Intelligent customer service intent understanding and response method and system based on big data analysis

Abstract

The application discloses an intelligent customer service intention understanding and response method and system based on big data analysis; through obtaining user consultation text and identification information, a standardized query sentence is generated through text cleaning, word segmentation, entity recognition and sentence completion, and time-aligned associated data packets are generated in combination with user static attributes and dynamic behavior data; after the query sentence and the associated data packets are encoded, cross-modal attention fusion is carried out, and an intention recognition result is output through a multi-task learning network; for non-complex problems, a rule base is matched to quickly respond, and for complex problems, an intention vector is used to search a knowledge base and the response is generated by a large language generation model; meanwhile, the portrait is updated according to user feedback behavior, and the model is incrementally fine-tuned under a trigger condition, so that the intention recognition accuracy, the complex problem response quality and the model self-adaptive evolution ability are systematically improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent customer service and artificial intelligence technology, specifically relating to an intelligent customer service intent understanding and response method and system based on big data analysis. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, intelligent customer service systems based on big data analytics have become a core infrastructure for enterprises to improve service efficiency and reduce operating costs. However, with the increasing complexity of business scenarios and the diversification and personalization of user inquiries, traditional customer service systems based on keyword matching and finite state machines are struggling to meet users' expectations for a natural, smooth, and intelligent interactive experience. In recent years, generative AI technologies, represented by large-scale pre-trained language models, have made groundbreaking progress, bringing a new technological paradigm to the field of intelligent customer service. An industry consensus is gradually forming: intelligent customer service systems need to evolve from passive question-and-answer tools to cognitive agents with deep semantic understanding, multi-turn dialogue management, and complex task reasoning capabilities, in order to truly understand user intent and efficiently resolve problems.

[0003] Most current mainstream intelligent customer service systems adopt an intent recognition and rule-based response architecture. This involves first classifying user input intent using a classification model, and then responding based on preset rules or simple dialogue flows. While this architecture excels in handling high-frequency, standardized consultation scenarios, it exhibits significant limitations when facing complex scenarios such as conversational expressions, multiple complex intents, and contextual dependencies. Specifically, intent recognition models suffer from a sharp drop in accuracy in real-world open-domain scenarios due to limitations in the quality and coverage of training data; rule-based dialogue management lacks flexibility and cannot handle open-ended questions that go beyond preset paths; and the system's utilization of contextual information such as user history and order information is severely insufficient, resulting in weak understanding of fuzzy queries. Furthermore, most of these systems are statically deployed, lacking effective data feedback and model iteration mechanisms, causing their capabilities to stagnate after deployment and making it difficult to learn and evolve from continuously generated real-world interaction data.

[0004] In summary, the fundamental bottlenecks of existing intelligent customer service systems can be attributed to the lack of three major capabilities: First, the lack of complex semantic understanding and generation capabilities, with models lacking the ability to deeply reason about open-domain issues and implicit intentions; second, the lack of multi-source context fusion capabilities, failing to effectively integrate cross-dimensional information such as user profiles, historical interactions, and business data to support accurate judgments; and third, the lack of continuous learning and self-evolution capabilities, resulting in the system freezing upon deployment and being unable to utilize massive amounts of real feedback data for closed-loop optimization. Summary of the Invention

[0005] (a) Technical problems to be solved In response to the problems in related technologies, this invention provides an intelligent customer service intent understanding and response method and system based on big data analysis to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: S1. Obtain the current query text input by the user and the user's user identification information, preprocess the current query text to obtain a standardized current query statement; S2. Based on the user identification information, retrieve the associated data corresponding to the user from the preset user data warehouse; S3. Using a multi-task learning network, the standardized current query statement and the associated data corresponding to the user are jointly encoded to generate an intent representation vector that integrates contextual features. The intent representation vector is then used to perform intent recognition on the current query text to obtain the intent recognition result. S4. Based on the intent recognition result, determine whether the current consultation text is a complex open-ended question; If not, then generate the first response result based on the rule base matching; If so, the standardized current query statement and the intent representation vector of the fused context features are used as retrieval conditions to retrieve candidate knowledge documents from the preset knowledge base. The retrieved candidate knowledge documents, the standardized current query statement, and the intent representation vector of the fused context features are then input into the large language generation model to generate a second response result. S5. Output the first response result or the second response result as the final response result to the user, obtain the user's feedback behavior data for the final response result, and associate and store the feedback behavior data, the standardized current query statement, the intent representation vector fused with context features and the final response result in the user data warehouse to update the associated data corresponding to the user and trigger incremental fine-tuning of the large language generation model and the multi-task learning network. This invention constructs a complete intelligent response closed-loop system, including user profile association, multi-task joint encoding, intent recognition and triage, rule-based and generation-based hybrid response, and feedback incremental learning. It abandons the single-point analysis mode of traditional customer service systems that relies solely on the current question-and-answer text for intent understanding, and instead utilizes user identification information to retrieve associated data, achieving context-aware and accurate intent recognition. Through a multi-task learning network, it performs cross-modal attention fusion of query statements and user features, enhancing the semantic richness of intent representation. For complex and open-ended questions, it employs a retrieval-enhanced generation mechanism, inputting candidate knowledge documents and intent representation vectors into a large language generation model, overcoming the limitations of pure rule matching in handling open-ended problems. By associating and storing feedback behavior data and triggering incremental model fine-tuning, it achieves continuous model evolution and personalized adaptation while maintaining the stability of existing knowledge.

[0007] Preferably, step S1 includes the following steps: S11. Obtain the current consultation text input by the user, wherein the current consultation text includes speech-converted text, keyboard input text and / or touch screen handwriting recognition text; S12. Obtain the user identification information of the user, the user identification information including user account, device unique identifier and / or biometric identifier; S13. Perform text cleaning on the current consultation text. The text cleaning includes removing invalid characters, correcting typos, unifying simplified and traditional Chinese characters and normalizing punctuation marks to obtain the cleaned text. S14. Perform word segmentation on the cleaned text and correct the word segmentation results using a preset industry-customized dictionary to obtain a word segmentation sequence. S15. Perform part-of-speech tagging and named entity recognition on the word segmentation sequence to extract key entity information from the current consultation text; S16. Based on the key entity information and the word segmentation sequence, perform sentence normalization and semantic completion, complete incomplete or omitted sentences into complete semantic units, and generate the standardized current query statement. By performing multi-level preprocessing on the current consultation text, including cleaning, word segmentation correction, part-of-speech tagging and named entity recognition, sentence structure standardization, and semantic completion, the input from multiple sources such as voice, keyboard, and handwriting is uniformly converted into standardized query statements. This eliminates semantic ambiguity and information loss caused by text noise, typos, mixed simplified and traditional characters, and omitted sentences, providing a high-quality text input foundation with unified format and complete semantics for subsequent cross-modal coding and intent recognition.

[0008] Preferably, step S2 includes the following steps: S21. Based on the user identification information, determine the unique index of the user corresponding to the user identification information from the preset user data warehouse; S22. Based on the user's unique index, retrieve the user's static attribute data and dynamic behavior data from the user data warehouse; S23. Perform data quality verification on the static attribute data and the dynamic behavior data, remove outlier and null value fields, and fill missing fields with default values ​​or mean interpolation to obtain the verified associated data. S24. The verified associated data is organized in a structured manner according to a preset feature field mapping table to generate an associated data packet that is time-aligned with the standardized current query statement, thereby obtaining the associated data corresponding to the user. By identifying a unique index for a user using user identification information, static attribute data and dynamic behavior data are retrieved from the user data warehouse. After data quality verification to remove outliers and null values ​​and fill in missing fields, the verified associated data is structured according to the feature field mapping table to generate an associated data package that is time-aligned with the current query statement. This eliminates the problems of feature loss and semantic deviation caused by the dispersed, inconsistent quality, and time misalignment of user data sources, ensuring that the user context information on which the intent recognition stage depends has completeness, accuracy, and temporal consistency, and providing a structured, aligned, and reliable user profile data foundation for cross-modal joint coding.

[0009] Preferably, step S24 includes the following steps: S241. Obtain a preset feature field mapping table, wherein the feature field mapping table is used to define the mapping relationship between each original field in the verified associated data and the preset standardized feature dimension. S242. Based on the feature field mapping table, convert each original field in the verified associated data into a corresponding standardized feature value to form a structured feature vector set; S243. Extract the receiving timestamp of the standardized current query statement, and determine the backtracking time window based on the receiving timestamp; S244. From the structured feature vector set, select a subset of feature vectors whose timestamps fall within the backtracking time window, as valid feature data aligned with the time of the standardized current query statement; S245. Assemble the effective feature data into a fixed-length feature matrix according to a preset feature dimension order, and normalize the fixed-length feature matrix to generate the associated data packet. S246. The associated data packet is timestamped and linked to the standardized current query statement to generate an associated data packet with a time tag, thereby obtaining the associated data corresponding to the user. By standardizing associated data through feature field mapping, the effective features for time-series alignment are determined by the query statement timestamp and the backtracking window is used to assemble them into a normalized fixed-length feature matrix and bind time tags to generate associated data packages. This eliminates data alignment deviations caused by field heterogeneity and time misalignment, and provides standardized inputs with consistent time and regular structure for cross-modal fusion.

[0010] Preferably, step S3 includes the following steps: S31. Input the standardized current query statement into the first encoder to extract semantic features and generate a query statement semantic vector; at the same time, input the associated data packet into the second encoder to extract user static features and dynamic behavior temporal features and generate a user context feature vector. S32. Align the semantic vector of the query statement and the user context feature vector in dimensions and map their feature spaces. Use the aligned semantic vector of the query statement as the query item and the aligned user context feature vector as the key item. Calculate the semantic association weight using a cross-modal attention fusion mechanism and weight the aligned user context feature vector to obtain a weighted fusion feature. S33. The semantic vector of the query statement, the user context feature vector, and the weighted fusion feature are concatenated and input into the shared hidden layer of the multi-task learning network. The intent representation vector of the fusion context feature is generated through nonlinear transformation. S34. Input the intent representation vector into the intent classification task branch and the sentiment analysis task branch of the multi-task learning network respectively to obtain the intent category probability distribution and sentiment polarity category. Take the intent category with the highest probability according to the intent category probability distribution as the preliminary intent recognition result, and correct the preliminary intent recognition result according to the sentiment polarity category to output the final intent recognition result. By using dual encoders to extract query semantic vectors and user context feature vectors respectively, and then dimensionally aligned, weighted fusion features are generated through cross-modal attention fusion. The three are then concatenated and fed into a multi-task sharing network to generate an intent representation vector, which is then input into the intent classification and sentiment analysis branches respectively. Sentiment polarity is used to correct the initial intent recognition results. This approach abandons the single-modal analysis mode that relies solely on text. By fusing user profiles and sentiment information, it overcomes the intent recognition bias caused by ignoring user history and sentiment tendencies, and significantly improves recognition accuracy.

[0011] Preferably, step S32 includes the following steps: S321. The semantic vector of the query statement is transformed linearly to generate an attention query matrix, and the user context feature vector is transformed linearly to generate an attention key matrix and an attention value matrix respectively. S322. Calculate the dot product of the attention query matrix and the attention key matrix to obtain the initial attention score matrix; S323. Scale the initial attention score matrix by dividing it by the square root of the dimension of the attention key matrix, and normalize it using the Softmax function to obtain the attention weight distribution matrix. S324. The attention weight distribution matrix and the attention value matrix are weighted and summed to obtain the preliminary weighted fusion feature; S325. The preliminary weighted fusion features are processed sequentially through residual connections and feedforward neural network layers, and then normalized to obtain the final weighted fusion features. The weight distribution is calculated by scaling the dot product attention of the query and the key matrix. After weighted summation, the weighted fusion features are generated by passing the residual connection, feedforward network and layer normalization in sequence. This achieves accurate cross-modal semantic alignment between the query and user features, effectively suppresses gradient vanishing, enhances the stability and nonlinear fitting ability of feature expression, and ensures high discriminativeness of the fusion features.

[0012] Preferably, step S4 includes the following steps: S41. Based on the intent recognition result, determine whether the intent category of the current consultation text falls into the preset set of complex open-ended question intents; if not, proceed to S42; if yes, proceed to S43. S42. Based on the intent recognition result, match the response rule template in the preset rule base, extract the key slot information in the standardized current query statement and fill it into the response rule template to generate the first response result; S43. Using the standardized current query statement and the intent representation vector of the fused context features as retrieval conditions, perform a dual-path parallel retrieval of keyword retrieval and vector similarity retrieval on the preset knowledge base. Use the inverse ranking fusion algorithm to fuse the two retrieval results and sort them according to the comprehensive relevance score. Remove documents with a comprehensive score lower than the preset threshold to obtain a candidate knowledge document set. S44. The selected candidate knowledge document, the standardized current query statement, and the intent representation vector fused with contextual features are concatenated according to a preset prompt word template to construct a prompt. The generated prompt is then input into the large language generation model to generate the second response result. The question type is determined by the intent recognition result. For non-complex questions, the first response is generated by matching the rule base. For complex questions, keyword and vector similarity are used for parallel retrieval. After weighted fusion and sorting, candidate documents are selected and concatenated with the query statement and intent vector before being input into the large language generation model to generate the second response. A hybrid decision mechanism combining rule matching and retrieval-enhanced generation is constructed, which overcomes the dual limitations of pure rule system in dealing with open questions and pure generation model in producing illusions.

[0013] Preferably, step S41 includes the following steps: S411. Obtain the final intent category from the intent recognition result. If the final intent category matches the complex open-ended question intent set, then determine that the current consultation text is a complex open-ended question and proceed to S43. S412. If the final intent category does not match the set of complex open-ended questions, then calculate the statement complexity index of the standardized current query statement. The statement complexity index includes one or more combinations of sentence length, number of clauses, number of interrogative words, and density of technical terms. S413. If the statement complexity index does not exceed the preset complexity threshold, then the current consultation text is determined to be a non-complex open-ended question, and proceed to S42. S414. If the statement complexity index exceeds the preset complexity threshold, the standardized current query statement and the intent representation vector fused with context features are input to the big language generation model. The big language generation model performs binary classification on the current query text based on preset complexity openness judgment prompts and outputs a complex or non-complex judgment result. S415. If the output of the large language generation model is complex, then the current consultation text is determined to be a complex open-ended question, and proceed to S43; if the output is non-complex, then the current consultation text is determined to be a non-complex open-ended question, and proceed to S42. Through a three-level progressive discrimination logic of intent set hit judgment, statement complexity threshold filtering, and large language model semantic judgment, the final judgment result of complex or non-complex is output. It avoids the risk of misjudgment of a single judgment method, realizes fine binary classification of threshold edge samples, ensures the accuracy and reliability of complex problem judgment, and provides a precise basis for the reasonable diversion of response strategies.

[0014] Preferably, step S5 includes the following steps: S51. Output the first response result or the second response result as the final response result to the user; S52. Obtain the user's feedback behavior data in response to the final response result. The feedback behavior data includes explicit feedback data and implicit feedback data. The explicit feedback data includes user ratings, user evaluation text, and / or useful / useless click markers. The implicit feedback data includes dwell time, subsequent operation behavior, and / or whether the question was asked repeatedly. S53. Generate initial feedback labels based on the feedback behavior data, and extract verifiable signals to perform consistency verification on the initial feedback labels, generating feedback labels with confidence weights. S54. The feedback samples whose confidence weight reaches a preset threshold, along with their corresponding standardized current query statement, the intent representation vector fused with contextual features, and the final response result, are associated and stored in the user data warehouse to update the associated data corresponding to the user; low-confidence samples are not added to the warehouse to update the user profile. S55. Based on the source path of the final response result and the feedback label, an incremental fine-tuning strategy is adopted to perform incremental fine-tuning: if the final response result comes from the rule base matching path and the feedback label indicates an invalid response, then the large language generation model and the multi-task learning network are not fine-tuned, but instead the periodic manual review and update process of the rule base is triggered; if the final response result comes from the large language generation model and the confidence weight is greater than or equal to the preset threshold, then the reason for invalidity is further diagnosed: if the semantic relevance between the retrieved candidate knowledge document and the current query statement is higher than the preset threshold, then the problem is determined to be in the generation module, and only the incremental fine-tuning of the large language generation model is triggered; if the retrieval relevance is lower than the preset threshold and the intent classification confidence is insufficient, then the root cause of the problem is determined to be in the intent understanding stage, and the incremental fine-tuning of the multi-task learning network is triggered; all incremental fine-tuning adopts the elastic weight consolidation method; By acquiring explicit and implicit user feedback behavior data, feedback labels with confidence weights are generated after consistency verification. Only high-confidence samples are added to the database to update user profiles. Incremental fine-tuning is triggered when the response is invalid and the confidence level is met. During fine-tuning, an elastic weight consolidation method is used to constrain parameter updates. A closed-loop learning mechanism is constructed to ensure sample quality through confidence screening and to prevent catastrophic forgetting through parameter constraints, thereby achieving continuous model evolution and personalized adaptation.

[0015] Preferably, an intelligent customer service intent understanding and response system based on big data analysis is used to implement the intelligent customer service intent understanding and response method based on big data analysis; the system includes a data acquisition and preprocessing module, a user profile management module, an intent understanding and recognition module, an intelligent response decision module, and a feedback learning and model update module; Data acquisition and preprocessing module: Acquires consultation text and user identifiers, and generates standardized query statements through cleaning, word segmentation, annotation, entity recognition and sentence completion; User profile management module: Retrieves static and dynamic behavioral data based on user identifiers, verifies and structures the data, and generates time-aligned associated data packets; The intent understanding and recognition module encodes the query statement and associated data packets separately, then fuses them through cross-modal attention, inputs them into a multi-task learning network to generate an intent representation vector, and outputs the intent recognition result after classification and sentiment branches. Intelligent response decision module: Based on the intent recognition results, it determines the question type. For non-complex questions, it generates the first response by matching the rule base. For complex questions, it retrieves the knowledge base and generates the second response by the big language generation model, and then outputs it. Feedback learning and model update module: Obtain user feedback, generate tags, associate and store them in the user data warehouse, and perform incremental fine-tuning when trigger conditions are met.

[0016] (III) Beneficial Effects The present invention has the following beneficial effects: This invention achieves accurate intent recognition based on user profile joint encoding. By introducing a multi-task learning network and a cross-modal attention fusion mechanism, during intent understanding, the system can perform deep semantic alignment between the user's current consultation text and historical static attributes and dynamic behavioral data, and assign differentiated attention weights based on the importance of different modal features. Once the user inputs an incomplete sentence or implicit request, the mechanism can automatically complete the semantic missing parts and correct intent classification bias based on the user's context features. This mechanism eliminates the technical defects of insufficient perception of the user's personalized context under traditional independent text analysis, ensuring the accuracy of intent recognition in complex scenarios such as short texts and colloquial expressions. This invention achieves adaptive deep response for complex open-ended questions. It constructs a hierarchical decision-making mechanism that combines rule-based rapid response with deep reasoning from a large model, and introduces a knowledge retrieval-enhanced generative architecture. During response processing, the system first accurately classifies the question type based on intent recognition results and complexity indicators. Once a complex open-ended question is identified, the system immediately switches to a higher-order processing link, using a standardized current query statement and an intent representation vector incorporating contextual features as retrieval conditions. It performs parallel keyword and vector similarity searches from the knowledge base, filters highly relevant candidate knowledge documents, and then collaborates with a large language generation model to generate a comprehensive, logically rigorous, and highly personalized customized response. This mechanism eliminates the technical shortcomings of traditional single-rule systems, such as low coverage of complex questions, rigid responses, and even inability to respond. It also overcomes the risk of factual illusions arising from pure large-model generation when lacking external knowledge support. This invention achieves continuous model evolution based on feedback. By introducing a feedback label verification mechanism with confidence weights and an incremental fine-tuning strategy with elastic weight consolidation, the platform can automatically capture users' explicit evaluations and implicit behavioral signals during system operation and maintenance. It then performs consistency screening on feedback samples based on verifiable signals, including only high-confidence samples in the training set to avoid noise contamination. Once sufficient effective feedback samples are accumulated, this mechanism can perform lightweight parameter updates on large language generation models and multi-task learning networks without forgetting existing mature knowledge. This mechanism eliminates the technical shortcomings of traditional static models, such as knowledge solidification after deployment and difficulty in adapting to dynamic business evolution. Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the intelligent customer service intent understanding and response method based on big data analysis according to the present invention. Figure 2 This is an architecture diagram of the intelligent customer service intent understanding and response system based on big data analysis according to the present invention; Figure 3 This is a flowchart illustrating the intent recognition process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0020] Please see Figure 1 This invention discloses an intelligent customer service intent understanding and response method and system based on big data analysis, including the following steps: S1. Obtain the current query text input by the user and the user's user identification information, preprocess the current query text to obtain a standardized current query statement; Step S1 includes the following steps: S11. Obtain the current consultation text input by the user, wherein the current consultation text includes speech-converted text, keyboard input text and / or touch screen handwriting recognition text; S12. Obtain the user identification information of the user, the user identification information including user account, device unique identifier and / or biometric identifier; S13. Perform text cleaning on the current consultation text. The text cleaning includes removing invalid characters, correcting typos, unifying simplified and traditional Chinese characters and normalizing punctuation marks to obtain the cleaned text. S14. Perform word segmentation on the cleaned text and correct the word segmentation results using a preset industry-customized dictionary to obtain a word segmentation sequence. S14 specifically involves: the preset industry-defined dictionary systematically collects historical customer service dialogue records, product manuals, and industry standard documents as raw corpora. After segmenting and denoising the raw corpora, the TextRank algorithm is used to extract candidate professional terms from them using high frequency. The sliding window size of the TextRank algorithm is set to 6 words, and the number of iterations is set to 200. These parameters were determined through grid search optimization on 100,000 historical customer service dialogue samples, with maximizing the F1 score of the extracted terms as the optimization objective. After extracting candidate words... The comprehensive weight value of each candidate word is calculated. This comprehensive weight value is determined by the weighted sum of TextRank score, term frequency-inverse document frequency score, and part-of-speech tagging score. The weight coefficients for TextRank score, term frequency-inverse document frequency score, and part-of-speech tagging score are 0.5, 0.3, and 0.2, respectively. These weight coefficients are obtained by logistic regression fitting of 2000 annotated corpora from each of five different business domains, with each coefficient representing the median of the fitting results for each domain. The top 30% of candidate words with a comprehensive weight value greater than 0.8 are selected as initial screening terms. After initial screening and manual review, a basic lexicon was formed. The basic lexicon was then compared with the default dictionary of a general word segmentation tool to detect conflicts. For conflicting entries, the frequency of the entries in the basic lexicon was adjusted to be at least 10 times higher than the frequency of homographs in the default dictionary. This 10-fold increase was achieved by randomly selecting 5000 statements from historical customer service conversations containing ambiguous segmentation errors from the default dictionary. Tests were conducted at five scaling factor levels (5x, 8x, 10x, 12x, and 15x) to assess the priority hit rate of entries from the basic dictionary and their normal recall rate within the general vocabulary. The experimental results showed: At a 10x speedup, the priority hit rate reached 96.2%, while the normal recall rate of general vocabulary only decreased by 0.8%, and the F1 score reached a peak of 0.962. Forced part-of-speech tagging matching the business scenario was implemented to ensure that entries in the basic lexicon had the highest segmentation priority during word segmentation. A synonym mapping table and a stop word list were constructed to assist in correction. The synonym mapping table was formed by mapping entries in the basic lexicon to 300-dimensional pre-trained word vectors and calculating cosine similarity. Pairs of entries with a cosine similarity greater than 0.85 were clustered into synonym groups and manually reviewed and confirmed.The 85% similarity threshold was determined by analyzing the precision-recall curves of 1000 manually annotated synonym pairs from each of the three business domains, selecting the threshold corresponding to the intersection of precision and recall. The stop word list was constructed using a joint term frequency-document frequency filtering method, removing function words and general verbs that appeared in all documents with a document frequency greater than 80% and whose semantic contribution was below a preset contribution threshold. This semantic contribution threshold was determined through an ablation experiment on 1000 randomly sampled sentences, manually deleting each word to assess the degree of semantic integrity retention. Words with a semantic integrity retention of less than 20% were ultimately marked as stop words. This dictionary is not static; it requires incremental updates to the basic thesaurus, synonym mapping table, and stop word list according to a preset cycle, based on online user feedback data and newly added business documents, ensuring continuous dynamic maintenance and maintaining its effectiveness and coverage as business evolves. S15. Perform part-of-speech tagging and named entity recognition on the word segmentation sequence to extract key entity information from the current consultation text; S15 specifically involves: after performing part-of-speech tagging on the segmented sequence, using a sequence labeling model combining a 256-dimensional bidirectional long short-term memory network and a first-order linear chain conditional random field for named entity recognition, with a model learning rate of 2×10⁻⁶. -5 The training rounds consisted of 10 rounds, with a batch size of 32. These parameters were determined through Bayesian hyperparameter optimization on 5000 labeled corpora. Preset entity types included order number, product model, mobile phone number, shipping address, order time, and expected delivery time. Order numbers were double-checked using regular expressions, and product models were corrected using a prefix tree dictionary constructed from a historical model list. Continuous word combinations labeled as B-entity type and I-entity type were extracted as key entity information, and their type labels were recorded. S16. Based on the key entity information and the word segmentation sequence, perform sentence normalization and semantic completion, complete incomplete or omitted sentences into complete semantic units, and generate the standardized current query statement. S16 specifically involves: performing dependency parsing on the segmented sequence; using a graph model-based dependency parser trained on 15,000 labeled corpora to identify core verbs and determine missing components; if missing components exist, a template filling strategy is used to complete them; the complete subtree coverage rate is calculated as follows: for the dependency parsing tree generated from the segmented sequence, the total number N of all non-root nodes in the tree is counted, and the number M of nodes with valid dependency relationship labels with their parent nodes and whose parent nodes are content words is counted. The complete subtree coverage rate is then calculated. The criteria for determining whether safe completion is impossible are as follows: The default component associated with the core verb has more than two high-scoring candidates in the candidate entity set with a confidence score difference of less than 0.3, making it impossible to resolve the completion ambiguity; the default component to be added has a logical semantic conflict with the existing negation word, tense particle, or modal particle in the word segmentation sequence; the key entity information required for completion is not found in the preset industry-customized dictionary and the user's corresponding association data, making it impossible to confirm the entity's legality; when the complete subtree coverage C is lower than the preset... If a threshold is set, or if any of the above conditions for unsafe completion are met, then the word segmentation sequence is not completed and the original sentence is retained; the word segmentation sequence after completion or retention of the original sentence is adjusted in word order and punctuation is rearranged according to a preset sentence pattern template to generate the standardized current query statement; the preset threshold is determined based on ablation experiments on the 15,000 annotated corpus. The experiments show that the F1 score of intent recognition of the completed text fluctuates by no more than ±1.5% when the coverage threshold is in the range of 35% to 45%, so the median value of 40% is taken as the preferred threshold for engineering; S2. Based on the user identification information, retrieve the associated data corresponding to the user from the preset user data warehouse; Step S2 includes the following steps: S21. Based on the user identification information, determine the unique index of the user corresponding to the user identification information from the preset user data warehouse; if there is no user data in the data warehouse, directly jump to S24 and use the default associated data; S21 specifically refers to: the user data warehouse includes a static attribute domain and a dynamic behavior domain, which are associated and stored with the user's unique index as the primary key; when the user's unique index is not retrieved, the default associated data includes a global statistical mean vector, a temporary behavior placeholder sequence for the current session, and a preset neutral sentiment value. After the current session ends, the default associated data triggers the generation of a cold start profile based on the collected feedback behavior data, so as to construct a new user's unique index and write it into the user data warehouse; S22. Based on the user's unique index, retrieve the user's static attribute data and dynamic behavior data from the user data warehouse; Specifically, S22 refers to: the static attribute data refers to stable features that are retained during user registration or obtained through long-term statistics and whose update frequency is less than monthly; the dynamic behavior data is an interaction sequence recorded in timestamp order; the dynamic data is added to the dynamic behavior domain of the user data warehouse in real time after each interaction. S23. Perform data quality verification on the static attribute data and the dynamic behavior data, remove outlier and null value fields, and fill missing fields with default values ​​or mean interpolation to obtain the verified associated data. S24. The verified associated data or default associated data is organized in a structured manner according to a preset feature field mapping table to generate an associated data packet that is time-aligned with the standardized current query statement, thereby obtaining the associated data corresponding to the user. Step S24 includes the following steps: S241. Obtain a preset feature field mapping table, wherein the feature field mapping table is used to define the mapping relationship between each original field in the verified associated data and the preset standardized feature dimension. S241 specifically refers to: the preset feature field mapping table is a pre-configured lookup table that records the one-to-one correspondence between each original field in the user data warehouse and the standardized feature dimensions inside the system, and specifies the data type and default value filling rules for each standardized dimension; by consulting the mapping table, the system uniformly converts the original data from different data sources with different field names into standardized feature expressions. S242. Based on the feature field mapping table, convert each original field in the verified associated data into a corresponding standardized feature value to form a structured feature vector set; S243. Extract the receiving timestamp of the standardized current query statement, and determine the backtracking time window based on the receiving timestamp; S243 specifically refers to the following: The retrospective time window ΔT is set to [T0-ΔT, T0) based on the current consultation time T0. Its typical value range varies depending on the business scenario: For e-commerce transactions, the default is 30 days, covering the user's orders, browsing, and customer service interaction records for the most recent month; for financial services, the default is 90 days, covering transaction flow and account change records for the past quarter; for technical support / product after-sales scenarios, the default is 7 days, focusing on recent visits and fault reporting records; for high-frequency interaction scenarios, the default is 14 days. The configuration of the above values ​​is based on the average resolution cycle of user issues and the decay cycle of historical behavior data in each business scenario. In the e-commerce scenario... The contribution of historical behavior exceeding 30 days to current intent judgment drops to below 5%. In financial scenarios, due to the long product holding period, coverage of 90 days is required. In technical support scenarios, historical fault records exceeding 7 days have significantly reduced reference value for current issues. The system also supports adaptive adjustment of user activity: high-frequency users with an average of ≥3 visits per day have their activity period shortened to 7 days, while low-frequency users with an average of <5 visits per month have their activity period extended to 60 days. This ensures that active users focus on recent behavior, while low-frequency users accumulate sufficient features, avoiding the unusable feature vectors due to data sparsity. When user activity and business scenario backtracking window rules are inconsistent, the rule with the higher number of days is prioritized to avoid missing important information. Before system deployment, offline experiments were conducted to verify the impact of different window lengths on intent recognition accuracy. The experiments employed A / B testing, running the intent recognition model on the same test set with five candidate window lengths: 7 days, 14 days, 30 days, 60 days, and 90 days, using the F1 score as the evaluation metric. Experimental results in the e-commerce scenario showed an F1 score of 0.812 for the 7-day window, 0.873 for the 30-day window, 0.858 for the 60-day window, and 0.841 for the 90-day window. Based on these results, 30 days was determined to be the default window length for e-commerce scenarios, representing an improvement of approximately 7.5% in F1 score compared to the 7-day window and approximately 3.8% compared to the 90-day window. Similar experiments in the financial scenario indicated that the 90-day window was optimal, while the 7-day window was optimal for technical support scenarios. After system deployment, the current window length was periodically evaluated and monitored to ensure it remained optimal. If significant changes occurred in the business model, the above verification process was re-executed, and the default configuration was automatically adjusted. S244. From the structured feature vector set, select a subset of feature vectors whose timestamps fall within the backtracking time window, as valid feature data aligned with the time of the standardized current query statement; S245. Assemble the effective feature data into a fixed-length feature matrix according to a preset feature dimension order, and normalize the fixed-length feature matrix to generate the associated data packet. S245 specifically refers to: the preset feature dimension order refers to a standardized feature arrangement sequence that the system has fixed in advance, which stipulates that the associated data of all users must be organized into feature vectors in the same order; based on this order, the system fills the selected valid feature data into the corresponding positions one by one, and fills the missing parts with default values, and finally generates a numerical matrix with fixed dimensions and uniform length, which is the structured associated data package for downstream models to use; S246. The associated data packet is timestamped and linked to the standardized current query statement to generate an associated data packet with a time tag, thereby obtaining the associated data corresponding to the user. S3. Using a multi-task learning network, the standardized current query statement and the associated data corresponding to the user are jointly encoded to generate an intent representation vector that integrates contextual features. The intent representation vector is then used to perform intent recognition on the current query text to obtain the intent recognition result. Please see Figure 3 Step S3 includes the following steps: S31. Input the standardized current query statement into the first encoder to extract semantic features and generate a query statement semantic vector; at the same time, input the associated data packet into the second encoder to extract user static features and dynamic behavior temporal features and generate a user context feature vector. Specifically, S31 is as follows: The first encoder uses a BERT-base-Chinese pre-trained model with 12 Transformer layers, 768-dimensional output, approximately 110 million parameters, and an output dimension of 768 dimensions; domain-adaptive fine-tuning is used for the customer service domain, with the learning rate set to 2×10 during the fine-tuning stage. -5 The AdamW optimizer was used, with a batch size of 32 and 3 training rounds. The second encoder consists of a two-layer fully connected network, with 256 neurons in the first layer and 128 neurons in the second layer, using ReLU activation function, and two layers of bidirectional GRU. The hidden layer has a dimension of 128, and dropout = 0.2. Static attributes are mapped to 128-dimensional vectors through the fully connected layers, and the dynamic behavior sequence is encoded by the bidirectional GRU and the hidden state at the last moment is taken as 256-dimensional. The two are concatenated and linearly projected to output a 512-dimensional user context feature vector. S32. Align the semantic vector of the query statement and the user context feature vector in dimensions and map their feature spaces. Use the aligned semantic vector of the query statement as the query item and the aligned user context feature vector as the key item. Calculate the semantic association weight using a cross-modal attention fusion mechanism and weight the aligned user context feature vector to obtain a weighted fusion feature. Step S32 includes the following steps: S321. The semantic vector of the query statement is transformed linearly to generate an attention query matrix, and the user context feature vector is transformed linearly to generate an attention key matrix and an attention value matrix respectively. S322. Calculate the dot product of the attention query matrix and the attention key matrix to obtain the initial attention score matrix; S323. Scale the initial attention score matrix by dividing it by the square root of the dimension of the attention key matrix, and normalize it using the Softmax function to obtain the attention weight distribution matrix. Specifically, S323 is: The mathematical expression of the Softmax function is as follows: For the input vector z=[z1,z2,...,z...] n The Softmax function maps it to a probability distribution vector σ(z) = [σ(z)1, σ(z)2, ..., σ(z)]. n ], where the i-th element is defined as: ; This function converts any real-valued vector into a probability distribution in which each element is in the interval (0,1) and the sum is 1. S324. The attention weight distribution matrix and the attention value matrix are weighted and summed to obtain the preliminary weighted fusion feature; S325. The preliminary weighted fusion features are processed sequentially through residual connections and feedforward neural network layers, and then normalized to obtain the final weighted fusion features. S33. The semantic vector of the query statement, the user context feature vector, and the weighted fusion feature are concatenated and input into the shared hidden layer of the multi-task learning network. The intent representation vector of the fusion context feature is generated through nonlinear transformation. Specifically, S33 refers to the following: The multi-task learning network is a neural network architecture with shared representations. Its core idea is to enable the same model to learn multiple related tasks simultaneously, thereby improving the generalization performance of each task through information sharing and mutual constraints between tasks. In this scheme, the network adopts a typical structure of shared hidden layers + task-specific output layers: the bottom shared hidden layer is universal for all tasks and is used to extract high-level shared semantic representations from the concatenated input vectors; above the shared hidden layer, intent classification task branches and sentiment analysis task branches are respectively attached. Each branch has an independent fully connected output layer and a corresponding loss function, and each outputs the intent category probability distribution and sentiment polarity category. Integrating intent recognition and sentiment analysis into a single multi-task learning network for joint training, rather than training two separate models, offers the following advantages: Firstly, the two tasks are inherently related at the semantic level; that is, the sentiment polarity of user queries often plays a supporting role in intent judgment. The shared hidden layer can capture this common representation across tasks, enabling the model to learn specific intents more fully. Secondly, during training, the multi-task network simultaneously receives gradient signals from both loss functions through backpropagation, which is equivalent to introducing additional regularization constraints to the shared layer. This helps to suppress model overfitting and improve generalization ability in sparse data scenarios. In this scheme, the output of the shared hidden layer is an intent representation vector that integrates contextual features. This vector carries both the information required for intent classification and the auxiliary knowledge of sentiment analysis. The specific structure of the shared hidden layer of the multi-task learning network is as follows: The shared hidden layer consists of three fully connected network layers stacked together. The first layer has an input dimension of 1792 dimensions, which is the sum of the query statement semantic vector dimension, the user context feature vector dimension, and the weighted fusion feature dimension. The output dimension is 1024 dimensions, the activation function is ReLU, and it is connected to a Dropout layer with a dropout rate of 0.3. The second layer has an input dimension of 1024 dimensions and an output dimension of 512 dimensions. The activation function is ReLU, and it is connected to a Dropout layer with a dropout rate of 0.3. The third layer has an input dimension of 512 dimensions and an output dimension of 256 dimensions. The activation function is ReLU. After the three fully connected network layers, a LayerNormalization layer is added for normalization, and the output is a 256-dimensional intent representation vector. The aforementioned Dropout rate of 0.3 was determined by performing a grid search on 50,000 labeled customer service dialogue samples. The search range was {0.1, 0.2, 0.3, 0.4, 0.5}, with the goal of maximizing the intent classification F1 score on the validation set. The experimental result showed that 0.3 was the optimal value. Furthermore, on top of the shared hidden layer, branches for intent classification and sentiment analysis are respectively attached; the intent classification branch uses the cross-entropy loss function. ,in For the true label of the i-th intent category, This represents the predicted probability for that category; the sentiment analysis task branch also uses the cross-entropy loss function. ,in For the true label of the j-th sentiment polarity category, The predicted probability for this category; the joint loss is obtained by weighted summation of the loss functions of the two branches. , These are the weighting coefficients. The closer the value is to 1, the more the model focuses on optimizing the intent classification task during training. The specific value can be determined according to the actual business needs. If you want to ensure balance, you can take 0.5. Perform end-to-end joint optimization training with this joint loss as the target, so that the intent representation vector learned by the shared hidden layer can serve two tasks at the same time. S34. Input the intent representation vector into the intent classification task branch and the sentiment analysis task branch of the multi-task learning network respectively to obtain the intent category probability distribution and sentiment polarity category. Take the intent category with the highest probability according to the intent category probability distribution as the preliminary intent recognition result, and correct the preliminary intent recognition result according to the sentiment polarity category to output the final intent recognition result. Specifically, S34 involves using the sentiment polarity category as a priori bias weight to weight and adjust the probability distribution of the intent category; let the probability distribution of the intent classification branch output be P=[p1,p2,...,p...]. nThe sentiment polarity output by the sentiment analysis branch is s∈{positive, neutral, negative}. The system assigns a corresponding adjustment coefficient to each intention category i according to the sentiment-intention correction coefficient table. Corrected intention probability distribution According to the formula Calculate, take The intent category corresponding to the highest probability is used as the final recognition result. Correction is triggered only when max(P) is below a preset threshold, initially set to 0.75. The above correction coefficients are entirely based on historical feedback data statistics: the system extracts historical interaction samples from the user data warehouse, groups them by sentiment polarity s and intent category i, and statistically analyzes the frequency percentage of responses confirmed as valid by users in each group. A Laplace smoothing strategy is then employed, according to the formula... Calculate the correction factor, where This represents the number of valid samples in this group. K represents the total number of samples in this group, and K represents the total number of intent categories. The baseline effective frequency of intent category i under neutral sentiment is used to ensure that the correction coefficient does not have zero or extreme values ​​when the sample is sparse. Its typical value range is 0.3~0.9, and it can be adjusted according to different business needs after the system goes live. Furthermore, the preset threshold is a configuration value that combines empirical and data-driven approaches. Its initial value is determined through offline evaluation before the system goes live. Specifically, a batch of historical consultation samples covering all intent categories and labeled with correct answers by business experts is selected. The intent classification task branch is run to output the intent category probability distribution for each sample. For each prediction result, its maximum prediction probability value and whether the corresponding prediction is correct are recorded, and a precision-confidence curve is plotted accordingly. Observing the curve, it can be seen that when the maximum prediction probability is below 0.6, the prediction accuracy is close to the level of random guessing, while when the probability exceeds 0.7, the prediction accuracy significantly improves and tends to stabilize. Furthermore, a grid search method is used to traverse the range of 0.5 to 0. All candidate threshold values ​​between 9 and 9 were selected, with the overall intent recognition accuracy and correction trigger rate after triggering correction being the comprehensive optimization objectives. Setting the threshold too low would lead to a large number of high-confidence samples being corrected, increasing the probability of introducing emotional noise and potentially reducing the overall accuracy. Setting the threshold too high would lead to a large number of low-confidence samples not being corrected, thus losing the auxiliary role of emotional information in intent determination. Experimental results show that 0.75 is the balance point: the marginal gain of accuracy after correction begins to slow down significantly around this value, and the proportion of samples triggering correction is about 20%~30%. This can effectively correct a considerable proportion of low-confidence misjudged samples while ensuring that more than 70% of high-confidence samples are not interfered with by emotional noise. Therefore, this value was selected as the initial configuration. After the system goes live, this threshold also supports dynamic adaptive adjustment: The system extracts the latest samples from the user data warehouse weekly, calculates the proportion of samples that are actually effective after correction when triggered at the current 0.75 threshold, and the proportion of high-confidence samples that are actually predicted incorrectly when not triggered. These two metrics are then substituted into a preset performance loss function for optimization. If the false negative rate is significantly high at the current threshold, the threshold is appropriately increased to allow more low-confidence samples to enter the correction process. If the effective gain rate after correction remains low, the threshold is appropriately decreased to reduce unnecessary correction attempts. This mechanism ensures that the threshold remains optimal as the distribution of business data shifts. 0.75 is only a typical initial value; in actual production, it may converge to the 0.65~0.85 range. An optimization loss function is defined. ,in The false negative rate is calculated as follows: the denominator is the total number of samples with incorrect predictions, and the numerator is the number of samples with incorrect predictions that were not corrected. To correct the effective gain ratio, the denominator is the total number of samples that triggered the correction, and the numerator is the number of samples that were incorrectly predicted before the correction but correctly predicted after the correction. α and β are weighting coefficients, with initial values ​​of α=β=0.5. The system calculates the loss value at the current threshold weekly and iterates through multiple candidate values ​​within the fluctuation range above and below the threshold, selecting the value that best suits the current threshold. The minimum threshold is used as the new configuration, thereby achieving optimal dynamic updates of the threshold; S4. Based on the intent recognition result, determine whether the current consultation text is a complex open-ended question; If not, then generate the first response result based on the rule base matching; If so, the standardized current query statement and the intent representation vector of the fused context features are used as retrieval conditions to retrieve candidate knowledge documents from the preset knowledge base. The retrieved candidate knowledge documents, the standardized current query statement, and the intent representation vector of the fused context features are then input into the large language generation model to generate a second response result. Step S4 includes the following steps: S41. Based on the intent recognition result, determine whether the intent category of the current consultation text falls into the preset set of complex open-ended question intents; if not, proceed to S42; if yes, proceed to S43. Step S41 includes the following steps: S411. Obtain the final intent category from the intent recognition result. If the final intent category matches the complex open-ended question intent set, then determine that the current consultation text is a complex open-ended question and proceed to S43. Specifically, S411 is as follows: The complex open-ended question intent set is a whitelist of intent categories preset by the system, used to quickly determine whether the current inquiry belongs to a complex open-ended question that requires a generative response from a large language model; this set includes all intent categories defined by business experts as being unable to be answered by a fixed rule template and requiring a generative response by integrating multi-source information; when the final intent category in the intent recognition result falls into this set, the system directly determines that the current inquiry is a complex open-ended question, skipping the subsequent sentence complexity calculation and large language generation model-assisted determination process and proceeding to S43; The initial version of the complex open-ended question intent set was manually annotated by business experts based on historical consultation logs and customer service ticket data. The annotation principle was: if the response content of a certain intent could not be covered by a fixed template and needed to be generated openly in combination with the user's personalized context, it would be included in this set. After the system went live, this set supported dynamic incremental updates. When the response results generated under the rule base matching path met any of the following conditions, the system would push the intent category corresponding to the sample to the review queue, where operations personnel would assess whether it needed to be moved to the complex open-ended question intent set: The proportion of invalid response samples marked by users under this intent category to the total number of samples in this intent category exceeded 15% in the last 7 calendar days. This threshold was determined based on the analysis of the distribution of invalid feedback samples in the historical rule base response samples before the system went live, and was obtained by shifting the historical inefficiency mean of 5%-8% upward by about two standard deviations, representing a statistically significant anomaly; The same user repeatedly asked the same or similar questions a total of 3 times within 30 minutes after responding to the rule base response under this intent category. The above thresholds are based on customer service industry user satisfaction inflection point data. Repeated questions three times within 30 minutes indicate that previous responses did not address the core issue. For this intent category, the cumulative number of invalid responses in the rule base within the last 30 calendar days must reach 50, with an invalidity rate of at least 10%. This threshold is determined based on statistical significance testing; with a sample size of 50, a 10% inefficiency rate allows for rejection of the random fluctuation hypothesis at a 95% confidence level. Conversely, if a certain intent category within the set performs poorly in the large language generation model path and the actual response content tends towards a fixed pattern, then when the average similarity between the LLM-generated response and the historical rule base template output ROUGE-L of this intent category consistently exceeds 0.85 within the last 14 calendar days, this threshold is determined based on semantic equivalence judgment experiments. When ROUGE-L is above 0.85, the manual judgment semantic equivalence consistency rate reaches over 90%, and the model call cost exceeds the preset multiple of the rule base call cost, it is pushed to the review queue for operations personnel to assess whether to remove it from the set and switch to the rule base matching path. S412. If the final intent category does not match the set of complex open-ended questions, then calculate the statement complexity index of the standardized current query statement. The statement complexity index includes one or more combinations of sentence length, number of clauses, number of interrogative words, and density of technical terms. Specifically, S412 involves the following: The statement complexity index is comprehensively quantified from four dimensions: sentence length, number of subordinate clauses, number of interrogative words, and density of technical terms. The specific calculation method is as follows: First, dependency parsing is performed on the standardized current query statement to identify the subject-verb-object structure and count the total number of Chinese and English characters contained in the sentence as the sentence length L. Simultaneously, the number of independent clauses split based on subordinating conjunctions and punctuation marks is counted as the number of subordinate clauses C. Second, the frequency Q of interrogative words in the query statement is counted based on a preset interrogative word dictionary. The word segmentation results are then matched with a preset industry terminology database, and the ratio of the number of matched terms to the total number of segmented terms is calculated. Let D be the density of technical terms. Then, Min-Max normalization is performed on the four original indicators, linearly mapping them to the [0,1] interval. That is, the normalized value of each dimension = (original value - historical minimum value of that dimension) / (historical maximum value of that dimension - historical minimum value of that dimension). The historical minimum and maximum values ​​are dynamically maintained by the system based on the distribution extreme values ​​of all user queries over the past 30 days across each dimension. Finally, the entropy weight method is used to determine the weight coefficients of each dimension. Information entropy is calculated based on the degree of variation of each dimension indicator over the past 30 days; the greater the degree of variation, the smaller the information entropy and the larger the corresponding weight coefficient. Based on this, a normalized weight vector is generated that satisfies the following conditions: The normalized scores of each dimension are weighted and summed with their corresponding weights to obtain the final comprehensive complexity index. ; S413. If the statement complexity index does not exceed the preset complexity threshold, then the current consultation text is determined to be a non-complex open-ended question, and proceed to S42. Specifically, S413 involves the following: The initial value of the complexity threshold is determined before the system goes live through A / B testing and grid search. A batch of labeled historical consultation samples is selected, and different threshold values ​​are iterated through. With classification accuracy and average response latency as comprehensive optimization objectives, the threshold that optimizes the F1 score is selected as the initial configuration. After the system is running, this threshold can be dynamically adjusted automatically based on business feedback. Each week, a verification period's worth of data samples is extracted from the user data warehouse, and the proportion of invalid user feedback among the total number of samples diverted to the rule base is recorded as follows: The proportion of user feedback that is valid but whose response latency exceeds a preset latency threshold among the total number of samples diverted to the large language generation model is denoted as . ,Will and Substitute into the preset loss function Bayesian optimization iterative calculation enables Minimize the new threshold, where The denominator is the total number of samples that were distributed to the rule base this week. The denominator is the total number of samples diverted to the large language generation model this week. γ and δ are weighting coefficients that can be adjusted according to business preferences. If priority is given to ensuring response quality, then γ > δ. If balance is ensured, then γ = δ = 0.5. S414. If the statement complexity index exceeds the preset complexity threshold, the standardized current query statement and the intent representation vector fused with context features are input to the big language generation model. The big language generation model performs binary classification on the current query text based on preset complexity openness judgment prompts and outputs a complex or non-complex judgment result. Specifically, in embodiment S414: the complex openness determination prompt is as follows: Judgment Rules: Complex open-ended questions are those that cannot be answered directly by fixed rules or a single knowledge document, requiring a large model to synthesize information, reason logically, perform multi-step analysis, or compare multiple sources to provide an answer; simple questions are those that can be answered directly by rule matching, querying a single knowledge document, or using a fixed template, such as status queries, operation guides, and factual questions and answers; when the complexity of a question is on the fuzzy boundary, please tend to classify it as complex; the judgment is based solely on the semantic content of the current query statement, and is not influenced by emotional tone; Please follow these steps to reason: Analyze the user's core needs and whether there are multiple intentions or implicit conditions; determine whether answering this question requires consulting multiple knowledge sources and conducting comprehensive reasoning; based on the above analysis, provide a final binary classification decision; the output format must be strictly JSON, with no additional content. {"is_complex":true / false,"reason":"Briefly explain the reasoning"}; Please determine whether the following user inquiry is a complex open-ended question: User query statement: {standardized current query statement}; Intent representation vector of encoded user context features: {intent representation vector fused with context features}; Please output the judgment result in JSON format; S415. If the output of the large language generation model is complex, then the current consultation text is determined to be a complex open-ended question, and proceed to S43; if the output is non-complex, then the current consultation text is determined to be a non-complex open-ended question, and proceed to S42. S42. Based on the intent recognition result, match the response rule template in the preset rule base, extract the key slot information in the standardized current query statement and fill it into the response rule template to generate the first response result; S42 specifically refers to the following: The response rule template in the rule base is a structured response framework stored in the preset rule base, used for quick and standardized responses to non-complex open-ended questions; each response rule template is bound to one or more intent categories, and consists of three core parts: first, fixed script text, which is the standard response statement backbone preset for the intent, in which replaceable slot placeholders are embedded; second, slot extraction rules, which define the method of extracting the corresponding slot value from the current query statement; and third, filling logic, which specifies the fallback strategy when slot value extraction fails; when the query statement hits a certain intent, the system pulls the corresponding template from the rule base according to the intent category, extracts key information from the standardized query statement according to the slot extraction rules, fills the extracted specific values ​​one by one into the placeholders of the template, and renders and generates a complete first response result; S43. Using the standardized current query statement and the intent representation vector of the fused context features as retrieval conditions, perform a dual-path parallel retrieval of keyword retrieval and vector similarity retrieval on the preset knowledge base. Use the inverse ranking fusion algorithm to fuse the two retrieval results and sort them according to the comprehensive relevance score. Remove documents with a comprehensive score lower than the preset threshold to obtain a candidate knowledge document set. S43 specifically refers to: the initial value of the preset threshold is determined through offline evaluation before the system goes online; a batch of standard test query sets marked with standard answers are selected, and a dual-path retrieval is performed on each query, traversing different threshold values, and using precision and recall as evaluation indicators, the threshold that makes the F1 value optimal and can remove obviously irrelevant documents is selected as the initial configuration. After the system goes live, the threshold supports dynamic optimization. Every week, real query and user feedback data from the user data warehouse are extracted. The comprehensive score of the search documents corresponding to the user-confirmed valid response is used as the positive sample, and the score corresponding to the invalid response is used as the negative sample. The optimal threshold is recalculated by maximizing the classification accuracy and updated to the configuration center. When the system detects that the candidate set is empty after multiple consecutive queries, the threshold is automatically temporarily lowered by one step to ensure that at least candidate documents are output. At the same time, an alarm is triggered to prompt the knowledge base administrator to supplement the relevant content. Furthermore, the pre-defined knowledge base employs a multi-source heterogeneous data fusion construction strategy: First, raw corpora are collected from product manuals, operation manuals, frequently asked questions documents, historical customer service dialogue records, work order resolution logs, and business training materials. Then, preprocessing is performed, including text cleaning, encoding standardization, and sentence boundary recognition. The cleaned text is segmented using a hybrid strategy based on semantic integrity. Documents with clear chapter structures are recursively segmented according to title levels, while texts without clear structures are segmented using a sliding window with a window size of 512 characters and a step size of 256 characters. Semantic boundary detection optimization is performed using BERT embedding vector cosine similarity, resulting in independent knowledge fragments. Each knowledge fragment is encoded into a 768-dimensional dense vector using the same model as the first encoder. Simultaneously, a keyword set is extracted to construct an inverted index, forming a triplet structure of text content, metadata, and vector representation, serving both vector similarity retrieval and keyword retrieval indexing systems. Finally, the knowledge base quality is continuously maintained through manual sampling and automated indicator monitoring, and an incremental update interface is supported. When new business documents are added or policies change, new knowledge can be added and old versions can be marked via incremental import, ensuring the timeliness and completeness of the knowledge base. S44. The selected candidate knowledge document, the standardized current query statement, and the intent representation vector fused with contextual features are concatenated according to a preset prompt word template to construct a prompt. The generated prompt is then input into the large language generation model to generate the second response result. Specifically, S44 is as follows: the large language generation model is Qwen3.5-2B-Base; temperature is set to 0.2. For intelligent customer service scenarios, accurate, stable, and illusion-avoiding answers are required, so 0.2 is a reasonable engineering default value; top_p is retained at the default of 0.9, and max_tokens is a typical value of 1024. The preset prompt word template is as follows: As a professional intelligent customer service response generation expert, the task is to generate an accurate, complete, friendly, and personalized customer service response based on given candidate knowledge documents, user queries, and user intent characteristics. Generation guidelines: All factual information must strictly originate from the candidate knowledge documents; fabrication or distortion of content is prohibited. The response should directly address the user's core needs, covering all key points in the user's question without omission. The tone, level of detail, and service strategy of the response should be appropriately adjusted based on the user profile characteristics implicit in the intent representation vector. For responses containing multiple steps or key points, use segmentation, numbering, or bullet points to organize them for easy user comprehension. The response should be friendly and professional, using polite and respectful language while maintaining professionalism and accuracy, avoiding excessive colloquialism. If the candidate knowledge documents lack the complete information needed to answer the user's question, the user should be informed truthfully, and further human assistance or alternative query methods should be suggested; fabricated answers are prohibited. Output requirements: Output only the response text content itself, without any additional explanations, tags, or JSON structure; the response text should be directly displayable to the end user; Please generate a customer service response based on the following information: User query statement: {standardized current query statement}; User intent features: {intent representation vector fused with context features}; Candidate knowledge document: {candidate knowledge document}; S5. Output the first response result or the second response result as the final response result to the user, obtain the user's feedback behavior data for the final response result, and associate and store the feedback behavior data, the standardized current query statement, the intent representation vector fused with context features and the final response result in the user data warehouse to update the associated data corresponding to the user and trigger incremental fine-tuning of the large language generation model and the multi-task learning network. Step S5 includes the following steps: S51. Output the first response result or the second response result as the final response result to the user; S52. Obtain the user's feedback behavior data in response to the final response result. The feedback behavior data includes explicit feedback data and implicit feedback data. The explicit feedback data includes user ratings, user evaluation text, and / or useful / useless click markers. The implicit feedback data includes dwell time, subsequent operation behavior, and / or whether the question was asked repeatedly. S53. Generate initial feedback labels based on the feedback behavior data, and extract verifiable signals to perform consistency verification on the initial feedback labels, generating feedback labels with confidence weights. Specifically, S53 is as follows: The method for generating the feedback label with confidence weight is as follows: Explicit feedback data and implicit feedback data are treated as two independent sources of evidence, and a weighted evidence fusion method is used to calculate the comprehensive confidence weight. Explicit feedback confidence calculation: Calculate the explicit confidence score C based on the user's proactive feedback behavior. exp If the user clicks the "helpful / useless" marker, the signal confidence level is set to 0.9; if the user provides a rating of 1-5 points, the rating is mapped to the [0,1] interval: C exp =(Rating - 1) / 4; If the user submits a review text, perform sentiment analysis on the review text. If the sentiment polarity is positive, then C exp =0.8, negative is C exp =0.2, neutral then C exp =0.5; when multiple explicit feedbacks exist simultaneously, the maximum value is taken; Implicit feedback confidence calculation: Calculate implicit confidence C based on user behavior signals.imp Duration T stay Normalized to C dwell =min(T stay / 30,1.0), where 30 seconds is the ideal dwell time threshold; in subsequent actions, if the user clicks the link provided in the response or performs the recommended action after responding, then C action =0.8, C if the user closes the session action =0.9 indicates the problem has been resolved. If the user raises a new question within 30 seconds of the response, then C... action =0.1; If the user does not ask the same question again after answering, then C repeat =0.9, otherwise C repeat =0.1; the implicit confidence level is obtained by weighted averaging of the above signals: Weight , , The correlation values ​​were determined by the correlation analysis between implicit signals and manually labeled feedback tags in historical data, and were set to 0.25, 0.45, and 0.30 respectively. Verifiable signal consistency check: Extract the set of verifiable signals, including: whether the user initiates a new question within 5 minutes after the response (signal V1; if not, V1=1, otherwise V1=0); whether the user enters the business completion page or order details page after the response (signal V2; if so, V2=1, otherwise V2=0); whether the link provided in the response is clicked (signal V3; if clicked, V3=1, otherwise V3=0); calculate the comprehensive score S of the verifiable signals. v =(V1+V2+V3) / 3; Fusion confidence weights: Initial feedback label L init The decision is directly determined by explicit feedback; if the user clicks "helpful," then L... init = Valid; if clicking is ineffective, then L init =Invalid; if there is no explicit feedback, the implicit signal is used for comprehensive judgment; comprehensive confidence weight C final Calculated using the following fusion formula: ; Where α, β, and γ are fusion weight coefficients, obtained by logistic regression fitting using historical labeled data, reflecting the predictive contribution of the three types of evidence to the true value of the feedback label, with typical values ​​of α=0.4, β=0.3, and γ=0.3; if C exp If C = 0, then its weight α is redistributed to β and γ, i.e., β' = β + α / 2, γ' = γ + α / 2; final With initial feedback label L init The associations are used to create feedback labels with confidence weights. S54. The feedback samples whose confidence weight reaches a preset threshold, along with their corresponding standardized current query statement, the intent representation vector fused with contextual features, and the final response result, are associated and stored in the user data warehouse to update the associated data corresponding to the user; low-confidence samples are not added to the warehouse to update the user profile. S54 specifically refers to the following: The confidence weight preset threshold is a pre-configured threshold value used to filter high-quality, high-confidence feedback samples into the user profile update and incremental fine-tuning process, so as to avoid low-quality or unreliable feedback samples from polluting the user profile and model parameters; the confidence weight is a value in the range of 0 to 1, generated by the consistency verification process in S53, and comprehensively reflects the credibility of the current feedback sample. The more consistent the initial feedback label is with the verifiable signal and the more sufficient the source of the verifiable signal, the higher the confidence weight; conversely, the lower the confidence weight. The system compares this confidence weight with the preset threshold: samples that are higher than or equal to the threshold are judged as high-confidence valid samples, associated and stored in the user data warehouse and trigger subsequent incremental fine-tuning; samples that are lower than the threshold are not stored in the warehouse, but are only kept in the original log for manual sampling review, and do not participate in user profile updates and model parameter adjustments. The initial value of the preset threshold is determined through offline evaluation before going live. A batch of test samples with manually labeled true values ​​are selected, and the confidence weight of each sample is calculated. Then, different threshold values ​​are iterated. The performance improvement of downstream tasks after the user profile is updated is taken as the optimization target, and the threshold that makes the performance gain optimal is selected as the initial configuration. After the system is running, the threshold supports periodic dynamic adjustment: the threshold is fluctuated up and down by several steps every month for offline verification. If the business prediction accuracy of the user profile is significantly improved after the adjustment and the number of effective training samples is not significantly reduced, then it is gradually adjusted upwards. If the incremental fine-tuning effect of the model is significantly improved after the adjustment and no noise is introduced, then it is appropriately adjusted downwards. S55. Based on the source path of the final response result and the feedback label, an incremental fine-tuning strategy is adopted to perform incremental fine-tuning: if the final response result comes from the rule base matching path and the feedback label indicates an invalid response, then the large language generation model and the multi-task learning network are not fine-tuned, but instead the periodic manual review and update process of the rule base is triggered; if the final response result comes from the large language generation model and the confidence weight is greater than or equal to the preset threshold, then the reason for invalidity is further diagnosed: if the semantic relevance between the retrieved candidate knowledge document and the current query statement is higher than the preset threshold, then the problem is determined to be in the generation module, and only the incremental fine-tuning of the large language generation model is triggered; if the retrieval relevance is lower than the preset threshold and the intent classification confidence is insufficient, then the root cause of the problem is determined to be in the intent understanding stage, and the incremental fine-tuning of the multi-task learning network is triggered; all incremental fine-tuning adopts the elastic weight consolidation method; Specifically, S55 involves the following: The elastic weight consolidation method uses a diagonal Fisher information matrix for approximation calculation. Specifically, 500 historical interaction samples are randomly sampled from the user data warehouse. For any model to be fine-tuned in the large language generation model or multi-task learning network, the expected value of the squared gradient of its loss function with respect to the model parameters is calculated as the Fisher diagonal element. For the large language generation model, the loss function uses an autoregressive language modeling loss, and the formula for calculating the Fisher diagonal element of the i-th parameter is as follows: In the formula The number of samples is 500. Let cross-entropy be the loss function. It is the complete parameter vector of the model. Let i be the i-th trainable parameter of the model. and Given the input and output of the nth historical interaction sample, the gradient is obtained through backpropagation; for a multi-task learning network, its loss function is a weighted sum of the loss from the intent classification task and the sentiment analysis task. The expected squared gradients of each task branch with respect to the shared hidden layer parameters are calculated separately and then summed according to the task weights to form the Fisher diagonal elements of the layer parameters; the regularization term weight coefficient λ is preset to 1.0, which is located in the center of the optimal interval [0.5, 2.0] verified by experiments; the incremental fine-tuning learning rate is set to 5 × 10⁻⁶. -6 The initial pre-training learning rate is 5×10. -5 The training frequency is 1 / 10, following the general rule of thumb for fine-tuning downstream tasks; the training rounds are limited to 2 epochs and early stopping is used. Experiments have shown that after more than 2 rounds, the accuracy of the old task drops by more than 1.2%, while within 2 rounds, the drop can be controlled within 0.5%; at the same time, the online elastic weight consolidation mode is preferred, and the decay factor γ is set to 0.95, which is the median of the recommended range [0.9, 0.99], in order to balance the contribution of recent important parameters and long-term knowledge accumulation. To better demonstrate the collaborative mechanism of the various modules of this invention in actual operation, the following uses an intelligent customer service system of an e-commerce platform as an example to provide a complete user interaction case. User background: Li Ming (pseudonym), a registered member of the platform, has 3 completed orders in the past 30 days. Account history data shows that he has consulted about return-related matters multiple times. User entered the following inquiry: Can I return the headphones I bought yesterday? After obtaining the text and user account identifier, the system performs text cleaning, removing the modal particle "that", word segmentation, and industry-customized dictionary correction. Then, it performs named entity recognition to extract the key entity information "earphone" and "yesterday". Based on dependency parsing, it identifies the core verb "return" and determines the missing components, which are the object and specific conditions of the return. Based on historical dialogue experience, the system completes the elliptical sentence into a standardized query statement: "User applies for return of earphones, earphones were purchased yesterday". Based on user identification information, the system retrieves the user's associated data from the user data warehouse; static attribute data includes: user registration level and historical return frequency; dynamic behavior data includes: order browsing records in the past 30 days and historical interaction logs with this customer service system; the system uses the current consultation timestamp as a benchmark to trace back the behavior sequence of the past 30 days, extracting effective features such as 3 completed orders in the past 30 days, 1 of which had a return application, etc., and after feature field mapping and normalization, generates time-aligned associated data packages; The preprocessed standardized query statement generates a semantic vector through the BERT encoder; the associated data packet generates a user context feature vector through the second encoder; the two are fused through cross-modal attention and then input into a multi-task learning network; in the intent classification task branch, the model identifies the intent of a return / exchange application, and the sentiment analysis branch determines that the user's sentiment is neutral; since the intent recognition confidence is 0.68, which is lower than the preset threshold of 0.75, the system triggers a sentiment correction strategy: based on the historical statistical intent correction coefficient under neutral sentiment, the output probability of the return / exchange application category is weighted and adjusted, and after correction, the probability of this intent is increased to 0.82, and the final output intent recognition result is a return / exchange application; The system determines that the return / exchange request is a non-complex open-ended question and directly matches it to the return policy response template in the rule base. Simultaneously, it extracts key slot information from the standardized query: Product Category = Headphones, Purchase Time = Yesterday; these two slot values ​​are then filled into the rule template to generate the first response: Hello, the headphones you purchased meet the 7-day no-reason return policy. Please ensure the product and packaging are intact. Log in to your account, find the corresponding order in My Orders, and click to apply for after-sales service. The system outputs the above response results to the user; the user then clicks the "helpful" button on the interface, and exits the session after a short period of time without initiating a new question; the system captures this user feedback behavior data and generates positive feedback labels with high confidence weights; these high-confidence feedback samples will be associated and stored in the user data warehouse to update the user's profile features; at the same time, they will be accumulated as incremental training samples, and when the triggering conditions are met, the multi-task learning network and language generation model will be elastically weighted and incrementally fine-tuned to continuously evolve the model's ability to recognize and respond to scenarios such as return and exchange requests.

[0021] For further details, please refer to Figure 2 A big data analytics-based intelligent customer service intent understanding and response system is used to implement the big data analytics-based intelligent customer service intent understanding and response method; the system includes a data acquisition and preprocessing module, a user profile management module, an intent understanding and recognition module, an intelligent response decision module, and a feedback learning and model update module; Data acquisition and preprocessing module: Acquires consultation text and user identifiers, and generates standardized query statements through cleaning, word segmentation, annotation, entity recognition and sentence completion; User profile management module: Retrieves static and dynamic behavioral data based on user identifiers, verifies and structures the data, and generates time-aligned associated data packets; The intent understanding and recognition module encodes the query statement and associated data packets separately, then fuses them through cross-modal attention, inputs them into a multi-task learning network to generate an intent representation vector, and outputs the intent recognition result after classification and sentiment branches. Intelligent response decision module: Based on the intent recognition results, it determines the question type. For non-complex questions, it generates the first response by matching the rule base. For complex questions, it retrieves the knowledge base and generates the second response by the big language generation model, and then outputs it. Feedback learning and model update module: Obtain user feedback, generate tags, associate and store them in the user data warehouse, and perform incremental fine-tuning when trigger conditions are met; This invention addresses the challenges traditional question-answering systems face in handling complex open-ended questions, constructing a comprehensive solution encompassing accurate identification, hierarchical decision-making, and generative responses. At the question identification level, it abandons single-rule judgment and employs a three-tiered progressive mechanism—intent category matching, dual filtering based on multiple indicators of statement complexity, and binary classification verification using a large language model—to accurately define the boundaries of complex open-ended questions. At the response generation level, addressing the deep reasoning and information integration needs of complex open-ended questions, it utilizes dual-path parallel retrieval to broadly recall candidate documents from the knowledge base. The retrieval results, along with standardized query statements and intent representation vectors incorporating contextual features, are input into a large language generation model, driving it to generate a second response with logical coherence and knowledge integration capabilities. By deeply coupling accurate judgment of complex open-ended questions with generative responses, it achieves efficient collaboration from simple rule matching to complex semantic generation, significantly improving the response quality and generalization ability of intelligent customer service in open-ended and unstructured consultation scenarios. In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A smart customer service intent understanding and response method based on big data analysis, characterized in that, Includes the following steps: S1. Obtain the current query text input by the user and the user's user identification information, preprocess the current query text to obtain a standardized current query statement; S2. Based on the user identification information, retrieve the associated data corresponding to the user from the preset user data warehouse; S3. Using a multi-task learning network, the standardized current query statement and the associated data corresponding to the user are jointly encoded to generate an intent representation vector that integrates contextual features. The intent representation vector is then used to perform intent recognition on the current query text to obtain the intent recognition result. S4. Based on the intent recognition result, determine whether the current consultation text is a complex open-ended question; If not, then generate the first response result based on the rule base matching; If so, the standardized current query statement and the intent representation vector of the fused context features are used as retrieval conditions to retrieve candidate knowledge documents from the preset knowledge base. The retrieved candidate knowledge documents, the standardized current query statement, and the intent representation vector of the fused context features are then input into the large language generation model to generate a second response result. S5. Output the first response result or the second response result as the final response result to the user, obtain the user's feedback behavior data for the final response result, and associate and store the feedback behavior data, the standardized current query statement, the intent representation vector fused with context features and the final response result in the user data warehouse to update the associated data corresponding to the user and trigger incremental fine-tuning of the large language generation model and the multi-task learning network.

2. The intelligent customer service intent understanding and response method based on big data analysis according to claim 1, characterized in that, S1 includes the following steps: S11. Obtain the current consultation text input by the user, wherein the current consultation text includes speech-converted text, keyboard input text and / or touch screen handwriting recognition text; S12. Obtain the user identification information of the user, the user identification information including user account, device unique identifier and / or biometric identifier; S13. Perform text cleaning on the current consultation text. The text cleaning includes removing invalid characters, correcting typos, unifying simplified and traditional Chinese characters and normalizing punctuation marks to obtain the cleaned text. S14. Perform word segmentation on the cleaned text and correct the word segmentation results using a preset industry-customized dictionary to obtain a word segmentation sequence. S15. Perform part-of-speech tagging and named entity recognition on the word segmentation sequence to extract key entity information from the current consultation text; S16. Based on the key entity information and the word segmentation sequence, perform sentence normalization and semantic completion, complete incomplete or omitted sentences into complete semantic units, and generate the standardized current query statement.

3. The intelligent customer service intent understanding and response method based on big data analysis according to claim 2, characterized in that, S2 includes the following steps: S21. Based on the user identification information, determine the unique index of the user corresponding to the user identification information from the preset user data warehouse; S22. Based on the user's unique index, retrieve the user's static attribute data and dynamic behavior data from the user data warehouse; S23. Perform data quality verification on the static attribute data and the dynamic behavior data, remove outlier and null value fields, and fill missing fields with default values ​​or mean interpolation to obtain the verified associated data. S24. The verified associated data is organized in a structured manner according to a preset feature field mapping table to generate an associated data packet that is time-aligned with the standardized current query statement, thereby obtaining the associated data corresponding to the user.

4. The intelligent customer service intent understanding and response method based on big data analysis according to claim 3, characterized in that, S24 includes the following steps: S241. Obtain a preset feature field mapping table, wherein the feature field mapping table is used to define the mapping relationship between each original field in the verified associated data and the preset standardized feature dimension. S242. Based on the feature field mapping table, convert each original field in the verified associated data into a corresponding standardized feature value to form a structured feature vector set; S243. Extract the receiving timestamp of the standardized current query statement, and determine the backtracking time window based on the receiving timestamp; S244. From the structured feature vector set, select a subset of feature vectors whose timestamps fall within the backtracking time window, as valid feature data aligned with the time of the standardized current query statement; S245. Assemble the effective feature data into a fixed-length feature matrix according to a preset feature dimension order, and normalize the fixed-length feature matrix to generate the associated data packet. S246. The associated data packet is timestamped and linked to the standardized current query statement to generate an associated data packet with a time tag, thereby obtaining the associated data corresponding to the user.

5. The intelligent customer service intent understanding and response method based on big data analysis according to claim 4, characterized in that, S3 includes the following steps: S31. Input the standardized current query statement into the first encoder to extract semantic features and generate a query statement semantic vector; at the same time, input the associated data packet into the second encoder to extract user static features and dynamic behavior temporal features and generate a user context feature vector. S32. Align the semantic vector of the query statement and the user context feature vector in dimensions and map their feature spaces. Use the aligned semantic vector of the query statement as the query item and the aligned user context feature vector as the key item. Calculate the semantic association weight using a cross-modal attention fusion mechanism and weight the aligned user context feature vector to obtain a weighted fusion feature. S33. The semantic vector of the query statement, the user context feature vector, and the weighted fusion feature are concatenated and input into the shared hidden layer of the multi-task learning network. The intent representation vector of the fusion context feature is generated through nonlinear transformation. S34. Input the intent representation vector into the intent classification task branch and the sentiment analysis task branch of the multi-task learning network respectively to obtain the intent category probability distribution and sentiment polarity category. Take the intent category with the highest probability according to the intent category probability distribution as the preliminary intent recognition result, and correct the preliminary intent recognition result according to the sentiment polarity category to output the final intent recognition result.

6. The intelligent customer service intent understanding and response method based on big data analysis according to claim 5, characterized in that, S32 includes the following steps: S321. The semantic vector of the query statement is transformed linearly to generate an attention query matrix, and the user context feature vector is transformed linearly to generate an attention key matrix and an attention value matrix respectively. S322. Calculate the dot product of the attention query matrix and the attention key matrix to obtain the initial attention score matrix; S323. Scale the initial attention score matrix by dividing it by the square root of the dimension of the attention key matrix, and normalize it using the Softmax function to obtain the attention weight distribution matrix. S324. The attention weight distribution matrix and the attention value matrix are weighted and summed to obtain the preliminary weighted fusion feature; S325. The preliminary weighted fusion features are processed sequentially through residual connections and feedforward neural network layers, and then normalized to obtain the final weighted fusion features.

7. The intelligent customer service intent understanding and response method based on big data analysis according to claim 5, characterized in that, S4 includes the following steps: S41. Based on the intent recognition result, determine whether the intent category of the current consultation text falls into the preset set of complex open-ended question intents; if not, proceed to S42; if yes, proceed to S43. S42. Based on the intent recognition result, match the response rule template in the preset rule base, extract the key slot information in the standardized current query statement and fill it into the response rule template to generate the first response result; S43. Using the standardized current query statement and the intent representation vector of the fused context features as retrieval conditions, perform a dual-path parallel retrieval of keyword retrieval and vector similarity retrieval on the preset knowledge base. Use the inverse ranking fusion algorithm to fuse the two retrieval results and sort them according to the comprehensive relevance score. Remove documents with a comprehensive score lower than the preset threshold to obtain a candidate knowledge document set. S44. The selected candidate knowledge document, the standardized current query statement, and the intent representation vector fused with contextual features are concatenated according to a preset prompt word template to construct a prompt. The generated prompt is then input into the large language generation model to generate the second response result.

8. The intelligent customer service intent understanding and response method based on big data analysis according to claim 7, characterized in that, S41 includes the following steps: S411. Obtain the final intent category from the intent recognition result. If the final intent category matches the complex open-ended question intent set, then determine that the current consultation text is a complex open-ended question and proceed to S43. S412. If the final intent category does not match the set of complex open-ended questions, then calculate the statement complexity index of the standardized current query statement. The statement complexity index includes one or more combinations of sentence length, number of clauses, number of interrogative words, and density of technical terms. S413. If the statement complexity index does not exceed the preset complexity threshold, then the current consultation text is determined to be a non-complex open-ended question, and proceed to S42. S414. If the statement complexity index exceeds the preset complexity threshold, the standardized current query statement and the intent representation vector fused with context features are input to the big language generation model. The big language generation model performs binary classification on the current query text based on preset complexity openness judgment prompts and outputs a complex or non-complex judgment result. S415. If the output of the large language generation model is complex, then the current consultation text is determined to be a complex open-ended question, and proceed to S43; if the output is non-complex, then the current consultation text is determined to be a non-complex open-ended question, and proceed to S42.

9. The intelligent customer service intent understanding and response method based on big data analysis according to claim 7, characterized in that, S5 includes the following steps: S51. Output the first response result or the second response result as the final response result to the user; S52. Obtain the user's feedback behavior data in response to the final response result. The feedback behavior data includes explicit feedback data and implicit feedback data. The explicit feedback data includes user ratings, user evaluation text, and / or useful / useless click markers. The implicit feedback data includes dwell time, subsequent operation behavior, and / or whether the question was asked repeatedly. S53. Generate initial feedback labels based on the feedback behavior data, and extract verifiable signals to perform consistency verification on the initial feedback labels, generating feedback labels with confidence weights. S54. The feedback samples whose confidence weight reaches a preset threshold, along with their corresponding standardized current query statement, the intent representation vector fused with contextual features, and the final response result, are associated and stored in the user data warehouse to update the associated data corresponding to the user; low-confidence samples are not added to the warehouse to update the user profile. S55. Based on the source path of the final response result and the feedback label, an incremental fine-tuning strategy is adopted to perform incremental fine-tuning: if the final response result comes from the rule base matching path and the feedback label indicates an invalid response, then the large language generation model and the multi-task learning network are not fine-tuned, but instead the periodic manual review and update process of the rule base is triggered; if the final response result comes from the large language generation model and the confidence weight is greater than or equal to a preset threshold, then the reason for invalidity is further diagnosed: if the semantic relevance between the retrieved candidate knowledge document and the current query statement is higher than a preset threshold, then the problem is determined to be in the generation module, and only the incremental fine-tuning of the large language generation model is triggered; if the retrieval relevance is lower than a preset threshold and the intent classification confidence is insufficient, then the root cause of the problem is determined to be in the intent understanding stage, and the incremental fine-tuning of the multi-task learning network is triggered; all incremental fine-tuning adopts the elastic weight consolidation method.

10. An intelligent customer service intent understanding and response system based on big data analysis, characterized in that: The system is used to implement the intelligent customer service intent understanding and response method based on big data analysis as described in any one of claims 1-9; the system includes a data acquisition and preprocessing module, a user profile management module, an intent understanding and recognition module, an intelligent response decision module, and a feedback learning and model update module; Data acquisition and preprocessing module: Acquires consultation text and user identifiers, and generates standardized query statements through cleaning, word segmentation, annotation, entity recognition and sentence completion; User profile management module: Retrieves static and dynamic behavioral data based on user identifiers, verifies and structures the data, and generates time-aligned associated data packets; The intent understanding and recognition module encodes the query statement and associated data packets separately, then fuses them through cross-modal attention, inputs them into a multi-task learning network to generate an intent representation vector, and outputs the intent recognition result after classification and sentiment branches. Intelligent response decision module: Based on the intent recognition results, it determines the question type. For non-complex questions, it generates the first response by matching the rule base. For complex questions, it retrieves the knowledge base and generates the second response by the big language generation model, and then outputs it. Feedback learning and model update module: Obtain user feedback, generate tags, associate and store them in the user data warehouse, and perform incremental fine-tuning when trigger conditions are met.

Images