Intelligent professional knowledge question and answer customer service system based on self-optimization mechanism

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a professional knowledge question and answer intelligent customer service system based on a self-optimization mechanism. The system comprises a user intention analysis module, a knowledge processing module, a self-optimization learning module and an interactive presentation module. A semantic understanding unit of the user intention analysis module generates a user intention signal; after the knowledge processing module receives the signal, a knowledge retrieval unit outputs related knowledge fragments and confidence, an answer generation unit forms candidate answers, and a quality evaluation unit determines an optimal answer according to the confidence; in the self-optimization learning module, a feedback analysis unit adjusts answer generation parameters according to user interaction data, a strategy adjustment unit optimizes a retrieval strategy in combination with an optimal answer and a historical dialogue, and a knowledge updating unit updates a knowledge base depending on an external knowledge source; and a multi-round dialogue management unit of the interactive presentation module adjusts a dialogue process, and a visual presentation unit outputs a natural language text and collects user feedback to a feedback analysis unit.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a professional knowledge question and answer intelligent customer service system based on a self-optimization mechanism. BACKGROUND

[0002] With the rapid development of information technology, intelligent customer service systems are increasingly widely used in various industries, becoming an important bridge for communication between enterprises and users. Traditional intelligent customer service systems rely on pre-set rule bases or simple machine learning models, and often have many limitations when dealing with professional knowledge questions and answers. In the field of professional knowledge, users' questions often have strong professional and complex characteristics, involving specific industry terms, concepts and deep logic. The semantic understanding unit of existing systems is often limited to surface text matching, and it is difficult to accurately capture the user's true intention, especially when the user's expression is ambiguous, has ambiguity or uses industry-specific expressions, which can easily lead to understanding deviation, resulting in a loss of direction for subsequent knowledge retrieval and answer generation. In the knowledge processing link, the knowledge retrieval unit of traditional systems mostly uses keyword matching or simple vector similarity calculation, which is difficult to accurately locate the knowledge fragments highly related to the user's intention from the vast knowledge base. At the same time, there is a lack of effective confidence assessment for the retrieved knowledge, so that the answer generation unit cannot distinguish the reliability of the knowledge when integrating information, and it is easy to include low-confidence knowledge in the answer, affecting the accuracy of the answer. In addition, the quality assessment unit often lacks dynamic adjustment capability, and it is difficult to develop targeted evaluation standards according to different fields and different types of problems, resulting in a lack of scientificity in the selection of optimal answers. In terms of system iteration, the self-optimization capability of existing intelligent customer service systems is weak. Most systems rely on manual updates to the knowledge base and adjustments to model parameters, making it difficult to adapt to rapid updates in professional knowledge and dynamic changes in user needs. User interaction data and feedback information cannot be effectively utilized, and problems in semantic understanding, knowledge retrieval and other aspects cannot be discovered in a timely manner, resulting in slow improvement in system performance. At the same time, the optimization of retrieval strategies lacks the support of historical dialogue data, making it difficult to form adaptive retrieval logic and leading to repeated errors when facing similar problems. In terms of interactive presentation, the multi-round dialogue management capability of traditional systems is insufficient, and it is difficult to flexibly adjust the dialogue process according to the user's subsequent questions, often resulting in dialogue interruption or answering questions that are not asked when the user asks follow-up questions or changes the topic. The answer form output by the visual display unit is single, mostly pure text information, and lacks structured presentation of professional knowledge, which is not conducive to the user's quick understanding of complex content, and the collection mechanism for user feedback is not perfect, and valuable feedback cannot be converted into a driving force for system optimization in a timely manner.

[0003] The existence of these problems makes it difficult for existing intelligent customer service systems to provide accurate, reliable and flexible services in professional knowledge question and answer scenarios, and cannot meet the efficient acquisition demand of users for professional information, which also limits the further application and development of intelligent customer service in the professional field. SUMMARY

[0004] The present application aims to provide a professional knowledge question and answer intelligent customer service system based on a self-optimization mechanism to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides a professional knowledge question and answer intelligent customer service system based on a self-optimization mechanism, which comprises: a user intention analysis module, which comprises a semantic understanding unit that generates a user intention signal; a knowledge processing module that receives the user intention signal, the knowledge processing module comprising a knowledge retrieval unit, an answer generation unit and a quality assessment unit, the knowledge retrieval unit processing the user intention signal to output relevant knowledge fragments and knowledge confidence, the answer generation unit processing the relevant knowledge fragments and the user intention signal to obtain a candidate answer, and the quality assessment unit defining an answer quality standard according to the knowledge confidence and outputting an optimal answer; a self-optimization learning module, which comprises a feedback analysis unit, a strategy adjustment unit and a knowledge updating unit, the feedback analysis unit adjusting the parameters of the answer generation unit according to user interaction data, the strategy adjustment unit optimizing the retrieval strategy according to the optimal answer and historical dialogue data, and the knowledge updating unit dynamically updating the system knowledge base according to external knowledge sources; an interactive presentation module, which comprises a multi-round dialogue management unit and a visual display unit, the multi-round dialogue management unit adjusting the dialogue process according to the optimal answer and user follow-up questions, and the visual display unit converting the optimal answer into natural language text and outputting it, while collecting user feedback signals and sending them to the feedback analysis unit.

[0006] Preferably, the user intention analysis module comprises a distributed semantic analysis unit and a context awareness unit, the user intention analysis module being deployed on a customer service interaction platform, a mobile terminal and a web interface and collecting user input information, the context awareness unit obtaining user conversation context data through dialogue history tracking technology and forming a context signal, the semantic understanding unit receiving the user input information and the context signal to realize multi-modal semantic fusion and form the user intention signal, and the user intention signal comprising a retrieval intention signal, a generation intention signal and an evaluation intention signal.

[0007] Preferably, the knowledge retrieval unit comprises an entity linking model based on a knowledge graph and a reinforcement learning retrieval strategy model, the retrieval intention signal comprises user question keywords, entity type data, context correlation data, historical retrieval records and domain knowledge classification data, the user question keywords, the historical retrieval records, the entity type data, the context correlation data and the domain knowledge classification data are input into the entity linking model and output the relevant knowledge fragments, and the context correlation data, the domain knowledge classification data and the knowledge confidence are fused through a deep reinforcement learning strategy model and output the knowledge confidence.

[0008] Preferably, the entity linking model comprises an entity feature extraction submodule, a context fusion submodule, an association degree calculation submodule and a retrieval result sorting module, the entity feature extraction submodule performs word vector embedding on the user question keywords, extracts entity features meeting an entity recognition threshold condition as candidate entity signals, the context fusion submodule processes the entity type data, the context correlation data and the domain knowledge classification data using an attention mechanism weighted bidirectional Transformer network and outputs the data to the association degree calculation submodule, the association degree calculation submodule defines an association degree scoring rule, and the retrieval result sorting module generates a priority order of the relevant knowledge fragments based on a multi-layer sorting strategy and uses different sorting positions to represent different association degree scores of the knowledge fragments.

[0009] Preferably, the answer generation unit processes the generation intention signal and the relevant knowledge fragments through a generative question answering model, outputs a candidate answer, and processes the candidate answer using a particle swarm optimization algorithm to obtain an optimal answer; the generation intention signal comprises question type features and answer style parameters, the generative question answering model is constructed through the question type features, the relevant knowledge fragments and the answer style parameters, the generative question answering model realizes text generation based on an encoder-decoder architecture, and adjusts decoder attention weights to optimize answer fluency and knowledge accuracy.

[0010] Preferably, the answer generation unit further comprises a professional corpus dynamic library, a syntax logic verification submodule and an answer style adaptation submodule, the professional corpus dynamic library stores corpus features of historical question and answer pairs, the syntax logic verification submodule detects syntax errors and logical contradictions of a candidate answer according to the question type features and the relevant knowledge fragments, and the answer style adaptation submodule obtains the optimal answer through a particle swarm optimization algorithm according to the answer style parameters, the generative question answering model and the user intention signal.

[0011] Preferably, the quality evaluation unit defines an answer quality indicator according to the evaluation intention signal, outputs an answer quality score, and dynamically adjusts the answer quality score to be not lower than a quality pass threshold through fuzzy PID control, the evaluation intention signal includes user satisfaction feedback and professional accuracy data, the user satisfaction feedback and the professional accuracy data define the answer quality indicator, and the fuzzy PID control dynamically adjusts decoding parameters and knowledge retrieval thresholds of the generative question and answer model to make the answer quality score not lower than the quality pass threshold.

[0012] Preferably, the interaction presentation module includes a finite state machine dialogue management unit and a multi-template response generation module, the finite state machine dialogue management unit receives the optimal answer and the user feedback signal and dynamically adjusts dialogue state transition rules, and the multi-template response generation module processes the answer quality score and the user intention signal according to an improved fuzzy comprehensive evaluation algorithm and outputs a natural language answer, when the answer quality score is lower than a quality pass threshold, the interaction presentation module controls the knowledge processing module to regenerate candidate answers, the multi-template response generation module supports multi-style answer template switching, can display answer generation progress and knowledge source annotations in real time, and simulates answer tone and expression under different user portraits based on a rule engine.

[0013] Preferably, the feedback analysis unit includes state identification of a multi-level feedback processing mechanism, which is a general improvement state, an important optimization state and an emergency repair state, respectively, when the satisfaction score in the user feedback signal drops to a first score threshold, the general improvement state is started, and the parameter weight of the answer generation unit is adjusted to a regular optimization range, when the satisfaction score in the user feedback signal drops to a second score threshold, the current dialogue process is suspended and the knowledge processing module is controlled to re-retrieve knowledge, and when the satisfaction score in the user feedback signal drops to a third score threshold, the feedback analysis unit sends a system knowledge defect signal and controls the self-optimization learning module to urgently update the knowledge base.

[0014] Preferably, the optimal answer includes but is not limited to problem solution text, related knowledge links, recommended question lists and next step operation suggestions, the knowledge updating unit includes an intelligent crawler component and a knowledge graph updating device, the intelligent crawler component can identify the update frequency of external knowledge sources and grab the latest professional knowledge on demand, the knowledge graph updating device receives the knowledge confidence and adjusts the entity association strength of the knowledge graph in real time based on an entity relationship reasoning algorithm, and the user feedback signal includes but is not limited to a satisfaction score, an answer usefulness mark, a problem solving state and a supplementary information input.

[0015] Compared with the prior art, the application has the following beneficial effects: The professional knowledge question and answer intelligent customer service system based on the self-optimization mechanism has multiple advantages in the field of professional knowledge question and answer through the cooperative operation of multiple modules. The semantic understanding unit in the user intention analysis module can deeply analyze the professional connotation in the user's question and accurately generate the user intention signal. This accurate grasp of the user's intention can reduce subsequent processing errors caused by understanding deviations, and enable the system to accurately capture the user's core needs when facing professional term-intensive and complex expression questions, providing a clear direction for subsequent knowledge processing.

[0016] The knowledge processing module processes the user intention signal through the knowledge retrieval unit, outputs the relevant knowledge segments and knowledge confidence, so that the answer generation unit has a basis when integrating information. Combined with the knowledge confidence, the quality evaluation unit can more reasonably define the answer quality standard, and the selected optimal answer is more in line with the user's accuracy requirements for professional knowledge. This mechanism avoids including unreliable knowledge in the answer, making the output content more in line with user expectations in terms of professionalism and credibility. The introduction of the self-optimization learning module enables the system to have the ability of autonomous iteration. The feedback analysis unit adjusts the parameters of the answer generation unit according to user interaction data, enabling the system to learn from actual conversations and continuously optimize the generation logic of the answer. The strategy adjustment unit optimizes the retrieval strategy based on the optimal answer and historical conversation data, making the knowledge retrieval more targeted and enabling faster positioning of relevant knowledge when facing similar problems. The knowledge update unit dynamically updates the system knowledge base based on external knowledge sources, ensuring that the professional knowledge mastered by the system is always synchronized with the development of the field, and avoiding answer failure caused by outdated knowledge. The multi-round conversation management unit of the interactive presentation module can flexibly adjust the conversation process according to the optimal answer and the user's subsequent questions, making the conversation more coherent and logical. When the user asks follow-up questions, provides additional information, or changes the topic, the system can naturally connect and maintain the smoothness of the conversation. The visual display unit converts the optimal answer into natural language text output and collects user feedback signals. This two-way interaction mechanism not only makes the information acquisition process more convenient for users, but also provides direct information input for the continuous optimization of the system, enabling the system to better adapt to the expression habits and demand characteristics of different users. The cooperative operation of the modules forms a closed-loop operation system, from accurate identification of user intention to efficient retrieval and processing of knowledge, to autonomous optimization and friendly interaction of the system, which comprehensively improves the service experience of professional knowledge question and answer, making the intelligent customer service system more adaptable and practical in professional fields. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The working principle diagram of the professional knowledge question and answer intelligent customer service system based on the self-optimization mechanism described in the present application; Figure 2 a detailed flow chart of the user intention analysis module; Figure 3 a detailed flow chart of the entity linking model; Figure 4 a dynamic adjustment flow chart of the quality evaluation unit. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0019] Please refer to Figure 1 The present application provides a professional knowledge Q&A intelligent customer service system based on a self-optimization mechanism, which comprises: The professional knowledge Q&A intelligent customer service system based on the self-optimization mechanism is composed of a user intention analysis module, a knowledge processing module, a self-optimization learning module and an interactive presentation module.

[0020] When the user intention analysis module is running, the semantic understanding unit contained therein processes the user input information to generate a user intention signal. After the signal is transmitted to the knowledge processing module, the knowledge retrieval unit in the knowledge processing module processes the user intention signal, filters out relevant knowledge segments and calculates the knowledge confidence degree. The answer generation unit generates a candidate answer in combination with the relevant knowledge segments and the user intention signal. The quality evaluation unit determines the answer quality standard according to the knowledge confidence degree and selects the optimal answer from the candidate answers.

[0021] After the self-optimization learning module receives the optimal answer and related data, the feedback analysis unit adjusts the parameters of the answer generation unit with the help of user interaction data. The strategy adjustment unit optimizes the retrieval strategy by referring to the optimal answer and historical dialogue data. The knowledge updating unit realizes the dynamic updating of the system knowledge base through external knowledge sources.

[0022] After the interactive presentation module obtains the optimal answer, the multi-round dialogue management unit flexibly adjusts the dialogue process according to the optimal answer and the user's subsequent questions. The visual display unit converts the optimal answer into a natural language text and outputs it, while collecting the user feedback signal and sending it to the feedback analysis unit.

[0023] Embodiment 1: Please refer to Figure 2The user intention analysis module is composed of a distributed semantic analysis unit, a context awareness unit, and a semantic understanding unit, and covers the deployment range of customer service interaction platforms, mobile terminals, and web interfaces. It can simultaneously receive information submitted by users in various forms such as text input, voice calls, image uploads, etc. In the running process, the context awareness unit relies on dialogue history tracking technology to capture all interactive content before and after the user's current session in real time, including the user's previous questions, the historical answers of the customer service system, the user's pause time during the conversation, repeated content, etc. After integrating these information, the context signal is formed. This context signal is continuously updated as the conversation progresses to fully reflect the user's intention evolution during the entire interaction process.

[0024] After receiving the user input information and the context signal, the semantic understanding unit starts the multi-modal semantic fusion process. For text information input by the user, the keywords, phrases, and sentence structures are directly extracted; for voice information, it is first converted into text, and then the same structure analysis is performed; for image information, entity elements and scene features are extracted through image recognition technology, and then converted into corresponding text descriptions. Then, these information from different modalities are associated and matched with the context signal, for example, when a pronoun appears in the user's current input question, the specific object referred to by the pronoun is determined through the context signal; when the user's input content is ambiguous, the historical dialogue theme in the context signal is combined to exclude irrelevant interpretations, and finally the user intention signal containing the retrieval intention signal, the generation intention signal, and the evaluation intention signal is generated. The retrieval intention signal points to the specific direction in which the system needs to find relevant information from the knowledge base, the generation intention signal specifies the type and structure of the answer that the system needs to build, and the evaluation intention signal provides a reference for subsequent judgment of whether the answer meets the user's needs.

[0025] The knowledge retrieval unit includes an entity linking model based on a knowledge graph and a reinforcement learning retrieval strategy model. The retrieval intention signal includes user question keywords, entity type data, context correlation data, historical retrieval records, and domain knowledge classification data. User question keywords are core words extracted from user input information, such as product names, technical terms, etc.; entity type data is used to define the category to which the keywords belong, such as "software" belonging to "product type" and "installation steps" belonging to "operation process type"; context correlation data quantifies the correlation between the current question and each part of the context signal, with a higher value indicating a closer correlation; historical retrieval records include content previously queried by the user in the system and the corresponding retrieval path; and domain knowledge classification data classifies the question into a specific professional field category, such as finance, medicine, education, etc.

[0026] After these data are input into the entity linking model, the entity linking model first preprocesses the keywords of the user question, removes stop words, and performs synonym replacement, and then matches the processed keywords with entities in the knowledge graph, while combining entity type data to filter out entities that meet the category requirements. Next, referring to the context correlation data and historical search records, the entities unrelated to the current dialogue topic are excluded, and finally the knowledge nodes related to the user's intention are determined, and the content corresponding to these nodes is extracted from the knowledge base to form relevant knowledge fragments.

[0027] The reinforcement learning retrieval strategy model fuses the context correlation data, domain knowledge classification data, and knowledge confidence. In the fusion process, the model determines the weight distribution of different professional fields according to the domain knowledge classification data, for example, the matching accuracy weight of professional terms is higher in the medical field; adjusts the influence degree of different historical dialogue content on the current retrieval according to the context correlation data; takes the knowledge confidence as an index to measure the reliability of the knowledge fragment, and dynamically adjusts the priority of the relevant knowledge fragment through multiple iterations, and finally outputs the optimized knowledge confidence, providing a basis for subsequent answer generation and quality evaluation.

[0028] Embodiment 2: refer to Figure 3 The entity linking model in the knowledge retrieval unit is composed of an entity feature extraction submodule, a context fusion submodule, a correlation calculation submodule, and a retrieval result sorting module, and each submodule cooperates to realize accurate retrieval of relevant knowledge fragments.

[0029] After receiving the keywords of the user question, the entity feature extraction submodule first performs standardization processing on the keywords, including removing special symbols, unifying the format of the words, etc., and then uses a pre-trained word vector model to perform embedding conversion on the processed keywords, mapping each keyword to a vector representation in a high-dimensional vector space. These vectors contain semantic information and context features of the keywords, which can more accurately reflect the association between words. After the vector conversion is completed, the submodule filters the vectors according to the preset entity recognition threshold condition, and extracts the entity features that meet the threshold requirements as candidate entity signals. The entity recognition threshold condition is set based on factors such as the frequency of entity appearance in domain knowledge and semantic importance, and through this process, entities with weak association with the user question can be preliminarily excluded, reducing the data volume for subsequent processing.

[0030] After receiving the entity type data, context relevance data, and domain knowledge classification data, the context fusion sub-module uses a bidirectional Transformer network with attention mechanism weighting to process the data. The bidirectional Transformer network can capture both forward and reverse dependencies in the data, allowing for a more comprehensive understanding of the context information. The attention mechanism assigns different weights to different data, for example, when the entity type data is highly matched with the core entity of the user's current question, its weight is increased; the historical dialogue content directly related to the current question in the context relevance data also gets a higher weight; the domain knowledge classification data adjusts the weight distribution according to the characteristics of the professional field to which the question belongs, ensuring that the knowledge features of different fields are fully reflected. The weighted data is processed in the bidirectional Transformer network through multi-layer encoding, and the semantic association between the data is fused during the encoding process to form a fused context feature vector, which is output to the relevance calculation sub-module.

[0031] After receiving the feature vector output by the context fusion sub-module, the relevance calculation sub-module calculates according to the pre-defined relevance scoring rules. The relevance scoring rules cover multiple dimensions, including the matching degree of entities and question keywords, the frequency of entity appearance in the context, the fit degree of entity and domain knowledge classification, etc. Each dimension has corresponding scoring standards, for example, when the entity and question keywords are completely matched, the score is full marks, and when they are partially matched, the score is calculated according to the matching proportion; the higher the frequency of entity appearance in the context, the higher the score of the corresponding dimension. The sub-module calculates the scores of each dimension comprehensively to obtain the relevance score of each candidate entity signal and user intent, and the scoring result is presented in the form of specific numerical values, and the higher the numerical value, the stronger the relevance.

[0032] After receiving the relevance score output by the relevance calculation sub-module, the retrieval result sorting module sorts the relevant knowledge fragments according to the priority based on the multi-layer sorting strategy. The multi-layer sorting strategy includes two stages of preliminary sorting and fine sorting. In the preliminary sorting stage, the knowledge fragments are roughly sorted according to the relevance score, and the knowledge fragments with higher scores are selected to the front; in the fine sorting stage, the sorting results are adjusted by combining historical retrieval records, user feedback data, etc. For example, if a knowledge fragment is selected multiple times in historical retrieval and receives positive feedback, even if its relevance score is slightly lower than that of other fragments, it may be promoted in the fine sorting. In the final sorting result, different sorting positions represent different relevance scores of knowledge fragments, and the knowledge fragments in the front will be used preferentially in the subsequent answer generation process, and the sorting result will be transmitted to the answer generation unit to provide a basis for the construction of candidate answers.

[0033] Embodiment 3: The operation of the answer generation unit begins with the generation of an intent signal and the reception of relevant knowledge fragments. The intent signal contains a question type feature and an answer style parameter. The former is used to define the attributes of the user's question, such as fact query, process consultation, troubleshooting, etc., and the latter involves the formal degree, detailed degree, and frequency of professional term use. The generative question and answer model is constructed based on these parameters and relevant knowledge fragments, and the encoder-decoder architecture is used to realize text generation. The encoder jointly encodes the relevant knowledge fragments and the generated intent signal, converting them into high-dimensional semantic vectors. In the process, the key information in the knowledge fragments is strongly encoded to ensure that the core content is fully captured. The decoder generates text sequences based on the semantic vectors output by the encoder. During the generation process, the output results are optimized by dynamically adjusting the attention weights, such as increasing the proportion of process-related knowledge fragments in attention allocation when generating answers involving specific steps, and enhancing the attention weight of domain knowledge-related fragments when generating answers containing professional terms, to balance the fluency and knowledge accuracy of the answers.

[0034] After generating the candidate answers, the particle swarm optimization algorithm is used to process them to obtain the optimal answer. The particle swarm optimization algorithm regards each candidate answer as a particle in the solution space, and optimizes by iteratively updating the particle position. In the initial stage of the algorithm, a random initial position and velocity are assigned to each candidate answer. The position represents the optimization state of the answer in the solution space, and the velocity represents the optimization direction and amplitude. During the iteration process, each particle adjusts its motion trajectory based on its historical best position and the historical best position of the group. The historical best position of the particle is the best state of the candidate answer in the previous iteration, and the historical best position of the group is the best state of all candidate answers. When the iteration reaches a predetermined number of times or the particle position changes less than a threshold value, the optimization stops, and the candidate answer corresponding to the historical best position of the group is the optimal answer.

[0035] The professional corpus dynamic library included in the answer generation unit stores the corpus features of historical question and answer pairs, including sentence structure, vocabulary selection, and logical connection methods. The sources of the corpus include system historical dialogue records, manually annotated high-quality question and answer samples, etc., and new corpus will be continuously supplemented as the system runs to ensure that the content in the library keeps pace with changes in user needs. When generating candidate answers, the generative question and answer model will refer to the features in the professional corpus dynamic library to ensure that the output answers are consistent with historical high-quality answers in terms of expression form and meet user expectations for the expression habits of question and answer in this field.

[0036] The syntax and logic checking submodule starts working after the candidate answers are generated, and detects the candidate answers according to the characteristics of the question type and the related knowledge fragments. For the detection of syntax errors, a rule-based and statistical method is adopted, the rule level covers basic syntax specifications such as subject-verb agreement, time matching, and punctuation usage, and the statistical level identifies the sentence structure that does not conform to common expression habits by comparing the sentence distribution in the dynamic library of professional corpus. For the detection of logical contradictions, the focus is on the logical coherence within the answer and the consistency with the related knowledge fragments, for example, when answering process type questions, the order of the steps is checked to see if it is reasonable; when answering factual questions, the answer content is verified to see if there is a conflict with the facts in the knowledge fragments. The detected syntax errors and logical contradictions are marked and fed back to the generative question answering model as the basis for adjusting the decoder parameters.

[0037] The answer style adaptation submodule receives the answer style parameters, the generative question answering model, and the user intent signal, and adjusts the style of the candidate answers through a particle swarm optimization algorithm. During the adjustment process, the answer style parameters are taken as the target, and the difference between the style features of the candidate answers and the target parameters is taken as the optimization objective function, as follows: wherein, represents the total difference degree of style adaptation, represents the actual style feature value of the candidate answer, represents the target value of the answer style parameter, represents the matching degree of the content of the candidate answer and the user intent signal, represents the target value of the content matching degree, and respectively represent the weight coefficients of the style feature difference and the content matching degree difference. By minimizing the value of , the candidate answers are adapted to the parameter requirements in style, while ensuring the matching of the content and the user intent. For example, when the answer style parameter requires “conciseness”, the submodule will shorten the sentence length and delete redundant adjectives through the algorithm; when the parameter requires “professionalism”, the proportion of domain terminology will be increased, and a more rigorous logical structure will be adopted. After the style adaptation processing, the candidate answers will be detected again by the syntax and logic checking submodule, and finally the optimal answer that meets the requirements is formed.

[0038] Example 4: see Figure 4When the quality evaluation unit is running, it receives an evaluation intention signal containing user satisfaction feedback and professional accuracy data. The quality evaluation unit defines answer quality indicators based on these data, which involve multiple aspects such as the relevance of the answer content to the user's question, the completeness of the information, the clarity of the expression, etc. After determining the indicators, the quality evaluation unit scores the generated answer and outputs the answer quality score. At the same time, the quality evaluation unit dynamically adjusts the answer quality score through fuzzy PID control, so that the score does not fall below the quality threshold. In the process of fuzzy PID control, the decoding parameters and knowledge retrieval threshold of the generative question and answer model are dynamically adjusted according to the real-time score. For example, when it is found that the answer quality score has a trend of falling below the threshold, the fuzzy PID control adjusts the focus parameter of the decoder on the knowledge fragments in the generative question and answer model, so that the decoder focuses more on the core knowledge; at the same time, it also adjusts the knowledge retrieval threshold, so that the knowledge retrieval unit can filter out more relevant knowledge fragments, so as to maintain the answer quality score above the threshold.

[0039] The finite state machine dialogue management unit in the interactive presentation module continuously receives the optimal answer and user feedback signals. The finite state machine contains multiple dialogue states, such as initial inquiry state, information provision state, question answering state, and end of dialogue state, etc. Each state has corresponding state transition rules. When receiving the optimal answer, the finite state machine dialogue management unit determines the current dialogue state according to the content of the optimal answer, and then judges whether state transition is needed in combination with the user feedback signal. For example, when the user feedback signal shows that the user still has questions about the current answer, the finite state machine dialogue management unit will transfer the dialogue state from the information provision state to the question answering state, and adjust the subsequent dialogue process to guide the user to further explain the question; if the user feedback signal indicates that the problem has been solved, the dialogue state may be transferred to the end of dialogue state, or whether there are other needs may be asked.

[0040] The multi-template response generation module uses an improved fuzzy comprehensive evaluation algorithm to process the answer quality score and the user intention signal, and then outputs a natural language answer. The improved fuzzy comprehensive evaluation algorithm considers the level of the answer quality score and the answer type and structure information specified by the generation intention signal in the user intention signal, and selects the most matching template from the pre-set multiple answer templates. These templates cover different expression methods, some are concise and clear, some are detailed and specific, some focus on professional terms, and some focus on popular explanations. For example, when the user intention signal shows that the user needs to quickly understand the general steps of a process, and the answer quality score is high, the multi-template response generation module will select a concise process overview template; if the user intention signal indicates that the user needs to understand a technical detail in depth, and the answer quality score also meets the requirements, the module will select a detailed technical analysis template.

[0041] When the quality score of the answer is below the quality threshold, the interaction presentation module sends a signal to the knowledge processing module, triggering the knowledge processing module to regenerate the candidate answer. The knowledge processing module restarts the workflow of the knowledge retrieval unit and the answer generation unit, the knowledge retrieval unit retrieves relevant knowledge fragments again, the answer generation unit regenerates the candidate answer based on the newly retrieved knowledge fragments and the user intent signal, and then the quality assessment unit evaluates the new candidate answer until the optimal answer that meets the quality requirements is generated, which is then output by the interaction presentation module.

[0042] The multi-template response generation module supports the switching of multi-style answer templates, which enables the system to flexibly adjust the style of the answer according to different scenarios and user needs. During the output of the answer, the module displays the progress of the answer generation in real time, such as showing the completion ratio of the current answer generation in the form of a progress bar, so that the user can understand the waiting state. At the same time, the source of knowledge is also marked, such as marking that a part of the information in the answer comes from a chapter of the system knowledge base or from a certain external authoritative material, so that the user can clearly understand the source of the information.

[0043] Based on the rule engine, the multi-template response generation module can simulate the answering tone and expression mode under different user portraits. The rule engine stores rules corresponding to different user portraits, which may be based on the user's age, occupation, communication style, etc. For example, for the portrait of a young user, the rule engine triggers the lively and colloquial tone rule, and some popular vocabulary may be used in the answer; for the user portrait marked as a professional, the rule engine will call the strict and professional tone rule, and the answer will use more professional terms, and the logical structure will be more rigorous. In this way, the generated natural language answer is more suitable for different users' preferences, and the user's experience in the interaction process is improved.

[0044] During the entire interaction process, the interaction presentation module continuously collects user feedback signals and transmits them to the feedback analysis unit in the self-optimization learning module, providing a basis for the self-optimization of the system, and also allowing the interaction presentation module to continuously adjust its own working mode according to user feedback to better complete the interaction presentation task with the user.

[0045] Embodiment 5: The feedback analysis unit is provided with state identifiers based on a multi-level feedback processing mechanism, including a general improvement state, an important optimization state, and an emergency repair state. These state identifiers correspond to different user feedback signal processing strategies to address different levels of user satisfaction. When the satisfaction score in the user feedback signal falls below the first score threshold, the system automatically enters the general improvement state. In this state, the feedback analysis unit retrieves the current parameters of the answer generation unit, including the attention weight distribution of the decoder and the training coefficients of the generative question and answer model, and then adjusts the weights of these parameters within the range of regular optimization according to the problem types reflected in the user interaction data. For example, if the user feedback shows that the satisfaction with product parameter-based answers is low, the feedback analysis unit will increase the weight proportion of knowledge fragments related to product parameters in the answer generation unit, while reducing the weight influence of other non-related domain knowledge.

[0046] When the satisfaction score in the user feedback signal falls below the second score threshold, the system enters the important optimization state, and the current dialogue process is paused. The feedback analysis unit sends a re-search instruction to the knowledge processing module, and the knowledge processing module receives the instruction. The knowledge retrieval unit adjusts the retrieval strategy, expands the retrieval range, and re-invokes the entity linking model based on the knowledge graph and the reinforcement learning retrieval strategy model to process the user intent signal again. During the secondary processing, the knowledge retrieval unit increases the weights of the context association data and the domain knowledge classification data, re-selects relevant knowledge fragments and calculates the knowledge confidence, and the answer generation unit generates new candidate answers based on the new knowledge fragments and user intent signals. The quality assessment unit evaluates the new candidate answers again until the optimal answer meeting the quality requirements is output, and the dialogue process resumes.

[0047] When the satisfaction score in the user feedback signal falls below the third score threshold, the system triggers the emergency repair state, and the feedback analysis unit sends a system knowledge defect signal. This signal is transmitted to the self-optimization learning module simultaneously, and the knowledge update unit in the self-optimization learning module immediately starts the emergency update program, while the strategy adjustment unit also pauses the regular retrieval strategy optimization and cooperates with the emergency update work of the knowledge update unit.

[0048] The presentation of the optimal answer is diverse, including not only the problem-solving text but also relevant knowledge links pointing to detailed documents in the knowledge base or external authoritative resources related to the current problem. The recommended question list lists questions that the user may be interested in subsequently based on the theme of the current question and the user's historical search records. For example, after the user asks about the installation method of a certain software, the recommended question list may include the uninstallation steps of the software and common installation troubleshooting. The next operation suggestion provides guidance based on the solution progress of the question, such as "if there are still problems with the installation, you can click here to submit a screenshot for further assistance."

[0049] The knowledge updating unit contains an intelligent crawler component that regularly monitors the updates of external knowledge sources, including industry official websites, professional databases, academic journals, etc. The intelligent crawler component analyzes the update logs and page modification records of the knowledge sources to identify their update frequency. For frequently updated knowledge sources, such as official websites that publish real-time policy changes, the monitoring period is shortened, and the latest professional knowledge is captured as needed, such as new clause content and modified regulations. For slower updating knowledge sources, the monitoring period is appropriately extended to reduce system resource consumption. The new knowledge captured is temporarily stored in the knowledge base after format conversion and redundant information removal.

[0050] After receiving the knowledge confidence, the knowledge graph updating device adjusts the entity association strength in the knowledge graph in real time based on the entity relationship reasoning algorithm. The entity relationship reasoning algorithm analyzes the association method of new knowledge and existing entities in the knowledge graph. If the association of two entities in the new knowledge is confirmed multiple times and the knowledge confidence is high, the association strength of these two entities in the knowledge graph is enhanced. If a conflict is found between the original entity association and the new knowledge, and the knowledge confidence of the new knowledge is higher, the original association is weakened or modified, and the reason and time of the association adjustment are recorded for subsequent tracing.

[0051] User feedback signals cover multiple aspects. Satisfaction scores are expressed in the form of stars or scores, reflecting the user's overall evaluation of the answer. Answer usefulness labels include "useful", "not useful", "partially useful", etc., which are used to specifically feedback the actual help degree of the answer. The problem solving status is divided into "solved", "unsolved", "partially solved", reflecting the processing progress of the user's problem. The supplementary information input allows users to provide more details related to the problem in the form of text, pictures, etc., such as "the step mentioned in the previous answer missed the XX link, which needs to be paid attention to in actual operation". These feedback signals are transmitted to the feedback analysis unit in real time, becoming an important basis for system self-optimization.

[0052] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0053] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. A professional knowledge question and answer intelligent customer service system based on a self-optimization mechanism, characterized in that, Comprise: a user intention analysis module comprising a semantic understanding unit that generates a user intention signal; a knowledge processing module that receives the user intention signal, the knowledge processing module comprising a knowledge retrieval unit, an answer generation unit, and a quality assessment unit, the knowledge retrieval unit processing the user intention signal to output relevant knowledge segments and a knowledge confidence level, the answer generation unit processing the relevant knowledge segments and the user intention signal to obtain a candidate answer, and the quality assessment unit defining an answer quality standard based on the knowledge confidence level to output an optimal answer; a self-optimizing learning module comprising a feedback analysis unit that adjusts parameters of the answer generation unit based on user interaction data, a strategy adjustment unit that optimizes retrieval strategies based on the optimal answer and historical dialogue data, and a knowledge updating unit that dynamically updates a system knowledge base based on external knowledge sources; an interactive presentation module comprising a multi-round dialogue management unit that adjusts dialogue processes based on the optimal answer and user follow-up questions, and a visual display unit that converts the optimal answer into natural language text and outputs it, while collecting user feedback signals and sending them to the feedback analysis unit.

2. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 1, characterized in that, The user intention analysis module comprises a distributed semantic analysis unit and a context awareness unit, and is deployed on a customer service interaction platform, a mobile terminal, and a web interface, and collects user input information, the context awareness unit obtains user conversation context data through dialogue history tracking technology and forms a context signal, the semantic understanding unit receives the user input information and the context signal to realize multi-modal semantic fusion and form the user intention signal, and the user intention signal comprises a retrieval intention signal, a generation intention signal, and an evaluation intention signal.

3. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 2, characterized in that, The knowledge retrieval unit comprises an entity linking model based on a knowledge graph and a reinforcement learning retrieval strategy model, the retrieval intention signal comprises user question keywords, entity type data, context correlation data, historical retrieval records, and domain knowledge classification data, the user question keywords, the historical retrieval records, the entity type data, the context correlation data, and the domain knowledge classification data are input into the entity linking model to output the relevant knowledge segments, and the context correlation data, the domain knowledge classification data, and the knowledge confidence level are fused through a deep reinforcement learning strategy model to output the knowledge confidence level.

4. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 3, characterized in that, The entity linking model comprises an entity feature extraction submodule, a context fusion submodule, a correlation degree calculation submodule and a retrieval result sorting module, the entity feature extraction submodule performs word vector embedding on the user question keywords, extracts entity features meeting an entity recognition threshold condition as candidate entity signals, the context fusion submodule processes the entity type data, the context correlation degree data and the domain knowledge classification data using a bidirectional Transformer network weighted by an attention mechanism and outputs the data to the correlation degree calculation submodule, the correlation degree calculation submodule defines a correlation degree scoring rule, and the retrieval result sorting module generates a priority order of the relevant knowledge segments based on a multi-layer sorting strategy and uses different sorting positions to represent different correlation degree scores of the knowledge segments.

5. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 2, characterized in that, The answer generation unit processes the generation intent signal and the relevant knowledge segments through a generative question answering model, outputs a candidate answer, and processes the candidate answer using a particle swarm optimization algorithm to obtain an optimal answer; the generation intent signal comprises a question type feature and an answer style parameter, the generative question answering model is constructed through the question type feature, the relevant knowledge segments and the answer style parameter, and the generative question answering model realizes text generation based on an encoder-decoder architecture, and adjusts the decoder attention weight to optimize the fluency and knowledge accuracy of the answer.

6. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 5, characterized in that, The answer generation unit further comprises a professional corpus dynamic library, a syntax logic verification submodule and an answer style adaptation submodule, the professional corpus dynamic library stores corpus features of historical question and answer pairs, the syntax logic verification submodule detects syntax errors and logical contradictions of a candidate answer according to the question type feature and the relevant knowledge segments, and the answer style adaptation submodule obtains the optimal answer through a particle swarm optimization algorithm according to the answer style parameter, the generative question answering model and the user intent signal.

7. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 2, characterized in that, The quality evaluation unit defines an answer quality index according to the evaluation intent signal, outputs an answer quality score, and dynamically adjusts the answer quality score to be not lower than a quality qualified threshold value through fuzzy PID control, the evaluation intent signal comprises user satisfaction feedback and professional accuracy data, the user satisfaction feedback and the professional accuracy data define the answer quality index, and the fuzzy PID control dynamically adjusts the decoding parameters of the generative question answering model and the knowledge retrieval threshold value to make the answer quality score not lower than the quality qualified threshold value.

8. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 1, characterized in that, The interactive presentation module includes a finite state machine dialogue management unit and a multi-template response generation module, the finite state machine dialogue management unit receives the optimal answer and the user feedback signal and dynamically adjusts the dialogue state transition rule, and the multi-template response generation module processes the answer quality score and the user intention signal according to an improved fuzzy comprehensive evaluation algorithm and outputs a natural language answer, when the answer quality score is lower than a quality qualified threshold, the interactive presentation module controls the knowledge processing module to regenerate a candidate answer, the multi-template response generation module supports multi-style answer template switching, can display answer generation progress and knowledge source annotation in real time, and simulates answer tone and expression mode under different user portraits based on a rule engine.

9. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 8, characterized in that, The feedback analysis unit includes state identification of a multi-level feedback processing mechanism, which is a general improvement state, an important optimization state and an emergency repair state respectively, when the satisfaction score in the user feedback signal drops to a first score threshold, the general improvement state is started, and the parameter weight of the answer generation unit is adjusted to a regular optimization range, when the satisfaction score in the user feedback signal drops to a second score threshold, the current dialogue process is suspended and the knowledge processing module is controlled to re-search knowledge, when the satisfaction score in the user feedback signal drops to a third score threshold, the feedback analysis unit issues a system knowledge defect signal and controls the self-optimization learning module to update the knowledge base urgently.

10. The self-optimization mechanism based expert Q&A intelligent customer service system according to claim 1, characterized in that, The optimal answer includes but is not limited to question answer text, related knowledge link, recommended question list and next step operation suggestion, the knowledge updating unit includes an intelligent crawler component and a knowledge graph updating device, the intelligent crawler component can identify the update frequency of external knowledge sources and grab the latest professional knowledge as needed, the knowledge graph updating device receives the knowledge confidence and adjusts the entity association strength of the knowledge graph in real time based on an entity relationship reasoning algorithm, and the user feedback signal includes but is not limited to satisfaction score, answer usefulness label, problem solving state and supplementary information input.

Images