Core System for Intelligent Insurance Purchase Decision-making Based on Dynamic Dialogue Strategy

Abstract

The present invention relates to the fields of artificial intelligence and insurance technologies, and discloses a core system for intelligent underwriting decision-making based on a dynamic dialogue strategy. The system includes a user interaction module that collects user risk attribute data through multi-modal inputs such as voice and text; an insurance product database that stores structured data of multi-source insurance products; an intelligent decision-making module that generates personalized underwriting suggestions by using multi-model fusion technologies such as random forest and gradient boosting tree. There is also a dynamic dialogue strategy engine that generates dynamic interaction sequences based on context-aware reinforcement learning. The data synchronization method of the system can obtain the latest insurance product data, the multi-model fusion technology improves the decision-making accuracy, and the dynamic dialogue strategy realizes intelligent interaction, which can effectively solve the problems of incomplete information collection, inaccurate decision-making, and inflexible interaction in the traditional underwriting decision-making process, and provide more efficient and personalized underwriting decision-making services for users.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and insurance, and specifically to a core system for intelligent insurance purchase decision-making based on a dynamic dialogue strategy. Background Art

[0002] In the traditional insurance business model, there are many pain points in the insurance purchase decision-making process, which seriously affect the efficiency, accuracy, and user experience of insurance services.

[0003] From the perspective of information interaction, previous insurance purchase consultations mainly relied on single text or limited manual communication methods. Text input poses a threshold for some users who are not good at expressing themselves in writing, such as the elderly or those with a low level of education. They are difficult to accurately describe their complex risk situations and protection needs, resulting in incomplete and inaccurate information transmission. Although manual communication can solve the expression problem to a certain extent, it is limited by the professional level differences of insurance agents and working time and energy constraints, and cannot provide timely and professional consultation services for a large number of users at any time. This makes it impossible for users to fully elaborate their situations when purchasing insurance, and it is also difficult for insurance companies to comprehensively understand the risk attributes of users, laying a hidden danger for subsequent accurate insurance purchase decisions.

[0004] In terms of insurance product data management, there are a large number of insurance products in the insurance market with frequent updates. The data formats of different insurance companies vary greatly, which brings great challenges to data integration and unified management. On the one hand, the heterogeneity of data leads to a large amount of manpower and time consumption for format conversion and data cleaning when collecting and sorting insurance product information, with low efficiency and easy errors. On the other hand, in the face of a large amount of insurance product data, there is a lack of effective data synchronization and update mechanisms, resulting in lagging product information and unable to reflect the latest market dynamics. For example, the terms and rates of newly launched insurance products cannot be reflected in the system in a timely manner, or the adjusted claim conditions are not updated in a timely manner, which may lead to inaccurate product information obtained by users and incorrect insurance purchase decisions.

[0005] There are also serious deficiencies in the insurance purchase decision-making analysis link. Traditional insurance purchase suggestions are often based on simple rules of thumb or general risk assessment models, lacking in-depth exploration and precise matching of users' personalized risk characteristics. For example, for users with special occupational risks (such as high-altitude workers, chemical industry practitioners, etc.) or complex family financial situations (such as multiple properties, multiple investment portfolios, etc.), traditional methods cannot fully consider the impact of these unique factors on insurance needs, and the insurance purchase suggestions provided may not truly meet the protection needs of users. Either the protection is insufficient and cannot provide sufficient economic compensation when a risk occurs, or there is over-insurance, resulting in an increased economic burden on users.

[0006] Meanwhile, throughout the insurance application process, there is a lack of a dynamic interaction mechanism. When users have questions about insurance products or situations change, the insurance application plan cannot be adjusted in a timely manner. For example, if a user discovers new health problems during the insurance application process or there is a sudden change in the family's economic income, the existing insurance application decision-making system is difficult to re-evaluate based on these real-time changes and give corresponding suggestions, resulting in a lack of flexibility and adaptability in the insurance application decision-making. Summary of the Invention

[0007] The purpose of the present invention is to provide a core intelligent decision-making system for insurance application based on dynamic dialogue strategies to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solutions: A core intelligent decision-making system for insurance application based on dynamic dialogue strategies, the system includes:

[0009] User interaction module: used to conduct dialogue interaction with users through multimodal input methods and collect user risk attribute data in real time;

[0010] Insurance product database: stores structured data of multi-source insurance products, including coverage scope, rate rules, and claim settlement conditions;

[0011] Intelligent decision-making module: uses multi-model fusion technology to match and analyze user risk attributes and insurance product data, and generates personalized insurance application suggestions;

[0012] The multi-model fusion technology of the intelligent decision-making module includes:

[0013] Construct a random forest model, screen high-risk correlation factors through feature importance ranking, and generate a preliminary risk rating;

[0014] Construct a gradient boosting tree model, optimize the risk score granularity based on residual fitting, and output a fine-grained risk probability distribution;

[0015] Design a dynamic weighted fusion layer, calculate the confidence weights of the outputs of each model according to the user risk attribute distribution, and generate a weighted decision vector.

[0016] Preferably, the user interaction module includes:

[0017] The multimodal input method includes a voice input channel and a text input channel. The voice input channel uses an end-to-end speech recognition model to generate a text intermediate representation;

[0018] Construct a natural language processing unit enhanced by domain knowledge, extract risk entities in the user input through an entity recognition network, and construct a semantic dependency graph in combination with historical dialogue records;

[0019] Design a double-buffer storage mechanism. The first buffer stores the raw data of the current conversation turn, and the second buffer stores the historical conversation context using a timestamp indexing mechanism to achieve cross-turn context association.

[0020] Preferably, the system further includes:

[0021] A dynamic dialogue policy engine: Based on a context-aware reinforcement learning model, it combines a preset policy template and the real-time dialogue state to generate a dynamic interaction sequence, and the dynamic interaction sequence includes risk information collection questions and feedback response logic.

[0022] Preferably, the dynamic dialogue policy engine includes:

[0023] Construct a hierarchical state space model, where the upper state represents the completeness of user risk information, and the lower state represents the confidence level of the dialogue intention;

[0024] Design an information gain evaluation function to dynamically adjust the priority of question generation based on the change in information entropy;

[0025] Adopt an asynchronous policy optimization algorithm to train the reinforcement learning model, and each policy iteration cycle integrates an offline experience replay pool and online real-time interaction data.

[0026] Preferably, the data synchronization method of the insurance product database includes:

[0027] Connect to the data interface of the insurance company, and convert the heterogeneous data format into a unified attribute matrix through a protocol parsing component;

[0028] Construct an incremental web crawler framework, use dynamic rendering technology to crawl the details of insurance products on the web page, and extract key fields in the product terms through a rule engine;

[0029] Design a version conflict detection mechanism. When detecting multiple versions of data for the same product, automatically select the latest version based on the timestamp and digital signature.

[0030] Preferably, the hierarchical state space model updates the state using an attention gating mechanism, including:

[0031] Define the dialogue state update rule, and perform a cosine similarity match between the keyword vector of the user's answer and the preset risk dimension;

[0032] Construct a gated recurrent unit network, fuse the similarity score with the historical state vector to generate the current dialogue state encoding;

[0033] Calculate the association strength between different risk dimensions through a multi-head self-attention mechanism, and dynamically adjust the state transition probability.

[0034] Preferably, the rule engine uses a semantic graph convolutional network to extract key fields, including:

[0035] Convert the insurance clause text into a dependency syntax graph, where the nodes in the graph represent clause statements and the edges represent logical association relationships;

[0036] Define a multi-hop graph attention layer to extract clause constraint conditions by iteratively aggregating adjacent node features;

[0037] Construct a conditional trigger rule library, and perform fuzzy matching on the extracted constraint conditions with the preset trigger patterns to generate standardized product attribute labels.

[0038] Preferably, the dynamic weighted fusion layer uses dual-channel confidence evaluation, including:

[0039] The first channel calculates the out-of-bag error rate of the random forest model as an indicator of model stability;

[0040] The second channel analyzes the loss function convergence curve of the gradient boosting tree model and calculates the model fitting degree indicator;

[0041] Through a fuzzy logic inference system, comprehensively consider the stability and fitting degree indicators to generate the weight distribution coefficients of each model.

[0042] Preferably, the gated recurrent unit network introduces a memory enhancement mechanism, including:

[0043] Construct an external memory matrix to store the characteristics of high-risk cases in historical conversations;

[0044] Design a memory retrieval module to perform similarity retrieval based on the current user attribute vector and the memory matrix to generate a memory enhancement vector;

[0045] Concatenate the memory enhancement vector with the hidden state of the gated recurrent unit and generate the final state encoding through a fully connected layer.

[0046] Preferably, the multi-hop graph attention layer adopts a residual diffusion strategy, including:

[0047] Define the node feature propagation rule, and after aggregating the features of each hop, superimpose the original node features as the residual term;

[0048] Design a diffusion attenuation coefficient to exponentially reduce the attention weight according to the logical hop count between nodes;

[0049] Eliminate the feature scale shift caused by multiple jumps through a feature normalization layer.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] At the user interaction level, the multi-modal input method has greatly improved the convenience of users' participation in the insurance purchase decision-making and the comprehensiveness of information collection. The voice input channel allows users to communicate with the system easily without manual typing, especially in busy or inconvenient scenarios for text input, which is particularly convenient for users who are not very proficient in operating electronic devices. The end-to-end voice recognition model can quickly and accurately convert voice into text intermediate representation, reducing the waiting time for users. The text input channel provides a way for users who are good at expressing themselves in words to accurately describe their own situations. The domain knowledge-enhanced natural language processing unit, combined with entity recognition network and semantic dependency graph construction technology, can not only accurately extract risk entities in the user input, but also understand the context semantics using historical conversation records, avoiding information omission and misunderstanding. The double-buffer storage mechanism realizes cross-turn context association, enabling the system to better track the user's thinking and providing a coherent information basis for subsequent interactions and decisions. For example, when the user mentions changes in their occupation and the health status of family members in different conversation turns, the system can automatically associate this information and conduct comprehensive analysis to provide insurance purchase suggestions that are more in line with the user's actual situation.

[0052] In terms of the insurance product database, its innovative data synchronization method ensures the timeliness, accuracy, and integrity of data. By docking with the insurance company's data interface and converting the data format through the protocol parsing component, it can quickly integrate insurance product information from different sources, breaking down data barriers. The incremental web crawler framework combined with dynamic rendering technology can capture the details of insurance products on the web page in real time and obtain the latest product information in a timely manner. The rule engine uses a semantic graph convolutional network to extract key fields and generates standardized product attribute labels through fuzzy matching with preset trigger patterns, improving the efficiency and accuracy of data processing. The version conflict detection mechanism automatically selects the latest version based on timestamps and digital signatures, avoiding information errors caused by chaotic data versions and enabling users to always obtain the latest and most reliable insurance product data to make more informed decisions when comparing and selecting products.

[0053] The multi-model fusion technology in the intelligent decision-making module enables highly personalized and precise insurance recommendations. The random forest model uses feature importance ranking to screen high-risk correlates, providing a preliminary, macro-level assessment of risk ratings and quickly identifying key factors influencing user risk. The gradient boosted tree model optimizes the granularity of risk scoring based on residual fitting, outputting a fine-grained risk probability distribution. This further refines risk assessment and accounts for the impact of various complex factors. The dynamic weighted fusion layer calculates the confidence weights for each model's output based on the user's risk attribute distribution, combining the strengths of both models to generate a more scientifically sound weighted decision vector. For example, for a user with a family medical history and a high-risk occupation, the random forest model may prioritize the impact of family medical history on health risks, while the gradient boosted tree model can further refine the specific risk probabilities for different disease and occupational risk combinations. The dynamic weighted fusion layer assigns appropriate weights to the two models based on the user's specific circumstances, ultimately providing insurance recommendations that better align with the user's actual risk profile.

[0054] The introduction of a dynamic dialogue strategy engine empowers the system with intelligent interaction capabilities. Based on a context-aware reinforcement learning model, it combines pre-set strategy templates with real-time dialogue status to generate dynamic interaction sequences. This allows for targeted questioning and feedback based on the user's current risk information collection progress and conversational intent. A hierarchical state-space model, representing the completeness of the user's risk information and the confidence level in the conversational intent, provides a clear state basis for strategy formulation. An information gain evaluation function dynamically adjusts question generation priorities, ensuring that the system prioritizes information most valuable for insurance decisions, improving information collection efficiency. An asynchronous strategy optimization algorithm integrates an offline experience replay pool with online real-time interaction data for model training, enabling the system to continuously learn and adapt to new situations. Over time, the interaction strategy becomes more intelligent and efficient. For example, if the system detects a user misunderstanding certain insurance terms, it can promptly adjust the interaction strategy based on the conversational status, providing more detailed explanations and instructions to guide the user in making the right insurance decision.

[0055] In addition, the system employs a series of advanced technologies in various key components, such as the memory enhancement mechanism of the gated recurrent unit network, the residual diffusion strategy of the multi-hop graph attention layer, and the dual-channel confidence assessment of the dynamic weighted fusion layer, all of which further optimize system performance from different perspectives. The memory enhancement mechanism allows the system to draw on historical high-risk case experience to better cope with complex situations; the residual diffusion strategy ensures the integrity and stability of features during the information extraction process; and the dual-channel confidence assessment makes the model weight distribution more scientific and reasonable. The combined application of these technologies has enabled the entire core system for intelligent insurance decision-making to achieve a high level of accuracy, flexibility, and adaptability, effectively improving the quality of insurance services, reducing users' insurance risks, and promoting the intelligent development of the insurance industry. Brief Description of the Drawings

[0056] Figure 1 This is the working principle diagram of the core system for intelligent insurance purchase decision-making based on dynamic dialogue strategies according to the present invention;

[0057] Figure 2 This is the flowchart of information processing in the user interaction module;

[0058] Figure 3 This is the working flowchart of data synchronization in the insurance product database;

[0059] Figure 4 This is the flowchart of state update in the hierarchical state space model. Detailed Embodiment

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0061] Please refer to Figures 1 - 4 , the present invention provides a core system for intelligent insurance purchase decision-making based on dynamic dialogue strategies. The overall implementation solution is as follows:

[0062] The system mainly includes a user interaction module, an insurance product database, and an intelligent decision-making module. The user interaction module conducts dialogue interaction with the user through multi-modal input methods, and real-time collects user risk attribute data. The multi-modal input methods can be voice input and text input, etc. When using voice input, the voice is converted into text through specific speech recognition technology; text input directly receives the text information input by the user. Through this multi-modal interaction method, it is possible to more conveniently obtain the risk-related information of the user, such as the attribute data of the user's age, occupation, health status, etc. that may affect the insurance purchase decision.

[0063] The insurance product database is used to store the structured data of multi-source insurance products, which covers key information such as the scope of protection, rate rules, and claim settlement conditions. These data provide an important reference basis for the subsequent generation of insurance purchase suggestions. The intelligent decision-making module uses multi-model fusion technology to match and analyze the user's risk attributes and insurance product data, and then generates personalized insurance purchase suggestions. Specifically, a random forest model is first constructed, and high-risk correlation factors are screened through feature importance ranking to generate a preliminary risk rating. Then a gradient boosting tree model is constructed, and the risk score granularity is optimized based on residual fitting to output a fine-grained risk probability distribution. Finally, a dynamic weighted fusion layer is designed to calculate the confidence weights of the outputs of each model according to the user's risk attribute distribution, generate a weighted decision vector, and combine the advantages of the two models to obtain a more accurate and reasonable insurance purchase suggestion.

[0064] The following further describes the implementation of the present invention in combination with Embodiments 1 to 8.

[0065] Embodiment 1:

[0066] In the specific process of the user interaction module, it includes multi-modal input processing, natural language processing enhanced by domain knowledge, and the construction and operation of a double-buffer storage mechanism. The specific implementation methods are as follows:

[0067] In the multi-modal input method, the voice input channel and the text input channel work together. For the voice input channel, the system uses an end-to-end speech recognition model to process the user's voice. When the user inputs voice information through an audio device such as a microphone, the voice signal first goes through a preprocessing stage, including operations such as noise reduction and normalization, to remove environmental noise and adjust parameters such as the volume and frequency of the voice signal, so as to improve the quality of the voice signal. The preprocessed voice signal is input into the end-to-end speech recognition model, which usually consists of multiple neural network layers. For example, a convolutional neural network (CNN) is used to extract local features of the voice signal, and a recurrent neural network (RNN) or its variants such as long short-term memory network (LSTM) and gated recurrent unit (GRU) are used to capture the temporal features of the voice signal. The model has learned the mapping relationship between voice and text based on a large amount of labeled voice data during the training stage. During operation, it directly maps the input voice signal to a text intermediate representation. Specifically, the model analyzes the voice signal frame by frame, predicts the corresponding phoneme or character for each frame, and then post-processes the prediction results through a language model to convert the phoneme sequence or character sequence into a text string that conforms to grammar and semantics, thus completing the voice-to-text conversion process. For the text input channel, the user directly inputs text information through methods such as a keyboard and touch input, and the system directly receives and stores this text data.

[0068] When building a domain knowledge-enhanced natural language processing unit, the system first uses an entity recognition network to extract risk entities from the text input by the user (including the text after speech conversion). The entity recognition network is built based on deep learning technologies. For example, a bidirectional encoder representation (BERT) model based on the Transformer architecture is used as the basic network, which is fine-tuned for insurance domain data on the basis of a pre-trained general language model. The training data comes from a large number of insurance-related text corpora, including insurance clauses, claim cases, industry reports, etc. Through manual annotation, the risk entities involved, such as age, occupation, health status, family situation, etc., are marked. During operation, the entity recognition network receives the user input text, analyzes each word in the text, and calculates the probability of each word belonging to different risk entity categories through the model, thereby identifying the risk entities in the text. After extracting the risk entities, the system constructs a semantic dependency graph in combination with the historical conversation records. The historical conversation records store the input content of the user in the previous conversation turns and the reply information of the system. The system uses natural language processing technologies, such as dependency syntax analysis, to analyze the user's current input text and the historical conversation records, and determines the semantic relationships between the words in the text, such as subject-predicate relationship, verb-object relationship, modification relationship, etc. Based on these semantic relationships, a semantic dependency graph is constructed. The nodes in the graph represent the words or phrases in the text, and the edges represent the semantic dependency relationships between the words or phrases. For example, in the conversation where the user mentions "I often travel on business and am worried about accidental risks", risk entities such as "travel on business" and "accidental risks" are identified. At the same time, through semantic analysis, the verb-object relationship between "travel on business" and "worried about" and the object relationship between "accidental risks" and "worried about" are determined, and are represented by corresponding nodes and edges in the semantic dependency graph, thus structuring the text information input by the user for subsequent semantic understanding and analysis.

[0069] When designing a double-buffer storage mechanism, the system creates two buffers with different functions. The first buffer is used to store the raw data of the current conversation turn, including the user's input information (voice or text), input time, input device, and other relevant metadata. When the user makes a new input, the first buffer immediately receives and stores this data, providing data support for the real-time processing of the current conversation turn. For example, it provides the raw input data for the natural language processing unit enhanced with domain knowledge, so as to extract risk entities and construct semantic dependency graphs in a timely manner. The second buffer uses a timestamp indexing mechanism to store the historical conversation context. Since the start of the conversation between the system and the user, every time a conversation turn is completed, the system will store the conversation content of that turn (including user input and system response) and the corresponding timestamp information in the second buffer. The timestamps accurately record the chronological order of the conversations. Based on this, the system can quickly index and retrieve the historical conversation context through timestamps during subsequent processing. When cross-turn context association needs to be implemented, for example, when the user's current input is related to the previous conversation, the system will, based on the historical conversation records stored in the second buffer and combined with the current input information, through semantic analysis and reasoning, understand the user's complete intention. For example, the user mentions "I have high blood pressure" in the first conversation turn and "Does this condition affect insurance application?" in a subsequent turn. By querying the historical conversation records stored in the second buffer and combining with the current question, the system can accurately understand that the user is asking about the impact of high blood pressure on insurance application, and thus generate a targeted response based on insurance domain knowledge and user risk attribute data, achieving effective cross-turn interaction and improving the user experience and the accuracy of information collection.

[0070] During the entire operation of the user interaction module, the user information obtained through the multimodal input method provides processing objects for the natural language processing unit enhanced with domain knowledge. The risk entities extracted by the natural language processing unit and the constructed semantic dependency graphs, in turn, provide more valuable content for the double-buffer storage mechanism. At the same time, the historical conversation context stored by the double-buffer storage mechanism provides data support for the natural language processing unit to achieve more accurate semantic understanding and cross-turn context association. In addition, the system also sets up an adaptive adjustment mechanism to optimize and adjust the end-to-end speech recognition model, entity recognition network, double-buffer storage strategy, etc., according to the characteristics of user input, the update of insurance domain knowledge, and the development of technology during the actual operation process, ensuring that the user interaction module always efficiently and accurately collects user risk attribute data.

[0071] Example 2:

[0072] Regarding the related technical solutions of the dynamic conversation policy engine, the specific implementation methods cover multiple parts such as the construction of a context-aware reinforcement learning model, hierarchical state space design, information gain evaluation, and asynchronous policy optimization training, as follows:

[0073] The dynamic dialogue policy engine operates based on a context-aware reinforcement learning model. This model views the entire conversation process as a sequential decision-making problem, where each response or question from the system represents a decision. The model first obtains the real-time conversation state through the user interaction module, including the user's current text and voice input, as well as historical conversation context recorded in a double-buffered storage mechanism. To process this information, a natural language processing unit, enhanced with insurance domain knowledge, deeply analyzes the user input, extracting risk entities and semantic relationships. For example, when a user mentions "I'm a high-altitude worker and want to learn about insurance," the natural language processing unit identifies the risk entity "high-altitude worker" and, based on historical conversations, determines the user's insurance intention. This information is then integrated into the current conversation state. The model uses pre-set policy templates, which include a library of common risk information collection questions, a library of standard response scripts, and a logical process to guide users in completing their information. These templates, such as risk inquiry templates for different occupations and FAQ templates, filter the pre-set policy templates based on the real-time conversation state to select a subset of potentially applicable policies. Next, the strategy is dynamically optimized using a reinforcement learning algorithm, aiming to maximize long-term cumulative rewards. Through continuous trial-and-error learning, the optimal interaction sequence for the current conversational state is determined. For example, after a user first mentions their occupation, the model uses reinforcement learning strategies to select appropriate follow-up questions, such as inquiries about years of work experience and whether they have experienced work-related injuries. This constructs a dynamic interaction sequence that includes risk information collection questions and feedback response logic.

[0074] For the specific implementation of the dynamic dialogue strategy engine, first, a hierarchical state space model is constructed. The upper-level state focuses on representing the completeness of user risk information, which is determined by quantitatively analyzing the risk entities extracted during the user interaction process. The system pre-defines key risk dimensions in the insurance field, such as age, occupation, health status, family income, pre-existing medical history, etc., and sets information completeness scoring rules for each dimension. For example, in the age dimension, if only a general age range is obtained, a lower completeness score is recorded; if the specific age is accurate, a higher score is given. As the dialogue progresses, whenever new risk entity information is obtained, the scores of each dimension are updated according to the scoring rules, and the overall user risk information completeness is calculated by summarization. The lower-level state is used to represent the confidence level of the dialogue intention, which is achieved by identifying the intention of the user input text. An intention recognition model based on deep learning, such as a sequence classification model based on Transformer, is trained on a large amount of labeled insurance dialogue corpus and can identify multiple intentions in the user input, such as consulting products, asking about the claims process, comparing rates, etc. The model outputs the confidence probability of each intention, and the intention with the highest probability is used as the main intention of the current dialogue, and its confidence level is the confidence level of the dialogue intention in the lower-level state. Through the hierarchical state space, the system can depict the dialogue state in more detail and provide rich information for subsequent decisions.

[0075] Then, an information gain evaluation function is designed to dynamically adjust the priority of question generation based on the change in information entropy. Information entropy is used to measure the uncertainty of user risk information. At the initial stage of the dialogue, due to less understanding of the user, the information entropy is at a relatively high level. As the dialogue progresses, after each question is asked and the user's answer is obtained, the system recalculates the information entropy. The information gain evaluation function quantifies the value brought by the question for obtaining user risk information by calculating the difference in information entropy before and after the question. For example, in a dialogue scenario where the user's occupation information is not yet known, asking "What is your occupation?" may significantly reduce the information entropy, and the information gain of this question is relatively high; while asking a question that repeats known information, the information gain approaches zero. Based on the size of the information gain, the system sorts all the candidate questions in the question library by priority, and preferentially selects questions with high information gain to ask the user, ensuring that each interaction can efficiently collect valuable risk information, optimize the dialogue process, and avoid redundant questions.

[0076] When training a reinforcement learning model, an asynchronous policy optimization algorithm is adopted, and an offline experience replay pool is integrated with online real-time interaction data. The offline experience replay pool stores a large amount of interaction experiences accumulated in historical conversations. Each experience sample contains information such as the conversation state, the executed action, the obtained reward, and the subsequent state. The online real-time interaction data comes from the immediate conversation between the system and the current user. In each policy iteration cycle, the algorithm randomly extracts a certain number of samples from the offline experience replay pool and merges them with the online real-time interaction data to form a training dataset. This method can not only utilize historical experiences to avoid overfitting of the model to new data but also learn the current user's interaction pattern in a timely manner and quickly adapt to different user needs. During the training process, based on the selected training data, the action value function (such as the Q function) is calculated to evaluate the advantages and disadvantages of each action in the corresponding state, and optimization methods such as gradient descent are used to update the model parameters and adjust the policy network, enabling the model to gradually learn to select the optimal action in various conversation states, generate more reasonable dynamic interaction sequences, improve the quality and efficiency of the dialogue interaction with users, and provide users with more accurate and effective risk information collection and insurance purchase advice guidance.

[0077] During the entire operation of the dynamic dialogue policy engine, the hierarchical state space model provides an accurate state representation for information gain evaluation and policy optimization; the information gain evaluation guides question generation and improves the information collection efficiency of the dialogue interaction; the asynchronous policy optimization training continuously enhances the model's decision-making ability based on actual interaction data.

[0078] Embodiment 3:

[0079] The specific implementation manner of the data synchronization method of the insurance product database is as follows:

[0080] In the link of docking with the insurance company's data interface, the system uses a protocol parsing component to achieve unified conversion of data formats. In actual application scenarios, the data interfaces of different insurance companies vary significantly due to different technical architectures and development standards, and the data formats they provide are also diverse, covering common forms such as XML and JSON, and the data fields are different in naming specifications and organizational levels. A complete and extensible parsing rule library is pre-constructed inside the protocol parsing component, and this rule library is based on in-depth analysis and summary of various data format protocols. When the system receives data from the insurance company, the protocol parsing component will automatically identify the format type of the data. For example, for XML format data, the component reads the label structure of the XML document through an XML parser and fills the specific data content contained in the label into the corresponding attribute matrix cells according to the attribute rules predefined by the system. During this conversion process, the protocol parsing component strictly follows the established data mapping relationship, performs field-by-field conversion and verification of the data, and ensures that the formed attribute matrix can accurately and completely reflect the original data information.

[0081] An incremental web crawler framework is built to comprehensively and accurately capture the details of insurance products on the web page, and the system introduces dynamic rendering technology. Currently, many insurance product web pages widely use scripting languages such as JavaScript to implement the dynamic loading of page content. If traditional static crawlers are used, only the initial static code of the web page can be obtained, and the complete product information after dynamic loading by the script cannot be obtained. The dynamic rendering technology in the system simulates the running environment of a real browser, creates a virtual environment with complete JavaScript execution capabilities, loads the insurance product web page in this environment, and runs various script codes step by step according to the execution logic of the web page script. After waiting for all the content on the page, including dynamically generated text, pictures, links, etc., to be fully presented, the data scraping operation is then performed. When the web page data is successfully captured, the rule engine immediately intervenes to extract key data. First, the rule engine converts the obtained insurance clause text into a dependency syntax graph through natural language processing technology. In this graph structure, each node corresponds to a statement in the clause, and the edges between the nodes intuitively represent the logical association relationships existing between the statements, such as causal relationships, parallel relationships, progressive relationships, etc., thereby converting the unstructured text information into structured graph data for convenient subsequent in-depth processing. Then, the rule engine defines a multi-hop graph attention layer based on graph neural network technology. This layer can accurately extract the constraint conditions in the clause by iteratively aggregating the features of adjacent nodes along the node paths in the graph structure. Taking the processing of clauses related to insurance claim conditions as an example, the multi-hop graph attention layer will follow the node paths related to claims in the graph structure, starting from the starting node, gradually aggregating the feature information of the surrounding nodes, and after multiple rounds of iterative calculations, finally accurately refine the specific constraint conditions for claims, including the time limit for claim applications, the supporting materials required to be submitted, the calculation method of claim amounts, etc. Finally, the rule engine constructs a condition trigger rule library according to the general standards and business requirements of the insurance industry, and performs fuzzy matching between the extracted clause constraint conditions and the preset trigger patterns. Even if the expressions of similar clauses by different insurance companies are different, through the fuzzy matching algorithm, they can be classified under the unified standardized product attribute labels, thus realizing the standardized processing of the web page scraped data.

[0082] To ensure the accuracy and consistency of data in the insurance product database, the system has designed a version conflict detection mechanism. During the actual data synchronization process, since insurance companies will update and iterate insurance products according to factors such as market demand, this may lead to multiple versions of data for the same insurance product at different times. The system uses a combination of timestamps and digital signatures to automatically screen the latest version of data. When each data is generated or updated, the system assigns it a timestamp accurate to the millisecond level to record the time attribute of the data; at the same time, using digital signature technology, based on a specific encryption algorithm, the data content is processed to generate a unique digital signature, which can verify the integrity of the data and the authenticity of the source. When the system detects multiple versions of data for the same insurance product, it first compares the versions of data based on the timestamps and preferentially selects the version with a newer timestamp as the candidate data; if the timestamps are the same, the digital signature is verified to determine whether the data is complete and has not been tampered with during transmission or storage, and only the data version that passes the digital signature verification will be finally determined to be adopted, so as to ensure that the insurance product database always stores the latest and most accurate product data, providing a solid and reliable information support for the intelligent decision-making module of the core system for intelligent insurance purchase decisions.

[0083] In the entire data synchronization process, the system also sets up a real-time monitoring module to continuously monitor various indicators and states during the data synchronization process. Once it detects changes in the data interface protocol of the insurance company or adjustments in the web page structure, etc., the system can quickly sense and trigger the adaptive adjustment mechanism to make corresponding modifications and optimizations to the rules of the protocol parsing component, the crawling strategy of the web crawler, and the processing logic of the rule engine, etc., to ensure the stability and accuracy of the data synchronization process and output accurate and effective personalized insurance purchase suggestions for users.

[0084] Example 4:

[0085] This example details the specific process of the hierarchical state space model using the attention gating mechanism to update the state. Define the dialogue state update rule and perform a cosine similarity match between the keyword vector of the user's answer and the preset risk dimensions. Let the keyword vector of the user's answer be , and the set of preset risk dimension vectors be . For the th preset risk dimension vector , the cosine similarity calculation formula is: , where represents the cosine similarity between the keyword vector of the user's answer and the th preset risk dimension vector. Through this similarity calculation, the degree of association between the user's answer and each preset risk dimension can be judged.

[0086] Construct a gated recurrent unit network to fuse the similarity score with the historical state vector and generate the current dialogue state encoding. Let the gated recurrent unit network be , the historical state vector be , and the similarity score vector be (composed of multiple calculated previously), then the current dialogue state encoding . The gated recurrent unit network can effectively process sequential data, combine the historical state and the current similarity score, generate a more accurate current dialogue state encoding, and better reflect the progress of the dialogue and the changes in user risk information.

[0087] Calculate the correlation strength between different risk dimensions through the multi-head self-attention mechanism and dynamically adjust the state transition probability. Let the multi-head self-attention mechanism be , and the input risk dimension feature matrix be , then the correlation strength matrix between different risk dimensions . According to this correlation strength matrix , the state transition probability can be dynamically adjusted, enabling the dialogue strategy to be more flexible in adjusting according to the relationship between different risk dimensions. For example, when it is found that the health risk dimension and the occupational risk dimension are closely related, relevant information can be more targeted in subsequent conversations.

[0088] Example 5:

[0089] In the process of the rule engine using the semantic graph convolutional network to extract key fields, the specific implementation method is as follows:

[0090] First, convert the insurance clause text into a dependency syntax graph. Insurance clauses are usually presented in the form of natural language text, which contains a large amount of complex information and diverse expression methods. The system uses natural language processing technology to perform syntactic analysis on the input insurance clause text. Specifically, the text is split into individual words through a word segmentation algorithm, and then a part-of-speech tagging tool is used to tag the part of speech of each word, such as nouns, verbs, adjectives, etc. On this basis, a dependency syntax analyzer is used to construct a dependency syntax graph according to the grammatical relationships between words. In the structure of this graph, each node corresponds to a sentence or a key phrase in the clause, and the node stores information such as the word sequence and part of speech of the sentence or phrase; the edges between the nodes represent the logical association relationships between the sentences or phrases, such as subject-predicate relationships, verb-object relationships, and modification relationships. Taking the sentence "If the insured person needs surgical treatment due to a serious illness during the insurance period, he / she can apply for compensation" in a health insurance clause as an example, in the dependency syntax graph, key contents such as "insured person", "insurance period", "serious illness", "surgical treatment", and "apply for compensation" will form nodes respectively. "Insured person" is connected to "apply for compensation" by an edge of the subject-predicate relationship, and "due to a serious illness" is connected to "apply for compensation" by an edge of the causal relationship, thus clearly presenting the logical context between the clause sentences and converting the unstructured insurance clause text into structured graph data.

[0091] Define a multi-hop graph attention layer to extract clause constraints. The multi-hop graph attention layer is based on the graph neural network architecture. Its core objective is to deeply mine the constraints in the clause by aggregating adjacent node features through multiple iterations along the node paths of the dependency syntax graph. During actual operation, first, initialize the feature vectors for each node. The feature vector contains the word vector information, part-of-speech information, and position information in the clause text of the node corresponding statement, etc. Then, according to the defined node feature propagation rule, starting from the starting node, in each hop of the calculation process, calculate the attention weight between the current node and adjacent nodes through the attention mechanism. This weight reflects the importance of the adjacent node features to the current node. Specifically, use the dot product operation between node feature vectors or other similarity calculation methods, combined with a learnable parameter matrix, to obtain the attention score, and then convert the attention score into a probability-form attention weight through a normalization function (such as the Softmax function). Based on the calculated attention weight, weightedly aggregate the features of adjacent nodes, and add the aggregated features to the original features of the current node to form new node features. This process is a feature propagation. By setting the number of multi-hops and continuously repeating the above feature propagation process, the node features can gradually spread and fuse in the graph structure, thereby capturing the logical relationships and semantic information between nodes that are farther apart. For example, when processing clauses regarding claim settlement conditions, the multi-hop graph attention layer can gradually aggregate the feature information of constraints such as claim settlement time limits, claim settlement material requirements, and claim settlement amount calculation methods scattered in different statements to key nodes through multiple feature propagations, and finally extract complete and accurate clause constraints.

[0092] Finally, construct a conditional trigger rule base, and perform fuzzy matching between the extracted constraint conditions and the preset trigger patterns to generate standardized product attribute labels. The conditional trigger rule base is constructed based on the professional knowledge and business experience of the insurance industry. A variety of common trigger patterns for insurance clause constraint conditions are predefined in it, and these patterns are stored in a structured form, including information such as the key features, logical relationships, and matching thresholds of the trigger conditions. When the rule engine extracts the constraint conditions of the insurance clause through the multi-hop graph attention layer, these constraint conditions are transformed into the form of feature vectors. Then, the constraint condition feature vectors are matched one by one with the preset trigger patterns in the conditional trigger rule base. Due to the diversity of insurance clause expressions, a fuzzy matching algorithm is adopted in the matching process, such as a matching method based on cosine similarity or edit distance. In the matching based on cosine similarity, the cosine value between the constraint condition feature vector and the preset trigger pattern feature vector is calculated. The closer the cosine value is to 1, the higher the similarity between the two; if the similarity reaches the preset threshold, it is considered that the constraint condition matches the corresponding preset trigger pattern. Once the matching is successful, the system will assign corresponding labels, such as "claim conditions", "coverage", "rate rules", etc., to the relevant content of the insurance clause according to the standardized product attribute labels corresponding to the preset trigger pattern, and further refine the label content, such as "major disease claim conditions", "accident coverage", etc. In this way, the extracted insurance clause constraint conditions are transformed into unified and standardized product attribute labels, providing standardized data support for the structured storage of the insurance product database and the subsequent analysis of the intelligent decision-making module, enabling the system to quickly and accurately understand and process the insurance product clauses of different insurance companies, and thus generating more accurate personalized insurance purchase suggestions for users.

[0093] When constructing the dependency syntax graph, the system will continuously optimize the parameters and rules of the natural language processing algorithm according to the characteristics of the insurance clauses actually processed to improve the accuracy of syntactic analysis; during the operation of the multi-hop graph attention layer, the parameters of the model are iteratively updated through the backpropagation algorithm, enabling the model to better adapt to the feature extraction requirements of different types of insurance clauses; for the conditional trigger rule base, the system also sets a dynamic update mechanism, and timely supplements and adjusts the preset trigger patterns according to the newly emerging insurance clause expression methods and changes in business requirements, ensuring that the rule engine always maintains an efficient and accurate keyword extraction ability, and guaranteeing the stable operation and good performance of the core system of intelligent insurance decision-making.

[0094] Example 6:

[0095] In the process of realizing the calculation of the confidence weights of the outputs of each model according to the user risk attribute distribution to generate a weighted decision vector, and the intelligent decision-making module performing matching analysis and generating personalized insurance purchase suggestions, the specific implementation methods are as follows:

[0096] First, the dynamic weighted fusion layer calculates the confidence weights of the outputs of each model according to the dual-channel confidence evaluation mechanism and in combination with the user risk attribute distribution. For the random forest model, the first channel determines its stability index by calculating the out-of-bag error rate. When constructing the random forest model, the system divides the training data set into multiple subsets for constructing each decision tree. Approximately two-thirds of the data in each subset is used to train a single tree, and the remaining one-third of the data that has not participated in the training is the out-of-bag data. After each decision tree is constructed, the out-of-bag data is used for testing, and the proportion of samples with prediction errors is calculated. The average value of the out-of-bag error rates of all decision trees is the out-of-bag error rate of this random forest model. When the user risk attribute data is obtained, the system inputs it into the random forest model and at the same time refers to the pre-calculated out-of-bag error rate. If the feature distribution of the user risk attribute data is less different from the feature distribution of the training data and the out-of-bag error rate is low, it indicates that the model has high stability on this data; otherwise, the stability is low.

[0097] For the gradient boosting tree model, the second channel calculates its fitting degree index by analyzing the loss function convergence curve. During the training process of the gradient boosting tree model, the system continuously records the loss function values during the training process to form a loss function convergence curve. When the user risk attribute data is input, the system observes the relationship between the loss function value of the model when processing this data and the convergence curve. If the loss function value is close to the end of the convergence curve and has little fluctuation after multiple iterations, it indicates that the model has high fitting degree for this data; if the loss function value deviates from the convergence curve or has large fluctuations, the fitting degree is low.

[0098] After obtaining the stability index of the random forest model and the fitting degree index of the gradient boosting tree model, the system generates the weight distribution coefficients of each model through a fuzzy logic inference system. The fuzzy logic inference system has predefined multiple fuzzy sets and inference rules. For example, the out-of-bag error rate is divided into fuzzy sets such as "low", "medium", and "high", and the loss function fitting degree is divided into fuzzy sets such as "excellent", "good", and "poor". According to the fuzzy sets in which the out-of-bag error rate of the random forest model and the fitting degree of the gradient boosting tree model are located, inferences are made according to the inference rules. For example, when the stability of the random forest model is "high" and the fitting degree of the gradient boosting tree model is "excellent", it is inferred that the weight coefficient of the random forest model is 0.4 and the weight coefficient of the gradient boosting tree model is 0.6; in this way, the weights of the two models in the decision-making are dynamically determined according to the user risk attribute distribution, and then a weighted decision vector is generated.

[0099] After generating the weighted decision vector, the intelligent decision-making module conducts a matching analysis between the user's risk attributes and the insurance product data to generate personalized insurance purchase recommendations. The multi-source structured insurance product data stored in the insurance product database has had its key fields extracted through a rule engine and standardized product attribute tags generated, covering key information such as the scope of coverage, rate rules, and claim settlement conditions. After the user risk attribute data is collected by the user interaction module, risk entity extraction and semantic analysis have been completed.

[0100] The intelligent decision-making module matches the user risk characteristics represented by the weighted decision vector with each insurance product in the insurance product database. Specifically, for the scope of coverage attribute of the insurance product, the system compares the risk types in the user risk attributes with the risk types covered by the scope of coverage. If the user has risk attributes related to high-risk occupations, the system retrieves products in the insurance products whose scope of coverage includes high-risk occupation protection; for the rate rule attribute, factors such as the user's age, health status, and income in the user risk attributes are matched and calculated with the parameters required for rate calculation to simulate and calculate the possible rates for the user under different insurance products; for the claim settlement condition attribute, the situations in the user risk attributes that may involve claims are compared with the claim settlement conditions of the insurance product to determine under what circumstances the user can obtain claim settlement.

[0101] After completing the matching analysis of all insurance products and user risk attributes, the intelligent decision-making module sorts the matching results according to preset screening rules and priorities. For example, it preferentially selects products whose scope of coverage fully covers the user's main risks; in the case of the same scope of coverage, it preferentially selects products with lower rates; and factors such as the leniency of the claim settlement conditions are also considered. Finally, based on the sorting results, several insurance products that best meet the user's needs are selected, and combined with the detailed information of the products, personalized insurance purchase recommendations are generated and fed back to the user through the user interaction module. Throughout the process, the intelligent decision-making module also continuously monitors the updates of the insurance product database and changes in the user risk attributes. If new insurance products are added to the database or the user supplements new risk attribute information, it will promptly re-conduct the matching analysis and adjust the insurance purchase recommendations to ensure that the insurance purchase recommendations provided to the user always fit their actual needs.

[0102] Example 7:

[0103] The specific implementation method of the memory retrieval module includes:

[0104] First, an external memory matrix is constructed, which is used to store the high-risk case features in historical conversations. When processing historical conversations, the system extracts key risk entities and features from the conversation content through a domain-knowledge-enhanced natural language processing unit. For each high-risk case, its features are converted into a vector representation and organized and stored in the external memory matrix according to a preset dimensional structure. For example, for a high-risk user case with a serious past medical history, features such as the type of medical history, treatment situation, and complications are extracted, these features are quantified into numerical values, and mapped to a low-dimensional vector space through an embedding algorithm to form the feature vector of this case. Each row in the matrix corresponds to the feature vector of a historical high-risk case, and metadata such as timestamps and risk levels are associated with each vector for subsequent retrieval and management.

[0105] After receiving the current user attribute vector, the memory retrieval module first preprocesses the vector. The user attribute vector is obtained by collecting user risk attribute data through the user interaction module and undergoing feature extraction and vectorization conversion. The preprocessing stage includes vector normalization operations to eliminate differences in dimension and numerical range among different feature dimensions, ensuring fair contributions of each dimension feature to similarity calculation. At the same time, the vector is dimension-reduced to remove noise and redundant information, retain the main feature information, and improve retrieval efficiency and accuracy.

[0106] When performing similarity retrieval, the memory retrieval module adopts a multi-stage retrieval strategy. In the first stage, coarse-grained retrieval is carried out, and the locality-sensitive hashing (LSH) algorithm is used to quickly screen out a subset of candidate historical cases that may be similar to the current user attribute vector. The LSH algorithm maps similar vectors to the same hash bucket by designing specific hash functions, thereby quickly narrowing the retrieval range in large-scale data. The system pre-hashes all vectors in the external memory matrix to construct a hash index. When receiving the current user attribute vector, it is mapped to the corresponding hash bucket through the same hash function, and all vectors in this hash bucket and adjacent hash buckets are extracted as the candidate set.

[0107] In the second stage, fine-grained retrieval is carried out, and the exact similarity between each historical case vector in the candidate set and the current user attribute vector is calculated. The similarity calculation uses the cosine similarity method, and the similarity between the two vectors is measured by calculating the cosine value of the included angle between the two vectors. At the same time, the system also considers the importance differences of vectors in different risk dimensions and introduces a dimension weight adjustment mechanism. For key risk dimensions in the insurance field, such as occupational risk and health risk, higher weights are assigned; for secondary dimensions, lower weights are assigned. The specific weight values are determined based on insurance domain expert knowledge and historical data statistical analysis. Through weighted cosine similarity calculation, the similarity scores between the current user attribute vector and each candidate historical case vector are obtained.

[0108] To further optimize the retrieval results, the memory retrieval module introduces a time decay factor. Considering that the risk characteristics of the insurance industry may change over time, recent historical cases have a higher reference value for current decision-making. The system multiplies the similarity score of each historical case vector by the time decay factor, which is calculated based on the timestamp of the case. The older the case, the smaller the decay factor.

[0109] After obtaining the similarity scores of all candidate historical cases, the memory retrieval module sorts them in descending order of score and selects the top K historical cases with the highest similarity. Then, a memory enhancement vector is generated based on the feature vectors of these cases. The generation process uses the attention-weighted aggregation method. For each selected historical case, the weight of its feature vector is determined by the similarity score. The higher the similarity, the greater the weight. By weighted summation, the feature vectors of the K historical cases are aggregated into a vector, which is the memory enhancement vector. This vector synthesizes the feature information of the historical high-risk cases most similar to the current user and provides additional context knowledge for the subsequent gated recurrent unit network.

[0110] The generated memory enhancement vector is concatenated with the hidden state of the gated recurrent unit. When processing the current dialogue state, the gated recurrent unit generates a hidden state vector that contains the short-term context information of the current dialogue. The memory enhancement vector is concatenated with this hidden state vector along the dimension to form a longer vector. Then, a linear transformation and non-linear activation are performed on the concatenated vector through a fully connected layer to generate the final state encoding. The weight parameters of the fully connected layer are optimized by the backpropagation algorithm during the model training phase, so that the final state encoding can fully integrate the historical high-risk case information and the current dialogue context information, providing a richer and more accurate representation for subsequent dialogue state prediction and decision-making.

[0111] Embodiment 8:

[0112] The multi-hop graph attention layer adopts a residual diffusion strategy, and the specific implementation is as follows:

[0113] When defining the node feature propagation rules, it is first necessary to convert the insurance clause text into a dependency syntax graph. Each node in the graph corresponds to a statement or phrase in the clause, and the node internally stores feature information such as the word vector, part of speech, and semantic category of the statement or phrase. The edges between the nodes represent the logical association relationships between the statements or phrases, such as causal, parallel, and modification relationships. During the operation of the multi-hop graph attention layer, each feature propagation (i.e., each hop) follows specific rules. When performing the first-hop calculation, for the target node, the system calculates the attention weights between it and the adjacent nodes based on the attention mechanism. Specifically, the feature vector of the target node and the feature vectors of the adjacent nodes are combined with a learnable parameter matrix through dot product or other similarity calculation methods to obtain attention scores, and then the attention scores are converted into probability-form attention weights through a normalization function. Based on this weight, the features of the adjacent nodes are weighted and aggregated to obtain a preliminary aggregated feature. At this time, the preliminary aggregated feature is superimposed on the original feature of the target node to form a new node feature, and the superimposed original node feature is the residual term. In this way, the information of the original node is retained during the feature propagation process, avoiding the loss of feature information caused by multiple aggregations. As the number of hops increases, the above process is repeated for each subsequent hop, performing a new round of attention calculation and feature aggregation between the node features updated in the previous hop and the adjacent nodes, and superimposing the original node features again to continuously update the node features, enabling the node to gradually integrate information from nodes at greater distances, thereby extracting more comprehensive clause constraints.

[0114] When designing the diffusion decay coefficient to adjust the attention weights, the system sets corresponding decay rules according to the logical hop count between the nodes. The logical hop count represents the number of the shortest paths connecting two nodes in the dependency syntax graph. To simulate the characteristic that the influence of logical relationships weakens with increasing distance in reality, the system determines the attenuation degree of the attention weights in an exponentially decreasing manner. In actual operation, after calculating the attention weights between a certain node and its adjacent nodes, according to the logical hop count between them, the attenuation coefficient is applied to adjust the attention weights. For example, if the logical hop count between two nodes is 1, indicating that they are directly connected, the attention weights are not attenuated or only slightly attenuated at this time; if the logical hop count is 2, the attention weights are reduced by a corresponding multiple according to the pre-set attenuation formula; as the logical hop count increases, the attenuation multiple grows exponentially, making the influence of distant nodes on the feature aggregation of the target node gradually decrease. This mechanism ensures that in the feature propagation process, the information of nearby nodes can be more fully utilized, while avoiding the over-influence of distant nodes on the target node features due to their relatively distant logical relationships, thereby more reasonably extracting the constraints with close logical associations in the clause and improving the accuracy and effectiveness of feature extraction.

[0115] During multiple feature jumps, due to continuous feature aggregation and calculation, the scale of node features is prone to shift, resulting in difficulties in subsequent processing. To solve this problem, the system introduces a feature normalization layer. When node features are updated through multiple propagations in the multi-hop graph attention layer, they enter the feature normalization layer for processing. The feature normalization layer adopts common normalization methods, such as Batch Normalization or Layer Normalization. Taking Batch Normalization as an example, in a batch of node feature data, the mean and variance of each feature dimension are calculated, and then the feature vectors of each node are normalized in each dimension, converting them into a standard distribution with a mean of 0 and a variance of 1. Through this normalization process, the problem of feature scale shift caused by multiple jumps is eliminated, enabling features of different nodes to be compared and processed on the same scale, ensuring the stability and consistency of subsequent model calculations, and helping to improve the accuracy and reliability of the model in extracting insurance clause constraint conditions.

[0116] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0117] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent decision-making core system for insurance application based on a dynamic dialogue strategy, characterized in that It includes: User Interaction Module: Used to conduct dialogue interaction with users through multi-modal input methods and collect user risk attribute data in real time; Insurance Product Database: Stores structured data of multi-source insurance products, including coverage scope, rate rules, and claim settlement conditions; Intelligent Decision-making Module: Adopts multi-model fusion technology to match and analyze user risk attributes and insurance product data, and generates personalized insurance purchase suggestions; The multi-model fusion technology of the Intelligent Decision-making Module includes: Construct a random forest model, screen high-risk correlation factors through feature importance ranking, and generate a preliminary risk rating; Construct a gradient boosting tree model, optimize the risk score granularity based on residual fitting, and output a fine-grained risk probability distribution; Design a dynamic weighted fusion layer, calculate the confidence weights of the outputs of each model according to the user risk attribute distribution, and generate a weighted decision vector; Match the user risk characteristics represented by the weighted decision vector with each insurance product in the insurance product database; The Dynamic Dialogue Policy Engine includes: Construct a hierarchical state space model, where the upper state represents the integrity of user risk information, and the lower state represents the confidence of dialogue intentions; Design an information gain evaluation function to dynamically adjust the problem generation priority based on the change in information entropy; Adopt an asynchronous policy optimization algorithm to train a reinforcement learning model, and each policy iteration cycle integrates an offline experience replay pool and online real-time interaction data; The hierarchical state space model updates the state using an attention gating mechanism, including: Define the dialogue state update rule, and perform cosine similarity matching between the keyword vector of the user's answer and the preset risk dimension; Construct a gated recurrent unit network, fuse the similarity score with the historical state vector, and generate the current dialogue state encoding; Calculate the correlation strength between different risk dimensions through a multi-head self-attention mechanism, and dynamically adjust the state transition probability.

2. The core system for intelligent insurance decision-making according to claim 1, characterized in that, The User Interaction Module includes: The multi-modal input method includes a voice input channel and a text input channel. The voice input channel uses an end-to-end speech recognition model to generate a text intermediate representation; Construct a natural language processing unit enhanced by domain knowledge, extract risk entities in the user input through an entity recognition network, and construct a semantic dependency graph in combination with historical dialogue records; Design a double-buffer storage mechanism. The first buffer stores the original data of the current dialogue turn, and the second buffer stores the historical dialogue context using a timestamp indexing mechanism to achieve cross-turn context association.

3. The core system for intelligent insurance decision-making according to claim 1, wherein The system also includes: Dynamic Dialogue Policy Engine: Based on a context-aware reinforcement learning model, combines a preset policy template and the real-time dialogue state to generate a dynamic interaction sequence, and the dynamic interaction sequence includes risk information collection questions and feedback response logic.

4. The core system for intelligent insurance decision-making according to claim 1, characterized in that, The data synchronization method of the Insurance Product Database includes: Connect to the insurance company's data interface, and convert the heterogeneous data format into a unified attribute matrix through a protocol parsing component; Construct an incremental network crawler framework, use dynamic rendering technology to crawl the details of insurance products on the web page, and extract the key fields in the product terms through a rule engine; Design a version conflict detection mechanism. When detecting multi-version data of the same product, automatically select the latest version based on the timestamp and digital signature.

5. The core system for intelligent insurance decision-making according to claim 4, characterized in that, The rule engine uses a semantic graph convolutional network to extract key fields, including: Convert the insurance clause text into a dependency syntax graph, where the nodes in the graph represent clause statements and the edges represent logical association relationships; Define a multi-hop graph attention layer to extract clause constraint conditions by iteratively aggregating adjacent node features; Construct a conditional trigger rule library, and perform fuzzy matching on the extracted constraint conditions and preset trigger patterns to generate standardized product attribute labels.

6. The core system for intelligent insurance decision-making according to claim 1, wherein The dynamic weighted fusion layer uses dual-channel confidence evaluation, including: The first channel calculates the out-of-bag error rate of the random forest model as an indicator of model stability; The second channel analyzes the loss function convergence curve of the gradient boosting tree model and calculates the model fitting degree indicator; Generate the weight allocation coefficients of each model through a fuzzy logic inference system by synthesizing the stability and fitting degree indicators.

7. The core system for intelligent insurance decision-making according to claim 1, characterized in that, The gated recurrent unit network introduces a memory enhancement mechanism, including: Construct an external memory matrix to store the high-risk case features in the historical conversation; Design a memory retrieval module to perform similarity retrieval based on the current user attribute vector and the memory matrix to generate a memory enhancement vector; Concatenate the memory enhancement vector with the hidden state of the gated recurrent unit and generate the final state encoding through a fully connected layer.

8. The core system for intelligent insurance decision-making according to claim 5, wherein The multi-hop graph attention layer adopts a residual diffusion strategy, including: Define the node feature propagation rule, and stack the original node features as the residual term after aggregating the features of each hop; Design a diffusion attenuation coefficient to exponentially reduce the attention weight according to the logical hop count between nodes; Eliminate the feature scale offset caused by multiple jumps through a feature normalization layer.

Images