Insurance APP renewal method and system based on dynamic user portrait

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an insurance APP renewal method and system based on a dynamic user portrait. The method comprises the following steps: acquiring a user identifier and loading a dynamic user portrait and a cross-session task state record of the user identifier; receiving current dialogue interaction data, performing context awareness analysis in combination with the dynamic user portrait and the cross-dialogue task state record, and identifying a current insurance renewal intention and an evolution state thereof relative to a historical intention; and generating an insurance renewal strategy of the current dialogue round according to the evolution state, the user portrait and the task record. According to the method, the real-time user state and the task process context are fused, accurate tracking of the evolution path of the insurance renewal intention of the user is achieved, and the insurance renewal strategy dynamically aligned with the historical process and the current service logic is driven to be generated accordingly. According to the method, the randomness of strategy generation is effectively inhibited, and the context continuity of the insurance renewal dialogue, the service alignment precision and the certainty of process propulsion are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an insurance app renewal method and system based on dynamic user profiles. Background Technology

[0002] As the digital transformation of the insurance industry deepens, intelligent renewal dialogues in insurance applications have become a key element in improving operational efficiency and customer experience. Current technical solutions generally face a fundamental technical challenge in achieving this goal: the fragmentation of effective contextual information between multi-turn dialogues and the lack of maintenance of key business states result in highly random renewal strategies generated by the system, making it difficult to align with complex and predetermined business logic.

[0003] Specifically, user needs during the insurance renewal process do not arise in isolation but evolve dynamically through continuous interactions and are closely related to specific business process nodes (such as quoting, underwriting, and payment). To maintain dialogue continuity, existing mainstream solutions typically rely on two methods: one is to store the original text records of the most recent rounds of dialogue; the other is to build a set of static attribute tags for users. However, both methods have significant limitations: linear dialogue history is essentially an unstructured text sequence, making it difficult for the system to extract structured information effective for current decisions, such as the core user intent that was agreed upon in the previous dialogue, the specific operational instructions that the system has successfully executed, and the business bottlenecks (task status) that urgently need to be resolved in the current process. This directly leads to a substantial break in the business context between different dialogue rounds. At the same time, static user tags cannot perceive and respond to real-time life events of users (such as car purchases, changes in family structure), which are precisely the key factors driving changes in insurance demand and requiring differentiated responses through strategies. Summary of the Invention

[0004] Therefore, the purpose of this invention is to overcome the problems of high randomness of renewal strategies and difficulty in aligning business logic caused by context breaks and state loss during renewal dialogues in existing insurance apps. This invention provides an insurance app renewal method and system based on dynamic user profiles. By loading the user's dynamic user profile and cross-session task state records, and integrating real-time updated user state and task process context, it achieves accurate identification and tracking of the user's renewal intention evolution path. Based on this, it drives the generation of renewal strategies aligned with historical operation sequences and current business states, effectively suppressing the randomness of strategy generation and ensuring that each round of strategy generation can inherit historical processes and match current business logic, effectively improving the contextual coherence of renewal strategies, business alignment accuracy, and process advancement determinism.

[0005] In a first aspect, to solve the above-mentioned technical problems, the present invention provides an insurance APP renewal method based on dynamic user profiles, the method comprising the following steps: Obtain the user's user identifier, and load the user's dynamic user profile and cross-session task status record based on the user identifier; wherein, the dynamic user profile includes intent tags updated based on the user's historical data and user status change characteristics; the cross-session task status record includes historical renewal intents related to the renewal task, the sequence of operation instructions executed by the system, and the status of tasks to be processed; Receive the user's current dialogue interaction data, combine dynamic user profiles and cross-session task status records to perform context-aware analysis on the current dialogue interaction data, and identify the user's current renewal intention and its evolution relative to historical renewal intentions; Based on the evolution status, dynamic user profiles, and cross-session task status records, a renewal strategy for the current dialogue round is generated.

[0006] Preferably, the user state transition features are generated based on the user's historical data, including the following steps: Based on an insurance scenario life event knowledge base, entities and relationships are extracted from unstructured text information in user historical data. The extracted entities and relationships are matched with event triggering conditions in the insurance scenario life event knowledge base to obtain one or more potential life event tags and a confidence score corresponding to each potential life event tag. Combining the potential life event tags and their corresponding confidence scores, user state transition features are generated. The event triggering conditions in the insurance scenario life event knowledge base are configured to simultaneously satisfy the association and temporal relationship between multiple semantic elements.

[0007] Preferably, the insurance scenario life event knowledge base is coupled with the rule engine. The insurance scenario life event knowledge base is configured to store multiple event trigger rule sets, each event trigger rule set corresponding to an insurance-related life event type. The event trigger rule set is composed of production rules, and the antecedent of the production rule is configured as a composite condition based on the association and temporal relationship between multiple semantic elements.

[0008] Preferably, context-aware analysis of the current dialogue interaction data is performed by combining dynamic user profiles and cross-session task state records to identify the user's current renewal intention and its evolution relative to historical renewal intentions. This includes: performing a first intention recognition on the current dialogue interaction data based on a large language model to obtain a current intention label; constructing fused prompt information based on the current intention label, dynamic user profile, and cross-session task state records; wherein the fused prompt information includes instruction information used to constrain the inference direction of the large language model; inputting the fused prompt information into the large language model to guide it to perform context-aware inference; and parsing the current renewal intention and its evolution state based on the inference output of the large language model.

[0009] Preferably, based on the current intent tag, dynamic user profile, and cross-session task status records, a fused prompt message is constructed, including: selecting a matching prompt template from a preset renewal scenario prompt template library based on the current intent tag; filling the corresponding placeholders in the matching prompt template with user status change features from the dynamic user profile and historical renewal intents and pending task statuses from the cross-session task status records to generate an initial fused prompt; optimizing the initial fused prompt to obtain fused prompt information; wherein, the optimization includes information compression or format conversion.

[0010] Preferably, before inputting the fused prompt information into the large language model, the method further includes: determining the logical reasoning depth level required for this inference based on the current intent label and the user state change characteristics in the dynamic user profile; configuring the corresponding inference parameters based on the logical reasoning depth level; and adjusting the inference generation process of the large language model according to the inference parameters to guide it to perform context-aware inference at the logical reasoning depth level.

[0011] Preferably, the inference parameters include a temperature parameter. Based on the inference parameters, the inference generation process of the large language model is adjusted to guide it to perform context-aware inference at the logical reasoning depth level. This includes: constructing a parameterized inference request by combining the fused prompt information and the set value of the temperature parameter; sending the parameterized inference request to the large language model service interface; wherein the set value of the temperature parameter is used to control the randomness of probability sampling in the model generation process at the service interface layer; and receiving the inference result returned by the large language model service interface as the output of the context-aware inference.

[0012] Preferably, the renewal strategy for the current dialogue round is generated based on the evolution status, dynamic user profile, and cross-session task status records. This includes: determining the strategy decision basis for the current dialogue round based on the evolution status and the status of pending tasks; matching a basic strategy template from a preset renewal strategy library based on the strategy decision basis; adapting the user state transition features in the dynamic user profile to the basic strategy template to generate strategy configuration data for instantiating the basic strategy template; and instantiating the basic strategy template based on the strategy configuration data to generate the renewal strategy for the current dialogue round.

[0013] Preferably, determining the strategy decision basis for the current dialogue round based on the evolution state and the task state to be processed includes: combining the evolution state and the task state to be processed into a state pair; matching the state pair with a preset strategy decision basis mapping table; wherein the strategy decision basis mapping table predefines the correspondence between different state pairs and strategy decision basis; and obtaining the corresponding strategy decision basis according to the matching query result.

[0014] Secondly, to solve the above-mentioned technical problems, the present invention provides an insurance APP renewal system based on dynamic user profiles, comprising: The data loading module loads the user's dynamic user profile and cross-session task status records based on the user identifier. The dynamic user profile includes intent tags updated based on the user's historical data and user status change characteristics. The cross-session task status records include historical renewal intents related to the renewal task, the sequence of operation instructions executed by the system, and the status of tasks to be processed. The context-aware analysis module receives the user's current dialogue interaction data, combines dynamic user profiles and cross-session task status records to perform context-aware analysis on the current dialogue interaction data, and identifies the user's current renewal intention and its evolution relative to historical renewal intentions. The renewal strategy generation module generates the renewal strategy for the current conversation round based on the evolution status, dynamic user profile, and cross-session task status records.

[0015] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: The insurance APP renewal method and system based on dynamic user profiles of the present invention loads the user's dynamic user profile and cross-session task status records, integrates the real-time updated user status and task process context, realizes accurate identification and tracking of the user's renewal intention evolution path, and drives the generation of renewal strategies that are aligned with historical operation sequences and current business status. This effectively suppresses the randomness of strategy generation and ensures that each round of strategy generation can inherit the historical process and match the current business logic, effectively improving the contextual coherence of the renewal strategy, the accuracy of business alignment, and the determinism of process advancement. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0017] Figure 1 This is a flowchart of an insurance APP renewal method based on dynamic user profiles in a preferred embodiment of the present invention.

[0018] Figure 2 This is a flowchart illustrating the generation of user state transition features in a preferred embodiment of the present invention.

[0019] Figure 3 This is a flowchart for identifying the evolution state in a preferred embodiment of the present invention.

[0020] Figure 4 This is a flowchart illustrating the renewal strategy for the current dialogue round in a preferred embodiment of the present invention.

[0021] Figure 5 This is a structural block diagram of an insurance APP renewal system based on dynamic user profiles in a preferred embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0023] The purpose of this invention is to overcome the problems of high randomness in renewal strategies and difficulty in aligning business logic caused by context breaks and state missing during insurance APP renewal dialogues in the prior art.

[0024] Example 1

[0025] Reference Figure 1 As shown in the figure, this invention discloses an insurance APP renewal method based on dynamic user profiles, including: S100. Obtain the user's user identifier, and load the user's dynamic user profile and cross-session task status record based on the user identifier; wherein, the dynamic user profile includes intent tags and user status change features updated based on the user's historical data; the cross-session task status record includes historical renewal intents related to the renewal task, the sequence of operation instructions executed by the system, and the status of tasks to be processed; S200: Receive the user's current dialogue interaction data, and perform context-aware analysis on the current dialogue interaction data in conjunction with the dynamic user profile and the cross-session task status record to identify the user's current renewal intention and its evolution status relative to the historical renewal intention. S300. Based on the evolution status, the dynamic user profile, and the cross-session task status record, generate a renewal strategy for the current dialogue round.

[0026] The following combination Figure 1 Each step is explained in detail: In the specific implementation step S100, this step is completed in collaboration between the user session management module and the data persistence service: when a user initiates a renewal conversation in the insurance APP, the front-end interface module obtains and transmits the user's unique identifier User ID. This identifier is usually bound to the APP's login account system to ensure the accurate association between the session and the user.

[0027] The user profiling service retrieves and loads the user's dynamic user profile from the user profile database based on the received User ID. The dynamic user profile is a data structure that contains at least updated intent tags and user state transition features. Intent tags are one or more tags that represent a user's main insurance needs in the near future, such as "car insurance renewal consultation" or "health insurance product comparison". These intent tags are updated periodically or triggered by the backend intent recognition engine based on the user's historical dialogue, browsing clicks and other interaction data through a classification model.

[0028] User state transition features are structured feature vectors or lists used to characterize changes in a user's life or state that affect their insurance needs. Their generation relies on a separate state transition analysis module, which, based on a knowledge base in the insurance field, extracts entities and relationships from the user's recent unstructured interaction data (such as customer service records and self-reported information). Through rule-based or lightweight model matching, it outputs structured features such as {Event: Car Purchase, Confidence: 0.92, Time: 2023-10}. User state transition features and intent tags are stored separately, but together they constitute a dynamic user profile.

[0029] The task status management service loads cross-session task status records from the task status database based on the same User ID and the task identifier associated with the current dialogue channel and business type. A cross-session task status record is a structured document for a specific renewal task chain, which includes at least the historical renewal intent, the sequence of operation instructions executed by the system, and the pending task status. Historical renewal intentions are records in chronological order of the user intention sequence that the system has identified in the current renewal task chain, such as: [“Inquire about car insurance quotes”, “Confirm coverage”].

[0030] The sequence of operation instructions executed by the system records the backend services or atomic operations that the system actually calls in response to the user's historical intent; for example: ["Call the policy query interface", "Call the rate calculation engine", "Send the quotation message"].

[0031] The status of a pending task is a status code or label that clearly identifies the bottleneck in the current task flow or the expected next action; for example: "Pending user confirmation of quotation" or "Pending supplementary vehicle information".

[0032] In step S100, by using two dedicated data structures—dynamic user profiles and cross-session task state records—and persistently loading them, the system has a complete and structured historical context about the user and task at the beginning of each conversation. This addresses the issue of missing state at the source of the data, providing a consistent and reliable factual basis for all subsequent analyses and directly suppressing the randomness of decisions caused by inconsistent or one-sided data.

[0033] In the specific implementation step S200, this step is executed by the context-aware analysis engine as the core; the context-aware analysis engine receives the current dialogue interaction data from the dialogue front end, which is usually text that has been initially cleaned and segmented.

[0034] The context-aware analysis engine does not directly process the original text, but first constructs a fused context feature representation, which encodes the following three into a structured context representation: (1) the dynamic user profile loaded from step S100; (2) the cross-session task state record loaded from step S100; and (3) the current user input text. Specifically, a prompt string or feature vector is constructed, for example, explicitly containing: "User profile: [Intent label: car insurance consultation; State transition: recent car purchase]; Task state: [Historical intent: asking for a quote; Pending state: pending confirmation]; The user currently says: I want to see the last quote, how do I add the new car information?"

[0035] This fused contextual representation is fed into a trained intent evolution analysis model, such as a large language model based on a sequence-to-sequence architecture or specific cue engineering; the task of the intent evolution analysis model is not to perform simple intent classification, but to perform relational reasoning, and its output includes at least the current renewal intent and the evolution state: Current renewal intent is the determination of the core intent of the user's current input within the current context, such as applying for insurance for a new vehicle. Evolutionary state is a label representing the relationship between the current renewal intent and recent historical renewal intents. Its determination logic is based on business rules and model reasoning, specifically: If the current intent text is highly similar to the historical intent, and the pending state is a progressive action, then it is determined to be an intent continuation; If the current intent is different from the historical intent, but the user state transition characteristics can provide a reasonable explanation, and the new intent is logically related to the transition characteristics, then it is determined to be an intent derivation; If the intentions are different and cannot be explained by known state transitions, then it is determined to be an intention shift.

[0036] The core of step S200 is to upgrade static intent recognition to the recognition of the dynamic trajectory and motivation of intent changes. Through context-aware analysis, the current moment and historical moments are semantically connected in a meaningful way, rather than just arranged in time. This ensures that the analysis conclusions are aligned with the business history logic.

[0037] When implementing step S300, this step is executed by the policy generator; the policy generator receives the current renewal intention and evolution status output from step S200, and combines them with the dynamic user profile and cross-session task status record loaded in step S100.

[0038] The strategy generator internally maintains a strategy decision basis matrix and a set of strategy templates. First, based on the evolution status and the status of the tasks to be processed, it queries the strategy decision basis matrix to determine the strategy decision basis for this round, such as completing the information binding and price update of newly added vehicles. Then, based on the strategy decision basis, it selects a predefined basic strategy template. The basic strategy template is an action framework containing pluggable parameters, such as [Action 1: Collect information {information type}, Action 2: Call service {service name}, Action 3: Push message {message template}].

[0039] Next, the user state transition features in the dynamic user profile are adapted to the selected basic strategy template to generate specific strategy parameters. For example, if the transition feature is "purchase of a car (model: Tesla Model Y)," the {information type} parameter in the basic strategy template is instantiated as "new vehicle identification number (VIN)," and {service name} is associated with "new vehicle insurance matching service."

[0040] The generated strategy parameters are filled into the basic strategy template to form an executable, specific sequence of instructions, which is the renewal strategy for the current dialogue round. For example: [Execute: Information Collection (Field: VIN), Call: Underwriting Engine (Vehicle Model: Tesla Model Y), Execute: Message Push (Template: New Car Insurance Quote)]. This strategy will be delivered to the dialogue execution engine, which will then translate it into specific system calls and user responses.

[0041] The insurance APP renewal method based on dynamic user profiles of the present invention loads the user's dynamic user profile and cross-session task status records, integrates the real-time updated user status and task process context, to achieve accurate identification and tracking of the user's renewal intention evolution path, and drives the generation of renewal strategies that are aligned with historical operation sequences and current business status. This effectively suppresses the randomness of strategy generation and ensures that each round of strategy generation can inherit the historical process and match the current business logic, effectively improving the contextual coherence of the renewal strategy, the accuracy of business alignment, and the determinism of process advancement.

[0042] In the aforementioned embodiments, user state transition features are generated based on a knowledge base. However, user expressions are highly arbitrary, and the same life event (such as "buying a car") may have dozens of different expressions (such as "bought a new car", "just drove the car home", "ordered a Model Y", etc.). Traditional methods based on keyword dictionaries or simple templates have low recall rates and are difficult to distinguish whether an event has already occurred, is planned, or is only mentioned. Furthermore, effective event recognition often requires a combination of multiple semantic elements as evidence (for example, recognizing a "car purchase" event requires multiple elements such as "purchase action", "vehicle entity", and "completion tense"). How to reliably discover these scattered elements from massive amounts of text and determine that they constitute a complete event is a challenge.

[0043] To address the issues of low accuracy and high false positive rates in extracting insurance-related life events from unstructured user text due to a lack of semantic constraints and domain logic, this invention discloses a method for generating user state transition features based on an insurance scenario knowledge base and multi-constraint logic matching. Specifically, referring to... Figure 2 As shown, generating user state transition features based on user historical data includes: extracting entities and relationships from unstructured text information in the user historical data based on an insurance scenario life event knowledge base; matching the extracted entities and relationships with event triggering conditions in the insurance scenario life event knowledge base to obtain one or more potential life event tags and a confidence score corresponding to each potential life event tag; and generating user state transition features by combining the potential life event tags and their corresponding confidence scores. The event triggering conditions in the insurance scenario life event knowledge base are configured to simultaneously satisfy both association relationships and temporal relationships between multiple semantic elements.

[0044] Specifically, the step of generating user state transition features based on historical user data is executed by the user state transition analysis module, and the specific implementation is as follows: Step 1: Obtain unstructured text information such as user conversation logs and notes as input. First, use natural language processing tools to preprocess the text (word segmentation, part-of-speech tagging, dependency parsing); then, load the insurance scenario life event knowledge base, which predefines entity types and relation types related to the insurance domain. Based on the definitions in the insurance scenario life event knowledge base, perform the following operations: Entity extraction: Identify and classify named entities in text that belong to predefined types; Relation extraction: Identify and classify semantic relationships between extracted entities.

[0045] The second step: In the insurance scenario life event knowledge base, in addition to entity and relationship definitions, multiple event patterns are also stored; each event pattern represents a life event that is meaningful to insurance business, and each event pattern is defined by an event trigger condition; the event trigger condition is a structured logical expression that specifies the association and temporal relationship between multiple semantic elements that need to be satisfied to determine the occurrence of the event. This condition requires that entity combination, specific relationship, and temporal requirements of the action be satisfied at the same time.

[0046] The user state transition analysis module matches all entities and relationships extracted in the first step with the triggering conditions of each event pattern in the insurance scenario life event knowledge base. This matching process involves logical fulfillment checks and evidence weight calculations. For each event pattern, it checks the extent to which the evidence extracted from the current text satisfies its triggering conditions. The more mandatory conditions satisfied and the more direct the evidence, the higher the calculated confidence score. The confidence score can be calculated based on factors such as the proportion of rule satisfaction, the confidence level of key entities / relationships, and temporal consistency. Finally, a list is output containing all matched potential life event tags and their corresponding confidence scores.

[0047] Step 3: The user state transition analysis module processes the list output from Step 2 to generate the final user state transition features; the processing methods include, but are not limited to: Threshold filtering: Only retain potential life event labels with confidence scores exceeding a preset threshold to control feature quality and avoid low-confidence noise.

[0048] Structured coding: Encode the filtered "event label-confidence score" pairs into a standardized data structure.

[0049] Finally, the structured data is used as user state transition features and updated and stored in the user's dynamic user profile.

[0050] The solution of this invention, based on a knowledge base of life events in insurance scenarios and using a multi-constraint logic matching method, stably, accurately, and interpretably extracts structured user state transition features from users' random text, thereby achieving reliable and interpretable structured recognition of users' life events and providing high-quality, highly reliable dynamic user state input for the generation of subsequent insurance renewal strategies.

[0051] Specifically, the insurance scenario life event knowledge base is coupled with the rule engine. The insurance scenario life event knowledge base is configured to store multiple event trigger rule sets, each of which corresponds to an insurance-related life event type. The event trigger rule set is composed of production rules, and the antecedent of the production rule is configured as a composite condition based on the association and temporal relationship between multiple semantic elements.

[0052] In the embodiments of the present invention, the specific implementation method for coupling the insurance scenario life event knowledge base and the rule engine is as follows: Step 1: The insurance scenario life event knowledge base is not a simple database, but a knowledge base specifically designed for rule engines. Its essence is a structured rule storage and management system, containing the following core components: (1) Rule engine integration interface: The knowledge base provides a standard rule engine integration interface, which allows the rule engine to load the rule set stored therein; the rule engine can adopt open source components or a self-developed lightweight inference engine; (2) Event trigger rule set storage: Multiple event trigger rule sets are stored in the knowledge base. Each event trigger rule set is bound to an insurance-related life event type through a unique identifier; (3) Production rule composition: Each event triggering rule set consists of one or more production rules; the production rules adopt the standard logical form of "IF (precedence) THEN (consequence)"; (4) Rule consequent: used to declare the conclusion that should be output when the rule is triggered, and to include a confidence score; (5) Rule antecedent: configured as a composite condition: the composite condition is defined based on the association and temporal relationship between multiple semantic elements; in the syntax of the rule engine, this is represented by a combination of a series of fact pattern matching and logical operators.

[0053] Step 2: Utilize the rule engine to execute event-triggered rule sets for matching and inference. When event identification is required based on user historical data, the system executes the following process: The results of the upstream entity and relation extraction steps are transformed into fact objects recognizable by the rule engine and inserted into the rule engine's working memory. The rule engine then performs pattern matching between all fact objects in its working memory and the antecedents of production rules in all event-triggered rule sets loaded from the insurance scenario life event knowledge base, identifying all rules whose antecedent conditions are satisfied by the current fact set. For each triggered rule, the rule engine executes the action defined in its consequent. The conclusions output by all triggered rules constitute the set of potential life event labels and their confidence scores.

[0054] The present invention introduces an architecture that integrates a rule engine with generative rules to decouple and optimize domain knowledge representation and inference execution. This solution stores domain knowledge in a knowledge base in the form of declarative and configurable rules, supporting no-code dynamic adjustment and agile iteration of business rules. At the same time, the rule engine is used to efficiently compile and batch infer complex logic, significantly improving event recognition performance and execution stability.

[0055] In the above embodiments, context-aware analysis of current dialogue interaction data is performed by combining dynamic user profiles and cross-session task state records to identify the user's current renewal intention and its evolution relative to historical renewal intentions. The conventional approach is to train a classification model, but this requires a large amount of labeled data, and the model itself is a black box, its decision logic being unexplainable, making it difficult to implement in financial scenarios that emphasize compliance and interpretability. In the embodiments of this invention, a Large Language Model (LLM) is used for analysis. However, when using a large language model for context-aware analysis, there is a contradictory technical problem: a lack of specialized domain knowledge in the large language model and an overload of dialogue history information. Specifically: Large language models lack a deep understanding of the professional knowledge in the insurance renewal field, which may lead to factual errors or inferences that do not conform to business logic when directly performing intent recognition and evolution analysis. Simply feeding dynamic user profiles and cross-session task state records as context input to LLM will cause the model's attention to be scattered, the computational cost to skyrocket, and key information may be overwhelmed, thereby reducing the accuracy and reliability of the analysis.

[0056] To solve this problem, refer to Figure 3As shown in the embodiment of the present invention, the method of identifying the evolution state using a large language model specifically includes: performing a first intent recognition on the current dialogue interaction data based on the large language model to obtain a current intent label; constructing fused prompt information based on the current intent label, the dynamic user profile, and the cross-session task state record; wherein the fused prompt information includes instruction information for constraining the inference direction of the large language model; inputting the fused prompt information into the large language model to guide it to perform context-aware inference; and parsing the current renewal intent and the evolution state based on the inference output of the large language model.

[0057] The above scheme is executed by the context-aware analysis engine, and the specific implementation process is as follows: The first step is for the context-aware analysis engine to send the current dialogue interaction data of the current round to the large language model service. The goal of this step is to quickly obtain a preliminary intent understanding focused on the current statement and output a standardized current intent label, which serves as the starting point and one of the key inputs for subsequent in-depth analysis.

[0058] The second step involves the context-aware analysis engine constructing fused prompts according to specific templates or rules, which includes the following parts: The instruction information section clearly specifies the specific tasks, output format, and constraints that the large language model needs to perform. Contextual data section: Key information from dynamic user profiles and cross-session task status records is integrated into prompts in a clear and concise natural language format; In this way, the integrated prompts organically combine instructions, user status, task progress, and current dialogue into a whole, providing the large language model with all the background and clear action guidelines needed for execution analysis.

[0059] Third, the context-aware analysis engine sends the constructed fusion prompt information to the large language model. After receiving the fusion prompt information, the large language model's generation process is strongly constrained by the instruction information and is guided to execute the specified analysis task, rather than engaging in free dialogue. The large language model infers based on all the information in the prompt and generates output text that meets the instruction requirements.

[0060] The fourth step involves the output returned by the large language model, which typically conforms to the format specified in the instructions or is easily parsed natural language; the output parsing module of the context-aware analysis engine then processes this output. If the output is in a specified format, the values ​​of the "current_intention" and "evolution_state" fields are extracted directly.

[0061] If the output is natural language, then structured current renewal intent and evolution status information are extracted from it using predefined parsing rules.

[0062] This invention's embodiment abandons the pre-trained classification model method for identifying evolutionary states. Instead, it utilizes a large language model (LLM) and a two-stage guided analysis scheme. In the first stage, LLM performs preliminary intent identification on the current dialogue interaction data to obtain the current intent label. In the second stage, based on the results of the first stage, and combining dynamic user profiles and cross-session task state records, it constructs fused prompt information. This fused prompt information explicitly includes instruction information to constrain the LLM's inference direction, focusing it on answering specific questions related to the renewal intent and evolutionary state, and utilizing provided structured background information for reasoning. By constructing fused prompt information, the model's input length and computational burden are significantly reduced, while ensuring that key context is not overlooked.

[0063] In the above scheme, the evolution status is analyzed through a two-stage guided analysis scheme and the construction of fusion prompt information. However, the following problems exist in the specific implementation process: the dynamic user profile and cross-session task status record itself contain a lot of details. If they are all filled into the template without processing, the generated initial fusion prompt can easily become lengthy, thereby triggering model performance degradation and response delay.

[0064] To address this issue, the present invention proposes a solution that constructs fused prompt information based on the current intent tag, dynamic user profile, and cross-session task status records. This includes: selecting a matching prompt template from a pre-defined renewal scenario prompt template library based on the current intent tag; filling the corresponding placeholders in the matching prompt template with user status change features from the dynamic user profile and historical renewal intents and pending task statuses from the cross-session task status records to generate an initial fused prompt; and optimizing the initial fused prompt to obtain fused prompt information. The optimization includes information compression or format conversion.

[0065] In the specific implementation process, the first step is to query a locally stored renewal scenario prompt template library based on the current intent tag obtained from the previous step. The renewal scenario prompt template library has multiple prompt templates designed for different typical renewal scenarios. Each prompt template is a text file containing a fixed instruction framework and variable placeholders. The matching logic can be keyword matching or vector similarity matching.

[0066] The second step is to retrieve dynamic user profiles and cross-session task status records from the data loading module, and then perform the population operation: The user state transition features are initially rendered as text, and the placeholder {user state transition features} is filled in. At this time, the rendering is in the form of a simple list. Fill in the corresponding placeholders with the historical renewal intentions and the status of pending tasks; Fill the {current user question} placeholder with the current user dialogue text; Once the fill is complete, you will get an initial merge hint; at this point, the initial merge hint is a complete, but potentially lengthy or awkward, text string.

[0067] The third step involves the engine calling the suggestion optimizer to process the initial fusion suggestion. Specific optimization strategies include information compression, summary generation, and irrelevant information removal. After one or more of the above optimization operations, the final fusion suggestion information is obtained.

[0068] Furthermore, before inputting the fused prompt information into the large language model, the following steps are also included: determining the logical reasoning depth level required for this inference based on the current intent label and the user state transition features in the dynamic user profile; configuring the corresponding inference parameters based on the logical reasoning depth level; and adjusting the inference generation process of the large language model according to the inference parameters to guide it to perform context-aware inference at the logical reasoning depth level.

[0069] In the embodiments of the present invention, firstly, the logical reasoning depth level is determined by analyzing the current intent tag and user state transition characteristics, thereby achieving a preliminary assessment of the inherent complexity of the task.

[0070] Next, by mapping different levels of logical reasoning depth to different sets of inference parameters, a link is established from business complexity cognition to model technology control. In particular, the differentiated setting of temperature parameters directly corresponds to the different requirements of different depth reasoning tasks for output determinism and creativity. Shallow tasks require high determinism, while deep tasks require a certain degree of creativity to explore solutions.

[0071] Ultimately, by submitting the inferred parameters along with the fused prompts to the large language model, precise guidance for this model invocation behavior is achieved. For simple queries, low temperature enables the model to quickly converge to the most likely answer; for complex queries, the corresponding parameter configuration provides the model with sufficient space and output length for deep inference, ensuring the quality and completeness of the analysis results in complex scenarios.

[0072] Furthermore, the inference parameters include temperature parameters. Based on these parameters, the inference generation process of the large language model is adjusted to guide it to perform context-aware inference at the logical reasoning depth level. This includes: constructing a parameterized inference request by combining fused prompts and the set value of the temperature parameter; sending the parameterized inference request to the large language model service interface; wherein the set value of the temperature parameter is used to control the randomness of probability sampling in the model generation process at the service interface layer; and receiving the inference result generated based on the parameterized inference request returned by the large language model service interface as the output of the context-aware inference.

[0073] The present invention's embodiment injects logical reasoning deep into the underlying temperature parameters of a single service call of a large language model. First, by constructing a parameterized inference request, the set value of the temperature parameter based on the scenario is encapsulated into a machine-readable instruction set that conforms to the industry standard API format. This ensures that regardless of which large language model service is used in the backend, this control strategy can be deployed and take effect in the same way.

[0074] Secondly, the set value of the temperature parameter is used to regulate the probabilistic sampling randomness during the model generation process at the service interface layer. The service interface layer explicitly regulates this through public, standard APIs, rather than through internal model modifications or proprietary protocols. Regulating the probabilistic sampling randomness during model generation controls the degree of randomness in the model's selection of the next word from the vocabulary, which is one of the most direct and fundamental mechanisms affecting the balance between determinism and creativity in model output.

[0075] This solution optimizes resource utilization and adapts output quality to different scenarios through standardized interface calls and parameter passing.

[0076] like Figure 4 As shown, based on the above implementation scheme, a renewal strategy for the current dialogue round is generated according to the evolution state, dynamic user profile, and cross-session task state records. This includes: determining the strategy decision basis for the current dialogue round based on the evolution state and the pending task state; matching a basic strategy template from a preset renewal strategy library based on the strategy decision basis; adapting the user state transition features in the dynamic user profile to the basic strategy template to generate strategy configuration data for instantiating the basic strategy template; and instantiating the basic strategy template based on the strategy configuration data to generate the renewal strategy for the current dialogue round.

[0077] In practice, the process of generating the renewal strategy for the current dialogue round is as follows: First, the strategy generator receives the evolution status from the upstream analysis module and the pending task status extracted from the cross-session task status record. The generator internally maintains a strategy decision rule table, which uses the joint condition formed by the evolution status and the pending task status as the query key, mapping it to the corresponding strategy decision basis, thus providing a clear decision foundation for the current dialogue round.

[0078] The second step involves the strategy generator having a built-in pre-defined renewal strategy library, which stores multiple basic strategy templates. Each template is associated with one or more strategy decision criteria and is defined as a structured, parameterized operation framework used to describe the standard sequence of operations and configurable decision points required to achieve specific business objectives. Based on the strategy decision criteria determined in the first step, the generator retrieves and matches the most suitable basic strategy template from the renewal strategy library.

[0079] The third step involves the policy generator acquiring user state transition features from the dynamic user profile and executing adaptation logic: based on predefined mapping rules, it transforms the information in the state transition features into specific values ​​for the parameters required by the basic policy template. The adaptation process outputs policy configuration data, organized as a set of key-value pairs, containing all parameters and optional action flags needed to instantiate the template, providing data support for template concretization.

[0080] Fourth, the strategy generator performs an instantiation operation, filling the corresponding parameter placeholders in the basic strategy template with the specific values ​​from the strategy configuration data, and activating or deactivating the corresponding logical branches in the template based on the condition flags in the configuration data. After instantiation, the executable renewal strategy for the current dialogue round is generated.

[0081] Through the template- and rule-driven strategy generation mechanism described above, this embodiment ensures the compliance of business processes while guaranteeing the structural consistency, maintainability, and scalability of the renewal strategy, thereby supporting flexible and accurate strategy output in complex insurance scenarios.

[0082] Specifically, based on the evolution state and the pending task state, the strategy decision basis for the current dialogue round is determined, including: combining the evolution state and the pending task state into state pairs; matching the state pairs with a preset strategy decision basis mapping table; wherein, the strategy decision basis mapping table predefines the correspondence between different state pairs and strategy decision basis; and obtaining the corresponding strategy decision basis according to the matching query results.

[0083] The solution of this invention, firstly, cleverly transforms a multi-condition decision-making problem into a single-key query problem by combining the evolution state and the state of the task to be processed into state pairs. This avoids writing complex, multi-layered nested condition judgment logic in the code, eliminating the risk of code corruption and potential logical errors. Secondly, it introduces a preset strategy decision-making basis mapping table to shorten the business iteration cycle and reduce change risks and maintenance costs.

[0084] Example 2

[0085] Reference Figure 5 As shown, this embodiment of the invention provides an insurance APP renewal system based on dynamic user profiles, including: The data loading module loads the user's dynamic user profile and cross-session task status records based on the user identifier. The dynamic user profile includes intent tags updated based on the user's historical data and user status change characteristics. The cross-session task status records include historical renewal intents related to the renewal task, the sequence of operation instructions executed by the system, and the status of tasks to be processed. The context-aware analysis module receives the user's current dialogue interaction data, combines dynamic user profiles and cross-session task status records to perform context-aware analysis on the current dialogue interaction data, and identifies the user's current renewal intention and its evolution relative to historical renewal intentions. The renewal strategy generation module generates the renewal strategy for the current conversation round based on the evolution status, dynamic user profile, and cross-session task status records.

[0086] The embodiments of the present invention are used to execute the insurance APP renewal method based on dynamic user profile in Embodiment 1. They are based on the same inventive concept as Embodiment 1 and have the same beneficial effects, which will not be repeated here.

[0087] In summary, the insurance APP renewal method and system based on dynamic user profiles of the present invention achieves accurate identification and tracking of the evolution path of user renewal intentions by loading and integrating real-time updated user status and task process context. Based on this, it drives the generation of renewal strategies that are aligned with historical operation sequences and current business status, effectively suppressing the randomness of strategy generation and ensuring that each round of strategy generation can inherit historical processes and match current business logic, effectively improving the contextual coherence of renewal strategies, business alignment accuracy, and process advancement determinism.

[0088] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for renewing insurance policies via an insurance app based on dynamic user profiles, characterized in that, Includes the following steps: Obtain the user's user identifier, and load the user's dynamic user profile and cross-session task status record based on the user identifier; wherein, the dynamic user profile includes intent tags updated based on the user's historical data and user status change characteristics; the cross-session task status record includes historical renewal intents related to the renewal task, the sequence of operation instructions executed by the system, and the status of tasks to be processed; Receive the user's current dialogue interaction data, combine dynamic user profiles and cross-session task status records to perform context-aware analysis on the current dialogue interaction data, and identify the user's current renewal intention and its evolution relative to historical renewal intentions; Based on the evolution status, dynamic user profiles, and cross-session task status records, a renewal policy for the current dialogue round is generated.

2. The insurance APP renewal method based on dynamic user profiles according to claim 1, characterized in that, User status transition characteristics are generated based on user historical data and include the following steps: Based on a knowledge base of life events in insurance scenarios, entities and relationships are extracted from unstructured text information in users' historical data. The extracted entities and relationships are matched with the event triggering conditions in the insurance scenario life event knowledge base to obtain one or more potential life event tags and the confidence score corresponding to each potential life event tag; By combining potential life event tags and their corresponding confidence scores, user state transition features are generated. Among them, the event triggering conditions in the insurance scenario life event knowledge base are configured to simultaneously satisfy the association and temporal relationship between multiple semantic elements.

3. The insurance APP renewal method based on dynamic user profiles according to claim 2, characterized in that, The insurance scenario life event knowledge base is coupled with the rule engine. The insurance scenario life event knowledge base is configured to store multiple event trigger rule sets, and each event trigger rule set corresponds to an insurance-related life event type. The event triggering rule set is constructed through production rules, and the antecedent of the production rules is configured as a composite condition based on the association and temporal relationship between multiple semantic elements.

4. The insurance APP renewal method based on dynamic user profiles according to claim 1, characterized in that, By combining dynamic user profiles and cross-session task state records, context-aware analysis is performed on the current dialogue interaction data to identify the user's current renewal intention and its evolution relative to historical renewal intentions, including: Based on a large language model, the first intent is identified from the current dialogue interaction data to obtain the current intent label; Based on the current intent label, dynamic user profile, and cross-session task state records, a fusion prompt message is constructed; wherein, the fusion prompt message includes instruction information used to constrain the inference direction of the large language model; The fused prompts are input into the large language model to guide it in context-aware inference. Based on the inference output of the large language model, the current renewal intention and evolution status are obtained through parsing.

5. The insurance APP renewal method based on dynamic user profiles according to claim 4, characterized in that, Based on the current intent tag, dynamic user profile, and cross-session task state records, a fusion prompt message is constructed, including: Based on the current intent tag, select a matching prompt template from the preset renewal scenario prompt template library; The user status change features in the dynamic user profile and the historical renewal intentions and pending task statuses in the cross-session task status records are filled into the corresponding placeholders in the matching prompt template to generate the initial fusion prompt. The initial fusion prompt is optimized to obtain fusion prompt information; wherein, the optimization includes information compression or format conversion.

6. The insurance APP renewal method based on dynamic user profiles according to claim 4, characterized in that, Before inputting the fused prompts into the large language model, the following steps are also included: Based on the current intent tag and the user state change characteristics in the dynamic user profile, determine the level of logical reasoning depth required for this inference; Configure corresponding inference parameters based on the depth level of logical reasoning; Based on the inference parameters, the inference generation process of the large language model is adjusted to guide it to perform context-aware inference at the depth level of logical reasoning.

7. The insurance APP renewal method based on dynamic user profiles according to claim 6, characterized in that, Inference parameters include temperature parameters. Based on these parameters, the inference generation process of the large language model is adjusted to guide it in context-aware inference at the depth level of logical reasoning, including: By combining the fusion prompts and the set values ​​of temperature parameters, a parameterized inference request is constructed; The parameterized inference request is sent to the large language model service interface; wherein, the set value of the temperature parameter is used to adjust the randomness of probabilistic sampling during the model generation process at the service interface layer; Receive the inference results returned by the large language model service interface and use them as the output of context-aware inference.

8. The insurance APP renewal method based on dynamic user profiles according to claim 1, characterized in that, Based on the evolution status, dynamic user profiles, and cross-session task status records, a renewal strategy for the current dialogue round is generated, including: Based on the evolution status and the status of tasks to be processed, determine the strategic decision-making basis for the current dialogue round; Based on the strategic decision-making criteria, a basic strategy template is matched from the preset renewal strategy library; Adapt the user state transition features in the dynamic user profile to the basic policy template to generate policy configuration data for instantiating the basic policy template; Based on the strategy configuration data, the basic strategy template is instantiated to generate the renewal strategy for the current dialogue round.

9. The insurance APP renewal method based on dynamic user profiles according to claim 8, characterized in that, Based on the evolution state and the status of tasks to be processed, the strategic decision-making criteria for the current dialogue round are determined, including: Combine the evolution state and the task state to be processed into a state pair; Match the state pairs with the preset strategy decision-making basis mapping table; The strategy decision basis mapping table predefines the correspondence between different state pairs and strategy decision basis; Based on the results of the matching query, obtain the corresponding strategy decision-making basis.

10. An insurance app renewal system based on dynamic user profiles, used to implement the insurance app renewal method according to any one of claims 1-9, characterized in that, include: The data loading module loads the user's dynamic user profile and cross-session task status records based on the user identifier. The dynamic user profile includes intent tags updated based on the user's historical data and user status change characteristics. The cross-session task status records include historical renewal intents related to the renewal task, the sequence of operation instructions executed by the system, and the status of tasks to be processed. The context-aware analysis module receives the user's current dialogue interaction data, combines dynamic user profiles and cross-session task status records to perform context-aware analysis on the current dialogue interaction data, and identifies the user's current renewal intention and its evolution relative to historical renewal intentions. The renewal strategy generation module generates the renewal strategy for the current conversation round based on the evolution status, dynamic user profile, and cross-session task status records.

Images