attribute weighting formula, retaining M. a Knowledge level ≥ 0.6; Timeliness adaptation is calculated based on user demand time T. u and the effective time range of knowledge [T] ds T de ], through the formula:

[0056] ,

[0057] The calculation yields ΔT. th For a time threshold, such as 30 days, if the time exceeds 30 days, F=0, indicating the knowledge is expired; knowledge with F≥0.7 is retained. Contextual relevance R is calculated by obtaining historical global semantic vectors and using the cosine similarity formula to determine the relevance between the historical global semantic vectors and the knowledge semantic vectors; knowledge with R≥0.7 is retained. Knowledge confidence C is obtained by extracting the manual verification status V and the last update day O from the knowledge base metadata, using the formula:

[0058] ,

[0059] The user demand matching degree SI is calculated, where K represents the number of the meeting dimensions, and K / 6 represents the dimension meeting rate.

[0060] S2.3: Define the meeting dimensions and the key dimensions, where the key dimensions are the semantic vector matching degree, the entity attribute matching degree, and the timeliness adaptation degree, and the key dimensions are determined by the following formulas:

[0061]

[0062] The user demand matching degree SI is calculated, where K represents the number of the meeting dimensions, and K / 6 represents the dimension meeting rate.

[0063] In this embodiment, it needs to be specifically pointed out that the meeting dimensions refer to the number of dimensions that meet the screening rules, and there are 6 dimensions. The dimension meeting rate reflects the overall adaptability of each dimension, and the average score of the key dimensions reflects the core adaptability. The user demand matching degree ranges from 0 to 1, and the higher the value, the higher the overall user demand matching degree.

[0064] It needs to be specifically pointed out that the candidate knowledge item set is preliminarily screened by the user demand matching degree, and the knowledge items with a user demand matching degree greater than or equal to 0.7 are retained. It is ensured that the candidate knowledge item set meets the semantic vector matching degree greater than or equal to 0.6, the entity attribute matching degree greater than or equal to 0.6, and the timeliness adaptation degree greater than or equal to 0.8. Based on the comprehensive matching degree, the knowledge confidence, and the update time, the candidate knowledge item set is arranged in descending order, and the basic information, the matching degree data, and the association identifier are output.

[0065] The optimal knowledge recommendation result needs to be obtained based on the candidate knowledge item set through multi-dimensional weighted decision and business rule verification. The user demand matching degree, the entity attribute matching degree, the context correlation degree, and the knowledge priority are weighted and calculated to obtain a score. The weights are 0.4, 0.3, 0.2, and 0.1, respectively. The knowledge with the highest score is selected as the optimal result. If the optimal knowledge has exceptional circumstances, it needs to be further compared with the user entity attribute for the second time to ensure that there is no conflict. The complete content of the knowledge, the recommendation basis, and the confidence level are output.

[0066] The context supplementary information needs to clearly indicate the matching relationship between the knowledge and the entity in the user demand. The core semantic key points of the user demand and the knowledge are extracted to explain the scene adaptation, and 1-2 pieces of knowledge that the user may need in the future are recommended. The association logic, the extended knowledge, and the supplementary prompt are output.

[0067] The dialogue strategy generation module: based on the semantic analysis result and the user demand matching result, a user interaction strategy is dynamically formulated, and a structured dialogue execution instruction set is output, including an interaction target instruction, a dialogue flow instruction, a dialogue style content instruction, and a feedback optimization instruction.

[0068] ​Further, the interaction target instruction includes a target type, a target description, a target priority, and a completion standard; the dialogue flow instruction includes a flow node, an exception handling rule, and a flow termination condition; the dialogue style content instruction includes a dialogue style, a core content element, a format requirement, and an entity embedding requirement; and the feedback optimization instruction includes a feedback listening type, an optimization strategy, an optimization priority, and a strategy update mechanism.

[0069] In this embodiment, it needs to be specifically pointed out that the interaction target instruction needs to be obtained based on the accurate intent classification result and the optimal knowledge recommendation result, the target type includes but is not limited to knowledge answer, intent clarification, and parameter supplement, and the target description refers to the specific action and content to be achieved; the dialogue flow instruction needs to be obtained based on the interaction target instruction and the context semantic association information, the flow node refers to the action node arranged according to the execution order; the dialogue style content instruction needs to be obtained based on the optimal knowledge recommendation result and the entity and attribute extraction result, the core content element refers to the knowledge points that must be included and the prohibited content, and the entity embedding requirement refers to how to naturally integrate the user entity information into the dialogue; the feedback optimization instruction needs to be obtained based on the candidate knowledge item set and the intent classification confidence, the feedback listening type refers to the user feedback scene that needs to be focused on, and the strategy update mechanism refers to the strategy parameter that is regularly optimized based on the feedback data.

[0070] The natural language generation module: receives the output semantic analysis result, the user demand matching result, and the structured dialogue execution instruction set, converts into a generation prompt, outputs a personalized reply text through a dialogue AI large model, and supports multi-modal reply output;

[0071] Further, the generation prompt includes a demand description, a knowledge support, and a context parameter.

[0072] The multi-modal reply output includes a voice reply output, a text-image reply output, and an interactive reply output.

[0073] In this embodiment, it needs to be specifically pointed out that the core appeal is based on the accurate intent classification result, and is used to clarify the core demand direction of the user; the knowledge support includes core knowledge content, knowledge association basis, and knowledge boundary limitation, the core knowledge content comes from the optimal knowledge recommendation result, the knowledge association basis comes from the context supplement information, and is used to explain the matching reason of the knowledge and the demand, and the context parameter comes from the candidate knowledge item set and the knowledge metadata, and is used to clarify the knowledge that cannot be used.

[0074] It needs to be further explained that the output of the personalized reply text needs to be connected to the external image system to receive the user portrait, and the user portrait data and historical interaction feedback data are optimized in depth, the user portrait data is optimized based on the age, membership level, historical interaction style and preferred language of the user portrait data, and the historical interaction feedback data is based on the historical interaction feedback of the user, and the current reply style is adjusted.

[0075] The service feedback work order module automatically generates a natural language feedback inquiry, receives user feedback information, performs an effect judgment based on the user feedback information, outputs a work order distribution instruction, and supports synchronous multi-channel pushing of work order progress information.

[0076] Further, the work order distribution instruction includes work order basic information, distribution target, execution requirement and associated data.

[0077] In this embodiment, it needs to be specifically explained that the execution effect judgment takes the user feedback information as the core basis, constructs a three-dimensional layered judgment rule, distinguishes four scenes of clear satisfaction, partial solution / fuzzy feedback, clear dissatisfaction and emergency / complaint, extracts the keywords in the user feedback information, matches the parsed feedback label with the preset rule, determines whether to generate a work order and the type of the work order, and outputs the conclusion of no work order / ordinary work order / emergency work order.

[0078] It needs to be specifically explained that the work order distribution needs to generate a work order distribution instruction based on the work order priority, demand type, user level and seat resource state, the emergency work order needs to add an emergency identifier and an upgrade mechanism based on the ordinary work order, and according to the demand type, the corresponding seat group is matched, the seat is filtered according to the preset priority rule, and the work order distribution instruction is sent to the seat workstation of the target seat.

[0079] As shown in Figure 2 The present application provides an intelligent interactive customer service method based on an AI large model, and the steps are as follows:

[0080] S1: receiving the multi-modal user interaction demand information uploaded by the user through the multi-modal input analysis module, and obtaining the structured multi-modal user interaction demand information;

[0081] S2: extracting the effective information in the structured multi-modal user interaction demand information through the intent preliminary detection module, and outputting the user preliminary intent data;

[0082] S3: performing deep semantic analysis on the user preliminary intent data based on the dialog AI large model through the large model semantic understanding module, and outputting the semantic analysis result;

[0083] S4: calculating the user demand matching degree based on the semantic analysis result and the dynamic knowledge base through the dynamic knowledge base association module, and outputting the user demand matching result;

[0084] S5: dynamically formulating the user interaction strategy based on the semantic analysis result and the user demand matching result through the dialogue strategy generation module, and outputting the structured dialogue execution instruction set;

[0085] S6: receiving the output semantic analysis result, the user demand matching result and the structured dialogue execution instruction set through the natural language generation module, and outputting the individualized reply text;

[0086] S7: automatically generating the natural language feedback inquiry through the service feedback work order module, receiving the user feedback information, executing the effect judgment based on the user feedback information, and outputting the work order distribution instruction.

[0087] Secondly, in the drawings of the disclosed embodiments, only the structures related to the disclosed embodiments are involved, other structures can refer to the general design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0088] Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An intelligent interactive customer service system based on an AI large model, characterized in that, Comprise multimodal input analysis module, intention preliminary detection module, large model semantic understanding module, dynamic knowledge base correlation module, dialogue strategy generation module, natural language generation module and service feedback work order module: The multimodal input analysis module comprises an input unit and an analysis unit, wherein the input unit is used for receiving user uploaded multimodal user interaction demand information, and the analysis unit is used for obtaining structured multimodal user interaction demand information; The intention preliminary detection module comprises an information screening unit and an intention anchoring unit, wherein the information screening unit is used for retaining valid information in the structured multimodal user interaction demand information, and the intention anchoring unit is used for outputting user preliminary intention data; The large model semantic understanding module performs deep semantic analysis on the user preliminary intention data based on a dialogue AI large model, and outputs a semantic analysis result, including a semantic vector representation, a precise intention classification result, an entity and attribute extraction result, and context semantic correlation information; The dynamic knowledge base correlation module constructs a dynamic knowledge base, calculates a user demand matching degree based on the semantic analysis result and the dynamic knowledge base, and outputs a user demand matching result, including a candidate knowledge item set, an optimal knowledge recommendation result, and context supplement information, and the calculation steps of the user demand matching degree are as follows: Based on the semantic analysis result and the dynamic knowledge base, user demand analysis parameters are obtained, including a user demand semantic vector, a knowledge base semantic vector, a user demand entity type set, a knowledge correlation entity type set, a user demand time, and a knowledge effective time; The semantic vector matching degree S, the entity type coincidence degree M, the entity attribute matching degree M, the timeliness adaptation degree F, the context correlation degree R, and the knowledge confidence C are calculated based on user demand analysis parameters t a a a a a a a a a a a a a a a a < Define the target dimension and the key dimension, wherein the key dimension is the semantic vector matching degree, the entity attribute matching degree, and the timeliness adaptation degree, and the user demand matching degree SI is calculated by the formula: ， The dialogue strategy generation module dynamically formulates a user interaction strategy based on the semantic analysis result and the user demand matching result, and outputs a structured dialogue execution instruction set, including an interaction target instruction, a dialogue flow instruction, a speech style content instruction, and a feedback optimization instruction; The natural language generation module receives the semantic analysis result, the user demand matching result, and the structured dialogue execution instruction set, converts them into a generation prompt, outputs an individualized reply text through the dialogue AI large model, and supports multimodal reply output; The service feedback work order module automatically generates a natural language feedback inquiry, receives user feedback information, performs effect judgment based on the user feedback information, outputs a work order distribution instruction, and supports synchronous multi-channel pushing of work order progress information. The multimodal user interaction demand information comprises text input information, voice input information, visual input information, and interaction context input information. 2.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: The user preliminary intention data acquisition needs the intention anchoring unit to predefine the core intention classification of the customer service scene, perform keyword matching and semantic similarity calculation, and output the user preliminary intention data, including valid information, a preliminary intention label, matching confidence, and associated key parameters. 3.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: The semantic analysis result is obtained by the following steps: 4.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: ​ S1.1: Transform the received user initial intention data into standardized text, transform the standardized text into a model computable Token sequence and initial word vector, and output the initial word vector matrix; S1.2: Import the initial word vector matrix into the dialog AI large model, inject position coding, identify the association relationship of the Token inside the text, and output the global semantic coding matrix; S1.3: Based on the global semantic coding matrix, generate semantic vector representation, calculate accurate intention classification, extract entities and attributes, and calculate upper and lower semantic association, and output the semantic analysis result. 5.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: The dynamic knowledge base supports multi-source data import, and can realize automatic knowledge update through timed crawling and API interface synchronization. The knowledge in the dynamic knowledge base is stored in the form of semantic vectors and text descriptions. 6.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: The interaction target instruction includes target type, target description, target priority, and completion standard; the dialogue flow instruction includes flow node, exception handling rule, and flow termination condition; the dialogue style content instruction includes dialogue style, core content element, format requirement, and entity embedding requirement; The feedback optimization instruction includes feedback listening type, optimization strategy, optimization priority, and strategy update mechanism. 7.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: The generation prompt contains requirement description, knowledge support, and context parameters; The multi-modal reply output includes voice reply output, graphic-text reply output, and interactive reply output. 8.The AI large model-based intelligent interaction customer service system according to claim 1, characterized in that: The work order allocation instruction includes work order basic information, allocation target, execution requirement, and associated data.

Images