Problem processing method and device based on artificial intelligence, computer equipment and medium

Abstract

The invention belongs to the technical field of artificial intelligence, and relates to an artificial intelligence-based question processing method and device, computer equipment and a storage medium, and the method comprises the steps: receiving question data inputted by a user; performing knowledge retrieval on the problem data based on a knowledge refining module to obtain a knowledge packet; analyzing the problem data based on a dialogue semantic modeling module to obtain a semantic state vector; performing prediction processing on the knowledge packet and the semantic state vector based on an entity prediction module to obtain an entity prediction result; processing the entity prediction result based on a dynamic prompt adjustment module to generate target prompt data; and the strategy-based decision-making module is used for making a decision on the problem data, the knowledge packet and the business target according to the target prompt data to generate reply data, and outputting the reply data. In addition, the invention also relates to a block chain technology, and reply data can be stored in a block chain. The method can be applied to a dialogue processing scene in the field of financial science and technology, and the accuracy and intelligence of problem processing are effectively improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and can be applied to the financial technology field, particularly to artificial intelligence-based problem-solving methods, devices, computer equipment, and storage media. Background Technology

[0002] In traditional intelligent dialogue service models in the financial insurance industry, existing systems mostly rely on static FAQ retrieval to handle customer inquiries. This model cannot dynamically track the evolution of customer intent across multiple rounds of dialogue, lacks precise knowledge filtering capabilities, resulting in low accuracy in problem handling and difficulty in achieving personalized recommendations. Specifically, in multi-round dialogue scenarios, traditional methods respond only based on the information of the current round, failing to fully consider the changes and accumulation of customer intent throughout historical dialogues, leading to discrepancies between the response and the customer's needs. For example, in the field of auto insurance, when inquiring about auto insurance, a customer might initially ask about the basic type of insurance, and in subsequent rounds gradually refine their inquiry to specific coverage clauses for particular car models, the scope of claims corresponding to different insured amounts, etc. Traditional systems cannot capture this process of intent refinement and may always respond with basic information, failing to provide accurate insurance recommendations and clause explanations tailored to the customer's deeper needs, thus reducing the efficiency and quality of customers obtaining effective information.

[0003] Therefore, there is an urgent need to provide an intelligent financial insurance dialogue system to improve the accuracy of problem handling, achieve personalized services, and enhance customer experience and insurance service efficiency. Summary of the Invention

[0004] The purpose of this application is to provide a problem-solving method, apparatus, computer device, and storage medium based on artificial intelligence, so as to solve the technical problem of low accuracy in problem-solving in existing dialogue service models.

[0005] Firstly, an artificial intelligence-based problem-solving method is provided, including: Receive user input for questions; The problem data is processed by a preset knowledge refinement module to obtain a corresponding knowledge package. The question data is parsed and processed based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector; The knowledge package and the semantic state vector are predicted based on the preset entity prediction module to obtain the corresponding entity prediction results. The entity prediction results are processed by a preset dynamic prompt adjustment module to generate corresponding target prompt data. Based on the preset strategy decision-making module, the system performs decision processing on the question data, the knowledge package, and the preset business objectives according to the target prompt data, and generates corresponding response data. The response data is then processed for output.

[0006] Secondly, an artificial intelligence-based problem-solving device is provided, comprising: The receiving module is used to receive user-input question data; The retrieval module is used to perform knowledge retrieval processing on the problem data based on a preset knowledge refinement module to obtain the corresponding knowledge package; The parsing module is used to parse and process the question data based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector; The prediction module is used to perform prediction processing on the knowledge package and the semantic state vector based on a preset entity prediction module to obtain the corresponding entity prediction result; The processing module is used to process the entity prediction results based on the preset dynamic prompt adjustment module to generate corresponding target prompt data; The generation module is used to perform decision processing on the question data, the knowledge package, and the preset business objectives based on the preset strategy decision module and the target prompt data, and generate corresponding response data. The output module is used to process the output of the response data.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described artificial intelligence-based problem-solving method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned problem-solving method based on artificial intelligence.

[0009] In the above-mentioned solution implemented by the AI-based problem-solving method, apparatus, computer equipment, and storage medium, the user-inputted problem data is first received; then, the problem data is processed by a preset knowledge refinement module to obtain a corresponding knowledge package; next, the problem data is parsed by a preset dialogue semantic modeling module to obtain a corresponding semantic state vector; then, the knowledge package and the semantic state vector are predicted by a preset entity prediction module to obtain a corresponding entity prediction result; subsequently, the entity prediction result is processed by a preset dynamic prompt adjustment module to generate corresponding target prompt data; further, a preset strategy decision module performs decision processing on the problem data, the knowledge package, and preset business objectives based on the target prompt data to generate corresponding response data; finally, the response data is output. Based on the above automated processing flow, this application, through the collaborative work of five core components—the knowledge refinement module, the dialogue semantic modeling module, the entity prediction module, the dynamic prompt adjustment module, and the strategy decision module—can automatically and accurately realize the entire process from information acquisition and semantic understanding to response generation, effectively improving the accuracy, intelligence, and adaptability of problem-solving, as well as the data accuracy of the generated response data. Attached Figure Description

[0010] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 This is a flowchart of an embodiment of the problem-solving method based on artificial intelligence according to this application; Figure 3 This is a schematic diagram of a structure of an embodiment of the artificial intelligence-based problem processing device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0013] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0015] like Figure 1 As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0016] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0017] Terminal device 101 can be various electronic devices with a display screen and support web browsing. In addition to laptops 1011, tablets 1012, or mobile phones 1013, terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer, and a desktop computer, etc.

[0018] Server 103 can be a server that provides various services, such as a backend server that provides support for the pages displayed on terminal device 101.

[0019] It should be noted that the problem-solving method based on artificial intelligence provided in this application is generally executed by a server / terminal device, and correspondingly, the problem-solving device based on artificial intelligence is generally set in the server / terminal device.

[0020] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0021] Continue to refer to Figure 2 This document illustrates a flowchart of an embodiment of the AI-based problem-solving method according to this application. The order of steps in the flowchart can be changed, and some steps can be omitted, depending on different needs. The AI-based problem-solving method provided in this application can be applied to any scenario requiring problem-solving, and therefore can be applied to products in these scenarios, such as problem-solving products in the financial insurance field. The AI-based problem-solving method includes the following steps: Step S201: Receive the question data input by the user.

[0022] In this embodiment, the problem-solving method based on artificial intelligence runs on an electronic device (e.g., Figure 1The server / terminal device shown can obtain user-inputted question data through wired or wireless connection. It should be noted that the above wireless connection methods may include, but are not limited to, 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other known or future wireless connection methods. The execution subject of this application is specifically a question processing system, or a dialogue system, which can be simply referred to as the system. The system can be composed of the following modules: (1) Data access module: responsible for accessing the financial insurance knowledge base, policy and regulation base, product terms base, and historical dialogue records. (2) Knowledge refinement module: performs relevance screening, noise removal, and weight reordering on the retrieved multi-source knowledge to generate a refined knowledge package. (3) Dialogue semantic modeling module: performs semantic parsing, intent recognition, and context tracking on multi-turn dialogues. (4) Entity prediction module: predicts potential key financial entities, such as insurance projects, claims conditions, and payment methods, based on historical context and knowledge packages. (5) Dynamic prompt adjustment module: adjusts the prompt content in real time according to changes in user status to optimize generative dialogue output. (6) Strategy Decision Module: Comprehensively analyzes user profiles, risk preferences, and product strategies to generate personalized responses and recommendations. (7) Feedback Learning Module: Models and transmits user feedback to achieve adaptive fine-tuning of each module.

[0023] The system can convert user voice input into text via an Automatic Speech Recognition (ASR) module, or directly receive user text input via a text parsing module, forming the raw input question data. Additionally, it can perform text processing on the question data and use the results for subsequent data processing. Text processing can include: Text correction: Correcting erroneous words in the raw input text, for example, correcting "critical illness display" to "critical illness insurance," resulting in corrected text data. Preliminary intent classification: Analyzing the corrected text to determine if the user's intent belongs to categories such as "terms inquiry" or "price inquiry," generating text data with intent tags. Entity extraction: Identifying key entities from the text with intent tags, such as "waiting period," forming structured text data containing intent tags and entities. Context initialization: If it's an intermediate round of dialogue, loading historical dialogue records from storage and combining them with the structured text data to form complete context data containing historical dialogue, current intent, and entities. If it's a new dialogue, initializing the user profile, for example, marking first-time inquiring customers as "newbies," and combining it with structured text data to form initial context data containing user profile, current intent, and entities.

[0024] Step S202: Based on the preset knowledge refinement module, the problem data is processed by knowledge retrieval to obtain the corresponding knowledge package.

[0025] In this embodiment, the specific implementation process of the above-mentioned knowledge retrieval processing of the problem data based on the preset knowledge refinement module to obtain the corresponding knowledge package will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0026] Step S203: Based on the preset dialogue semantic modeling module, the question data is parsed and processed to obtain the corresponding semantic state vector.

[0027] In this embodiment, the specific implementation process of parsing and processing the question data based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0028] Step S204: Based on the preset entity prediction module, perform prediction processing on the knowledge package and the semantic state vector to obtain the corresponding entity prediction result.

[0029] In this embodiment, the specific implementation process of the above-mentioned entity prediction module predicting the knowledge package and the semantic state vector to obtain the corresponding entity prediction result will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0030] Step S205: Based on the preset dynamic prompt adjustment module, the entity prediction result is processed to generate corresponding target prompt data.

[0031] In this embodiment, the specific implementation process of the above-mentioned dynamic prompt adjustment module processing the entity prediction result based on the preset method to generate corresponding target prompt data will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0032] Step S206: Based on the preset strategy decision module, the problem data, the knowledge package, and the preset business objectives are processed according to the target prompt data to generate corresponding response data.

[0033] In this embodiment, the specific implementation process of the preset strategy decision-making module, which performs decision processing on the question data, the knowledge package, and the preset business objectives based on the target prompt data to generate corresponding response data, will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0034] Step S207: Output the response data.

[0035] In this embodiment, the generated response data can be sent to the user to provide feedback on the issue, thereby completing the output processing of the response data. The system also has feedback learning and system optimization functions (feedback learning module), including: User feedback modeling: collecting explicit user feedback (1-5 star ratings) and implicit feedback (interaction duration, referral rate to human assistant, etc., which can be obtained from system interaction records). For example, if a user gives a low rating to "explanation of waiting period" and has a high referral rate to human assistant, it is marked as "points requiring optimization," forming user feedback data. Adaptive fine-tuning: continuously optimizing knowledge weights and dialogue strategies based on user feedback data, forming a self-closing learning mechanism of "knowledge refinement → semantic tracking → entity prediction → strategy generation → feedback optimization." For example, increasing the retrieval priority of high-satisfaction terms; optimizing dialogue strategies, such as strengthening the "predictive guidance" logic, reducing repeated user questions, forming system optimization parameter data.

[0036] This application first receives user-input question data; then, based on a preset knowledge refinement module, it performs knowledge retrieval processing on the question data to obtain a corresponding knowledge package; next, based on a preset dialogue semantic modeling module, it parses the question data to obtain a corresponding semantic state vector; then, based on a preset entity prediction module, it performs prediction processing on the knowledge package and the semantic state vector to obtain a corresponding entity prediction result; subsequently, based on a preset dynamic prompt adjustment module, it processes the entity prediction result to generate corresponding target prompt data; further, based on a preset strategy decision module, it performs decision processing on the question data, the knowledge package, and preset business objectives according to the target prompt data to generate corresponding response data; finally, it outputs the response data. Based on the above automated processing flow, this application, through the collaborative work of five core components—the knowledge refinement module, the dialogue semantic modeling module, the entity prediction module, the dynamic prompt adjustment module, and the strategy decision module—can automatically and accurately realize the entire process from information acquisition and semantic understanding to response generation, effectively improving the accuracy, intelligence, and adaptability of question processing, as well as the data accuracy of the generated response data.

[0037] In some alternative implementations, step S202 includes the following steps: Based on the knowledge refinement module, knowledge document fragments corresponding to the problem data are retrieved from a preset multi-source knowledge base.

[0038] In this embodiment, the aforementioned multi-source knowledge base includes at least an insurance product database, a financial regulations database, and an industry case database. The retrieval process based on the multi-source knowledge base includes: 1. Determining the retrieval scope. Knowledge source analysis: Clarifying the specific content scope of the three knowledge sources: the insurance product database, the financial regulations database, and the industry case database. The insurance product database contains detailed terms, characteristics, and applicable groups for various insurance products; the financial regulations database covers laws, regulations, and regulatory policies related to insurance; and the industry case database collects actual cases in insurance business, including claims cases and sales cases. User question analysis: Conducting in-depth analysis of the question data submitted by users to extract key information, such as the type of insurance product, business process, and regulatory points involved in the question. For example, for the question "How long is the waiting period for critical illness insurance?", the key information "critical illness insurance" and "waiting period" are extracted. 2. Constructing a retrieval strategy. Keyword extraction: Extracting keywords from the question data as the main basis for retrieval. Keyword extraction algorithms in natural language processing technology, such as the TF-IDF algorithm, can be used to identify words with high weight and representativeness in the question. Semantic Expansion: Considering the potential diversity in the expression of user questions, the extracted keywords are semantically expanded. For example, for "waiting period," synonyms or related terms such as "observation period" and "exclusion period" can be derived. 3. Perform the search operation. Database Search: Based on the constructed search strategy, searches are conducted in the insurance product database, financial regulations database, and industry case database. In each database, methods such as keyword matching and semantic similarity calculation are used to filter out document paragraphs related to the user question. Result Integration: The search results from the three knowledge sources are initially integrated, and duplicate document paragraphs are removed to form a preliminary result set containing all relevant documents, i.e., the aforementioned knowledge document fragments.

[0039] Obtain the dialogue context data corresponding to the question data.

[0040] In this embodiment, the historical dialogue records corresponding to the aforementioned user can be loaded, and the dialogue context data associated with the current text data can be extracted from the historical dialogue records.

[0041] Based on the knowledge document fragments and the dialogue context data, a knowledge relevance assessment is performed to obtain a knowledge list sorted by relevance.

[0042] In this embodiment, the aforementioned knowledge relevance assessment includes: 1. Design objective. After multi-source knowledge retrieval, it is necessary to quantify the matching degree between each candidate knowledge and the current dialogue context to ensure that the knowledge entering subsequent processes is "both relevant and accurate". 2. Method principle. A knowledge scoring matrix is ​​constructed, where each row represents a candidate knowledge document fragment. Columns represent semantic units of the dialogue context. (Such as the most recent user question, context summary, system response, etc.). Each element of the matrix represents a semantic similarity score. ,in: This represents the vector representation encoded using a Transformer (such as FinBERT or DeBERTa). Cosine similarity is usually used. 3. Actual Construction Steps. Encode the knowledge and dialogue context separately: Divide the knowledge document into several semantic blocks (each block containing 1-3 sentences); divide the dialogue context into semantic units (including current user input, historical summaries, system response summaries, etc.); input each into the Transformer model to obtain semantic vectors. , Calculate pairwise similarity and construct a matrix: iterate through all pairs of pairs. Calculate similarity scores and construct a matrix. The overall knowledge score is obtained by aggregation: the relevance of each knowledge block is aggregated (maximum value, weighted average, or attention aggregation can be used): or: Among them, weight Based on the semantic importance of the dialogue (such as keywords recently mentioned by the customer), a knowledge list sorted by relevance is finally obtained, providing input for subsequent filtering and fusion.

[0043] The knowledge list is subjected to knowledge noise filtering and fusion processing to obtain processed knowledge data.

[0044] In this embodiment, the aforementioned knowledge noise filtering and fusion includes: 1. Definition of Information Gain Ratio. "Information Gain Ratio (IGR)" originates from information theory and is used to measure the quality of a piece of knowledge. To explain the intent of the current dialogue The degree of contribution. Defined as follows: ,in: Information gain; Entropy function; denominator Used to penalize knowledge fragments that are only partially relevant but incidentally useful. 2. Calculation approach. First, estimate the knowledge fragments based on historical interaction logs or model simulation feedback. When it occurs, it increases the probability of a correct answer or a satisfactory dialogue outcome. Information gain is obtained by calculating the entropy difference; then normalization is performed by combining the entropy of the knowledge itself (information diversity). Intuitive understanding: if a piece of knowledge significantly improves the accuracy of the system's response in multiple similar scenarios, its information gain is high and it should be retained; if it appears frequently but contributes little or causes confusion, it should be removed. 3. Filtering and fusion strategies. Filtering stage: Setting thresholds. ,like If the similarity is high but the information gain is low, it is considered noise knowledge removal; redundancy items with high similarity but low information gain are deduplicated. Fusion stage: hierarchical clustering is performed on the retained knowledge, and the similarity measure is the cosine distance of sentence vectors; summary fusion is performed on each cluster (e.g., using BART or LLM to generate a summary) to form knowledge summary units (KSUs), or knowledge data; each KSU becomes the basic building block of the "knowledge package".

[0045] The knowledge data is processed based on a preset knowledge package generation strategy to obtain the corresponding knowledge package.

[0046] In this embodiment, the specific implementation process of generating knowledge data based on the preset knowledge package generation strategy to obtain the corresponding knowledge package will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0047] Based on the above processing flow, this application retrieves knowledge document fragments corresponding to the question data from a pre-set multi-source knowledge base using a knowledge refinement module; then, it obtains the dialogue context data corresponding to the question data; next, it performs knowledge relevance assessment based on the knowledge document fragments and dialogue context data to obtain a knowledge list sorted by relevance; subsequently, it performs knowledge noise filtering and fusion processing on the knowledge list to obtain processed knowledge data; finally, it performs data generation processing on the knowledge data based on a pre-set knowledge package generation strategy to obtain the corresponding knowledge package. Thus, this application, through the use of the knowledge refinement module, can efficiently and accurately complete the knowledge retrieval processing of question data, improving the processing efficiency of knowledge retrieval and ensuring the accuracy of the obtained knowledge package data.

[0048] In some optional implementations of this embodiment, the step of generating a corresponding knowledge package by processing the knowledge data based on a preset knowledge package generation strategy includes the following steps: Obtain the semantic context features and intent category features corresponding to the problem data.

[0049] In this embodiment, the aforementioned semantic context features may refer to the dialogue summary (topic, emotion, keywords) of the current round. Additionally, the user's intent category characteristics can be identified through an intent recognition model (classifier or prompt-based inference), such as whether the user is in a "consultation / hesitation / purchase / complaint" stage.

[0050] Obtain the user's profile features.

[0051] In this embodiment, user personal information, historical conversation records, and purchasing behavior data are collected in advance to construct a corresponding customer profile. The customer profile may include information such as the customer's age, gender, occupation, risk preference, and insurance needs. Then, key user characteristics (such as risk preference, age group, product history, and region) are extracted from the customer profile to obtain profile features. These features can reflect the user's personalized needs and preferences. For example, younger customers may be more concerned with the innovation and flexibility of insurance products, while older customers may be more concerned with the stability and coverage of insurance products.

[0052] The semantic context features, intent category features, profile features, and knowledge data are integrated using a preset dynamic knowledge fusion processor to obtain the corresponding integrated data.

[0053] In this embodiment, the system uses a dynamic knowledge fusion engine for feature integration. Its core algorithm can be described as follows: ,in: Typically, this is achieved using a weighted attention network or an instruction-driven generative model; the output is a structured summary package, including: a summary of relevant terms; supplementary explanations adapted to the user profile; and possible computational examples (such as premium calculation cases). The final generated knowledge package is a JSON structure, for example: { "intent": "compare_products", "key_points": ["Critical Illness Insurance Waiting Period: 90 days", "Premium Installment: Monthly Payment Supported"], "context_fit": 0.92, "recommendation_basis": "The client is 35 years old and prioritizes family protection." The integrated data is used as the knowledge package.

[0054] Based on the above processing flow, this application acquires semantic context features and intent category features corresponding to the problem data, as well as user profile features. Then, it integrates these features using a pre-defined dynamic knowledge fusion tool to obtain integrated data. This integrated data is subsequently used as a knowledge package. Thus, by using a dynamic knowledge fusion tool to integrate the acquired semantic context features, intent category features, profile features, and generated knowledge data, this application can automatically and accurately generate the required knowledge package, ensuring the accuracy of the obtained knowledge package.

[0055] In some alternative implementations, step S203 includes the following steps: Based on the dialogue semantic modeling module, semantic embedding and extraction are performed on the question data to obtain the corresponding current semantic vector.

[0056] In this embodiment, the text of each round of dialogue (such as the text data mentioned above) can be input into a pre-trained language model for word segmentation and encoding. For example, the tokenizer of the BERT model can be used to segment the text into sub-word units and convert them into numerical codes that the model can recognize. Alternatively, the FinBERT model can be used. FinBERT is a BERT model specifically fine-tuned for the financial field (including insurance). It is pre-trained on financial text data and can better understand the professional terminology and semantics of the financial field. If the business focus is on the insurance field, FinBERT may be a more suitable choice. Then, the encoded text is input into the forward propagation process of the model to extract the semantic embedding vector of the text. The semantic embedding vector is a high-dimensional vector that can represent the semantic features of the text. For example, for a user input "How much is the premium for this insurance?", the model will generate a corresponding semantic embedding vector, which contains semantic information about the concept of "premium" and the entire question.

[0057] The current semantic vector and the corresponding historical semantic vector are processed based on a preset cross-wheel attention mechanism to obtain the corresponding fused semantic representation data.

[0058] In this embodiment, intent evolution refers to the process by which a user's intent gradually refines, transforms, or expands from an initial goal during a multi-turn dialogue. For example: Turn 1: I want to learn about health insurance → Turn 2: What diseases are covered by critical illness insurance? → Turn 3: How is the premium calculated? At this point, the intent evolves from "understanding the type of insurance" to "understanding the details of the terms," ​​and then to "focusing on the price." Capturing this dynamic change is crucial for generating context-consistent and accurate responses. Therefore, to capture intent evolution, a cross-turn attention mechanism / structure (which can be based on Transformer or GRU) is introduced. Each round of dialogue is encoded as a vector. The model applies to the current round vector. With historical wheel vector Establish attention weights:

[0059] Fusion of historical semantic representation (fusion of semantic representation data):

[0060] in, Temperature coefficient. Fusion vector. This represents the "current intent state" and also reflects the historical intent trajectory. j is an index variable used to traverse all historical rounds, and k can be t 1. t Level 2 refers to any round preceding the current round. This is achieved through continuous time series... The system can detect intent drift, intent refinement, or intent reversal, and dynamically adjust the strategy accordingly.

[0061] Furthermore, the semantic embedding vectors from previous rounds are weighted and summed according to assigned attention weights to obtain a context vector that incorporates historical context information. This context vector reflects the background and relevance of the current round's dialogue within the historical dialogue. The semantic embedding vector of the current round is then integrated with the context vector, for example, through concatenation or weighted averaging, to obtain a vector representation that comprehensively considers information from both the current and historical rounds—that is, fused semantic representation data. This vector can better capture the evolution of customer intent.

[0062] Obtain the preset semantic memory unit.

[0063] In this embodiment, a semantic memory unit is pre-introduced to store key dialogue facts (such as customer age, premium budget, and insurance preferences) and is updated in each round. Specifically, the introduction process of the semantic memory unit includes: 1. Memory unit structure design. Content definition: Determine the key dialogue facts that the semantic memory unit needs to store, such as customer age, premium budget, and insurance preferences. This information is crucial for understanding the customer's intent and needs and needs to be continuously saved and updated throughout multiple rounds of dialogue. 2. Storage format planning: Design the storage format of the memory unit to efficiently store and retrieve key dialogue facts. For example, information can be stored in key-value pairs, where the key represents the type of information (e.g., "customer age"), and the value represents a specific numerical value or description (e.g., "30 years old"). 3. Memory unit initialization. Initial information filling: At the start of the dialogue, the semantic memory unit is initialized and filled based on the initial information of the dialogue (e.g., user registration information, information explicitly mentioned in the first round of dialogue, etc.). For example, if the user mentions their age as 30 years old in the first round of dialogue, the information "customer age: 30 years old" is stored in the memory unit.

[0064] The semantic memory unit is updated based on the fused semantic representation data to obtain the corresponding updated data.

[0065] In this embodiment, the update process of the semantic memory unit includes: Information extraction: In each round of dialogue, information that may update the memory unit is extracted from the dialogue text and semantic embedding vector of the current round. For example, the semantic understanding model identifies that the user's mentioned premium budget is 5,000 yuan and the insurance type preference is critical illness insurance. Information verification and updating: The extracted information is verified to ensure its accuracy and rationality. If the information is valid, the corresponding key-value pair in the semantic memory unit is updated. For example, "Premium budget: 5,000 yuan" and "Insurance type preference: critical illness insurance" are updated to the memory unit. Vector construction: Information integration: The semantic embedding vector of the current round, the context information vector generated by the cross-round attention mechanism, and the key dialogue fact information in the semantic memory unit are integrated. This can be done by concatenating these information into a longer vector, or by using a specific fusion method (such as weighted average) to fuse them into a comprehensive vector. Feature extraction: The integrated vector is further subjected to feature extraction and dimensionality reduction to remove redundant information and retain the most critical features. For example, Principal Component Analysis (PCA) or other dimensionality reduction techniques can be used to transform high-dimensional vectors into low-dimensional structured semantic state vectors. Vector Standardization: Numerical Range Adjustment: The eigenvalues ​​in the structured semantic state vector are standardized to ensure their values ​​fall within a uniform range (e.g., between 0 and 1). This facilitates subsequent modules in processing and comparing the vectors. Vector Format Standardization: Ensures the output structured semantic state vectors conform to predefined format requirements so that subsequent modules can correctly reference and use them. For example, specifying the vector's dimension, data type, etc. Vector Output and Storage: Real-time Output: At the end of each round of dialogue, the generated structured semantic state vectors are output to subsequent modules in real time, such as the strategy decision-making module. This ensures that subsequent modules can obtain the latest semantic state information in a timely manner to make appropriate decisions. Historical Storage: The structured semantic state vectors generated in each round of dialogue are stored historically for retrospective analysis when needed. For example, when handling customer complaints or conducting dialogue quality assessments, historical semantic state vectors can be consulted to understand the development process of the dialogue and changes in customer intent.

[0066] The updated data is used as the semantic state vector.

[0067] Based on the above processing flow, this application extracts the current semantic vector by semantic embedding of the question data using the dialogue semantic modeling module, and processes the current semantic vector and the corresponding historical semantic vector using the cross-turn attention mechanism to obtain fused semantic representation data. Then, the semantic memory unit is updated based on the fused semantic representation data, and the updated data is used as the corresponding semantic state vector. This enables efficient and accurate parsing of the question data and ensures the accuracy of the generated semantic state vector.

[0068] In some alternative implementations, step S204 includes the following steps: The entity prediction module invokes a preset dual-channel network structure.

[0069] In this embodiment, the entity prediction module simulates a doctor's "symptom prediction" behavior to predict potential "entities of interest" or "decision variables" in financial scenarios. Entity types are defined in five categories: insurance type, amount, time, conditions, and discounts. Insurance type: In insurance, the insurance type is a core concept. Different insurance types have different coverage, premiums, and applicable populations. When consulting or purchasing insurance, customers may be interested in specific types, such as critical illness insurance, accident insurance, and life insurance. Amount: Amounts are ubiquitous in financial transactions and insurance. Customers may be concerned about premiums, coverage amounts, claim amounts, and payment amounts. This monetary information directly affects customer decisions and interests. Time: Time-related information is also very important in insurance. For example, the effective date of the insurance, payment period, claim processing time, and waiting period. Customers need to understand this time information to plan their insurance plans. Conditions: Insurance products are usually accompanied by various conditions, such as underwriting conditions, claim conditions, and exclusion clauses. Customers need to understand these conditions to determine if they meet the underwriting requirements and under what circumstances they can receive compensation. Promotions: To attract customers and enhance competitiveness, insurance companies offer various promotional activities, such as discounts, gifts, and value-added services. Customers may pay attention to these promotional information to obtain more benefits.

[0070] In addition, the above dual-channel network structure can adopt a dual-channel Transformer structure, which includes a semantic state vector channel for receiving the structured semantic state vector output from the semantic modeling module, and a knowledge package embedding channel for receiving the knowledge package output from the knowledge refinement module.

[0071] Based on the dual-channel network structure, the knowledge package and the semantic state vector are predicted and processed respectively to obtain the corresponding knowledge features and semantic features.

[0072] In this embodiment, the semantic state vector described above can be input into the Transformer encoder of the dual-channel network structure. The Transformer encoder consists of multiple self-attention layers and feedforward neural network layers, capable of capturing long-term dependencies and semantic features in the vector. Through the self-attention mechanism, the model can focus on the relationships between different parts of the vector, thereby better understanding the semantic state of the current dialogue. Furthermore, the knowledge packet embedding is input into another Transformer encoder of the dual-channel network structure. Similarly, through the self-attention mechanism and feedforward neural network layers, the knowledge packet embedding is encoded, and its knowledge features are extracted.

[0073] The knowledge features and semantic features are fused based on a preset gating mechanism to obtain the corresponding fused features.

[0074] In this embodiment, the gating mechanism is a mechanism for controlling the flow of information. It can dynamically adjust the transmission and fusion of information based on the importance of the input information. In the dual-channel Transformer structure, the gating mechanism is used to fuse the outputs of the semantic state vector channel and the knowledge packet embedding channel. The fusion process includes: calculating gating weights: inputting the output vectors of the two channels into a gating network. The gating network is usually composed of a simple neural network, which can calculate the gating weights based on the similarity and importance of the two channel vectors. The gating weights represent the contribution ratio of the two channels' information in the fusion process. Vector fusion: based on the calculated gating weights, the output vectors of the two channels are weighted and summed to obtain the fused vector (i.e., the fused feature). This fused vector contains both the semantic information of the current dialogue and relevant knowledge information, and can more comprehensively reflect the potential entity distribution.

[0075] The fused features are classified based on a preset classifier to obtain the corresponding classification results.

[0076] In this embodiment, the fused features are input into a classifier, such as a softmax classifier. The classifier calculates the probability distribution for each entity type based on the vector features. This probability distribution represents the likelihood of each potential entity type appearing in the current dialogue state. It then outputs the potential entity distribution, i.e., the probability value corresponding to each entity type. For example, for the entity type "insurance type," the model might output a higher probability value, indicating that the customer might be interested in a specific insurance type in the current dialogue. Furthermore, based on the output potential entity distribution, the entity type with the higher probability value is selected as the prediction result, i.e., the classification result. These entity types represent the questions or product attributes that the customer might be interested in next in the current dialogue state.

[0077] The classification result is used as the entity prediction result.

[0078] In this embodiment, the prediction results can be further filtered and optimized by combining the logic of financial scenarios and insurance business. For example, if the model predicts that a customer may be interested in the "discount" entity, but according to the current dialogue content and business rules, the insurance product may not have relevant discount activities, then the prediction results can be adjusted to select other more suitable entity types.

[0079] Based on the above processing flow, this application invokes a preset dual-channel network structure based on the entity prediction module; and performs prediction processing on the knowledge package and semantic state vector respectively based on the dual-channel network structure to obtain corresponding knowledge features and semantic features; then, based on a preset gating mechanism, the knowledge features and semantic features are fused to obtain corresponding fused features; subsequently, based on a preset classifier, the fused features are classified to obtain corresponding classification results; and finally, the classification results are used as entity prediction results. Thus, by combining the dual-channel network structure, gating mechanism, and classifier, this application can efficiently and accurately complete the prediction processing between the knowledge package and semantic state vector, ensuring the accuracy of the generated entity prediction results.

[0080] In some optional implementations of this embodiment, step S205 includes the following steps: The adaptive generator is invoked based on the dynamic prompt adjustment module.

[0081] In this embodiment, the aforementioned adaptive generator is a pre-built adaptive Prompt generator, used to select and adjust appropriate Prompt templates based on input signals (customer intent shifts, mood changes, entity prediction results, and knowledge updates). The adaptive generator defines a series of rules that trigger corresponding template selection when specific conditions are met. A rich pre-built Prompt template library contains various types and structures, such as guiding question templates, explanatory templates, and comparative clause templates. Each template is designed for a specific dialogue context and customer needs.

[0082] Get the preset prompt template.

[0083] In this embodiment, the above-mentioned prompt template refers to a prompt template that matches the user's current dialogue context and needs, and the prompt template can be obtained from the above-mentioned Prompt template library.

[0084] Based on the adaptive generator, the prompt template is adjusted according to the entity prediction result and the semantic state vector to obtain the adjusted first prompt template.

[0085] In this embodiment, the Prompt template structure can be adjusted according to different situations: 1) Decreased user interest. Detection mechanism: By monitoring user participation indicators in the conversation, such as response time, response length, and number of proactive questions, it can be determined whether user interest has decreased. If user participation is significantly reduced, it is determined that user interest has decreased. Adjustment strategy: When a decrease in user interest is detected, a template containing "risk mitigation language" is selected from the template library. For example, when recommending insurance products, phrases such as "Our insurance products can effectively reduce your economic risks in unexpected situations and provide more comprehensive protection for you and your family" can be added to rekindle user interest. 2) User hesitation. Detection mechanism: Observe the user's conversation content. If the user expresses hesitation when expressing their needs or choices, such as "I'll think about it" or "I'm not sure," it is determined that the user is hesitant. Adjustment strategy: When user hesitation is detected, a template with "comparative clause explanation" is selected. For example, when introducing insurance products, the coverage, premiums, and claims conditions of different products can be compared to help users better understand the differences between the products and make a decision. 3) Repeated Questions by Users. Detection Mechanism: Records the questions users ask in the conversation. If a user is found to be asking the same or similar questions repeatedly, it is considered a repeated question. Adjustment Strategy: When a repeated question is detected, a "FAQ Supplementary Block" is automatically introduced. Frequently asked questions and answers related to the question are extracted from the knowledge base and presented to the user as supplementary explanations to help the user gain a more comprehensive understanding of the relevant information.

[0086] One approach is to use reinforcement learning (RLHF) to optimize the Prompt strategy, using "dialogue completion + user satisfaction" as the reward function to dynamically adjust the strategy parameters. The specific implementation process includes: 1. Building the reinforcement learning environment. State definition: Define various information during the dialogue as states, including customer intent, emotions, entity prediction results, and current prompt content. These state information collectively constitute the state space of the reinforcement learning environment. Action definition: Define various adjustment actions in the prompt reconstruction strategy as actions, such as selecting different Prompt templates or adding specific types of dialogue. The action space contains all possible prompt adjustment methods. Reward function design: Design a reward function using "dialogue completion + user satisfaction". Dialogue completion can be evaluated by measuring whether the dialogue has achieved the expected goal (e.g., the customer completes insurance purchase, understands the required information, etc.); user satisfaction can be measured by the customer's feedback after the dialogue ends (e.g., satisfaction rating, willingness to continue the dialogue, etc.). 2. Training the reinforcement learning model. Policy initialization: Initialize a reinforcement learning policy model, which can select appropriate actions based on the current state. The initial policy can be a random selection of actions or a selection based on simple rules. Interactive Learning: The reinforcement learning model learns interactively in a simulated dialogue environment. During each dialogue, the model selects an action based on the current state and then observes the reward signal from the environment. By continuously trying different actions, the model gradually learns which actions yield higher rewards. Policy Optimization: Based on the observed reward signals, reinforcement learning algorithms (such as Q-learning, policy gradient, etc.) are used to optimize the policy model. By adjusting the parameters of the policy model, it can select better actions in future dialogues, thereby improving dialogue completion and user satisfaction. 3. Dynamic Adjustment of Policy Parameters. Real-time Feedback: During actual dialogues, customer feedback information, such as dialogue completion and user satisfaction, is collected in real time as real-time feedback for the reinforcement learning model. Parameter Update: Based on real-time feedback, reinforcement learning algorithms are used to dynamically adjust the parameters of the policy model. By continuously adjusting the parameters, the policy model can adapt to different customer interaction paths and intent changes, achieving precise and personalized dialogue guidance.

[0087] The first prompt template is optimized based on a preset adaptation strategy to obtain the corresponding second prompt template.

[0088] In this embodiment, the language style of the adjusted first prompt template can be adjusted using the user profile of the user mentioned above, such as making it more professional or more colloquial, to form the final adapted Prompt template data, i.e., the second prompt template.

[0089] Use the second prompt template as the target prompt data.

[0090] Based on the above processing flow, this application calls an adaptive generator based on the dynamic prompt adjustment module; then obtains a preset prompt template; subsequently, based on the adaptive generator, the prompt template is content-adjusted according to the entity prediction result and semantic state vector to obtain an adjusted first prompt template; subsequently, the first prompt template is optimized based on a preset adaptation strategy to obtain a corresponding second prompt template; finally, the second prompt template is used as the target prompt data. Thus, this application, by using an adaptive generator, adjusting the prompt template according to the entity prediction result and semantic state vector to obtain a first prompt template, then optimizing the first prompt template based on an adaptation strategy, and using the obtained second prompt template as the target prompt data, can efficiently and accurately complete the generation of target prompt data, ensuring the accuracy and quality of the generated target prompt data.

[0091] In some optional implementations of this embodiment, step S206 includes the following steps: The corresponding semantic confidence is obtained based on the semantic state vector.

[0092] In this embodiment, the generated semantic state vector consists of the current intent category (e.g., consultation / hesitation / purchase) corresponding to the question data, semantic slot vectorization (e.g., insurance type, amount, period), semantic encoding of the dialogue context, customer sentiment score (positive, negative), the model's confidence in the current understanding (semantic confidence), and the current dialogue turn. Information can then be extracted from the above semantic state vector to obtain the required semantic confidence.

[0093] The strategy decision module obtains the response generation strategy corresponding to the semantic confidence level.

[0094] In this embodiment, if the semantic confidence level is higher than a preset confidence threshold, it is considered high confidence, and the response generation strategy corresponding to high confidence is a high confidence response generation strategy. If the semantic confidence level is lower than the preset confidence threshold, it is considered low confidence, and the response generation strategy corresponding to low confidence is a low confidence response generation strategy. The value of the aforementioned confidence threshold is not specifically limited and can be set according to actual business needs.

[0095] Based on the response generation strategy, decision-making processes are performed on the question data, the knowledge package, and the business objective according to the target prompt data to generate corresponding decision data.

[0096] In this embodiment, the decision-making process for the response generation strategy under high confidence conditions includes: Policy Network Activation: When the semantic confidence level is higher than a preset confidence threshold, the Policy Network is activated. The Policy Network is a trained neural network model that can generate recommendation strategies based on the dialogue state, knowledge results (i.e., knowledge packages), and business objectives (the business processing objectives of the current scenario). Policy Generation: The Policy Network takes the dialogue state, knowledge results, and business objectives as inputs and performs calculations and reasoning through its internal neural network structure to generate specific recommendation strategies. For example, in an insurance scenario, if a user shows strong interest in a certain insurance product and has high semantic confidence, the Policy Network may generate a strategy recommending that the user purchase the product, along with detailed product advantages and a purchase link. Policy Evaluation and Optimization: The generated recommendation strategy needs to be evaluated to ensure that it meets business objectives and compliance requirements. The strategy can be evaluated through simulation testing or actual small-scale applications. Based on the evaluation results, the Policy Network can be optimized and adjusted to improve the accuracy and effectiveness of its generated strategies.

[0097] While basic recommendations are already provided, to enrich and personalize them, making them conform to specific expression styles or include more relevant information, the generated target prompt data can be used to further generate responses that meet user needs and contexts. For example, the generated target prompt data (Prompt templates) may include recommendation script style templates tailored to different customer groups (such as young customers and elderly customers). When generating a recommendation like "Recommended monthly payment method, single-period premium reduced by 40%", appropriate script styles can be selected from the Prompt templates to supplement and improve the recommendation response based on the customer's age, consumption habits, and other characteristics, making the recommendation response more closely aligned with customer needs.

[0098] The decision-making process for response generation strategies in cases of low confidence or customer objections includes: Rule Engine Triggering: The rule engine is triggered when the semantic confidence level falls below a confidence threshold or when the customer objects to the system's response. The rule engine is a predefined set of rules that executes corresponding operations based on specific conditions and logic. Remedial Strategy Selection: The rule engine selects an appropriate remedial strategy from a pre-defined remedial strategy library based on the current dialogue state, knowledge outcomes (i.e., knowledge packages), and business objectives. For example, if a user has doubts about the insurance terms and has low semantic confidence, the rule engine might choose to explain the terms clearly and understandably. If the user is dissatisfied with the system's response and is emotionally agitated, the rule engine might choose to transfer the user to a human agent for further communication and processing. Strategy Execution and Feedback: The selected remedial strategy is executed, and user feedback is collected. For example, after explaining the terms, the user is asked if they understand; after transferring to a human agent, the communication between the user and the human agent and the final processing result are recorded. The rules engine was optimized and improved based on user feedback to enhance its ability to handle low confidence levels and customer objections.

[0099] When the rules engine is triggered to interpret clauses or transfer to human assistance, it may only determine the basic direction of processing, but the specific response needs to be more flexible and diverse. In this case, target prompt data can be used to guide the generation of a more appropriate response. For example, if the rules engine determines that a certain clause needs interpretation, the generated target prompt data may contain a structure template for clause interpretation, including introducing the clause content, explaining its meaning, and finally describing its impact. Based on this prompt template and the specific clause content, a clearer and easier-to-understand explanation response can be generated.

[0100] The decision data is used as the response data.

[0101] In this embodiment, the system also supports dynamic switching of strategy subnetworks based on different insurance scenarios (such as insurance application, claims, renewal, complaints, etc.) to ensure compliance and scenario specificity. The specific implementation process includes: 1. Scenario recognition. Keyword and intent analysis: By extracting keywords and analyzing intent from the user's input text, the insurance scenario of the current dialogue is identified. For example, if the user's input contains keywords such as "insurance application" or "purchase insurance," and the intent is to inquire or purchase, then the current scenario is determined to be an insurance application scenario; if the user's input contains keywords such as "claims" or "apply for compensation," and the intent is to apply for or inquire about claims-related matters, then the current scenario is determined to be a claims scenario. Context association: The scenario is further confirmed by combining the context information of the dialogue. For example, if the user has been inquiring about the details of a certain insurance product in the previous rounds of dialogue, and the user expresses the intention to purchase in this round of dialogue, then it can be more accurately determined that the current scenario is an insurance application scenario; if the user has previously submitted a claims application, and this round of dialogue revolves around the claims progress or claims result, then it is determined to be a claims scenario. 2. Strategy subnetwork switching. Sub-network selection: Based on the identified insurance scenario, select the corresponding sub-network from multiple pre-built strategy sub-networks. Each strategy sub-network is trained and optimized for a specific insurance scenario, enabling it to better handle decision-making problems within that scenario. For example, the insurance application scenario strategy sub-network focuses on recommending suitable insurance products based on user needs and guiding users through the application process; the claims scenario strategy sub-network focuses on quickly and accurately processing claims and answering user questions about claims. 3. Data adaptation and adjustment: Adapt and adjust the input data, such as dialogue states, knowledge results, and business objectives, according to the requirements of the selected strategy sub-network. Different strategy sub-networks may have different requirements for the format and content of input data, so appropriate processing is required to ensure that the data can be correctly understood and used. For example, the claims scenario strategy sub-network may require more detailed claims application information as input, so relevant information in the dialogue state needs to be organized and supplemented before inputting data into this sub-network. 4. Decision execution and monitoring: Make decisions using the selected strategy sub-network and monitor the execution and effectiveness of the decisions in real time. For example, in the insurance application scenario, the system monitors whether users have completed the application process according to the recommended strategy; in the claims scenario, it monitors the processing progress of claims and user satisfaction. Based on the monitoring results, the strategy subnetwork is dynamically adjusted and optimized to improve its decision-making performance in different scenarios.

[0102] Based on the above processing flow, this application obtains the corresponding semantic confidence level based on the semantic state vector; then, it obtains the response generation strategy corresponding to the semantic confidence level based on the strategy decision module; subsequently, based on the response generation strategy, it performs decision processing on the problem data, knowledge package, and business objectives according to the target prompt data, generating corresponding decision data; finally, the decision data is used as the response data. Thus, this application obtains the response generation strategy corresponding to the semantic confidence level through the use of the strategy decision module, and then, based on the use of the response generation strategy, it can perform decision processing on the problem data, knowledge package, and business objectives according to the target prompt data, thereby automatically and accurately generating the corresponding response data, improving the intelligence and accuracy of response data generation.

[0103] In some alternative implementations, the user information obtained is subject to user consent and complies with relevant laws and policies.

[0104] Furthermore, any software tools or components not belonging to our company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0105] Furthermore, this application boasts the following significant advantages: 1. Strong contextual adaptability: Through a dynamic prompt adjustment mechanism, the system can optimize dialogue content in real time based on changes in customer intent, improving the naturalness of interaction and conversion rate. 2. High knowledge accuracy: The knowledge refinement module can eliminate irrelevant or outdated clauses, significantly improving the accuracy and credibility of insurance knowledge citations. 3. Predictive guidance capability: The entity prediction module can identify potential concerns in advance, proactively guiding customers to advance decision-making and enhancing the system's "predictive ability." 4. Complete learning loop: The system possesses a continuous feedback optimization mechanism, enabling adaptive evolution in multiple scenarios and achieving long-term co-optimization of knowledge and strategies. 5. Adaptability to multiple financial insurance scenarios: The system can be extended to multiple sub-fields such as life insurance, health insurance, auto insurance, and accident insurance, possessing industry universality and engineering feasibility.

[0106] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0107] It should be emphasized that, to further ensure the privacy and security of the above response data, the response data can also be stored in a blockchain node.

[0108] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0109] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0110] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0111] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0112] Further reference Figure 3 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of an artificial intelligence-based problem-solving device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0113] like Figure 3 As shown, the AI-based problem-solving device 300 described in this embodiment includes: a receiving module 301, a retrieval module 302, a parsing module 303, a prediction module 304, a processing module 305, a generation module 306, and an output module 307. Wherein: The receiving module 301 is used to receive user input question data; The retrieval module 302 is used to perform knowledge retrieval processing on the problem data based on a preset knowledge refinement module to obtain the corresponding knowledge package; The parsing module 303 is used to parse and process the question data based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector; The prediction module 304 is used to perform prediction processing on the knowledge package and the semantic state vector based on a preset entity prediction module to obtain the corresponding entity prediction result. The processing module 305 is used to process the entity prediction results based on the preset dynamic prompt adjustment module to generate corresponding target prompt data; The generation module 306 is used to perform decision processing on the question data, the knowledge package and the preset business objectives based on the target prompt data, according to the preset strategy decision module, and generate corresponding response data. The output module 307 is used to output the response data.

[0114] In some optional implementations of this embodiment, the retrieval module 302 includes: The retrieval submodule is used to retrieve knowledge document fragments corresponding to the problem data from a preset multi-source knowledge base based on the knowledge refinement module. The first acquisition submodule is used to acquire dialogue context data corresponding to the question data; The evaluation submodule is used to evaluate the knowledge relevance based on the knowledge document fragments and the dialogue context data, and obtain a knowledge list sorted by relevance. The first processing submodule is used to perform knowledge noise filtering and fusion processing on the knowledge list to obtain processed knowledge data. The generation submodule is used to perform data generation processing on the knowledge data based on a preset knowledge package generation strategy to obtain the corresponding knowledge package.

[0115] In some optional implementations of this embodiment, the generation submodule includes: The first acquisition unit is used to acquire semantic context features and intent category features corresponding to the question data; The second acquisition unit is used to acquire the user's profile features; The integration unit is used to integrate the semantic context features, intent category features, profile features and knowledge data based on a preset dynamic knowledge fusion device to obtain corresponding integrated data. A determining unit is used to treat the integrated data as the knowledge package.

[0116] In some optional implementations of this embodiment, the parsing module 303 includes: The extraction submodule is used to perform semantic embedding extraction on the question data based on the dialogue semantic modeling module to obtain the corresponding current semantic vector; The second processing submodule is used to process the current semantic vector and the corresponding historical semantic vector based on a preset cross-wheel attention mechanism to obtain the corresponding fused semantic representation data. The second acquisition submodule is used to acquire preset semantic memory units; The update submodule is used to update the semantic memory unit based on the fused semantic representation data to obtain the corresponding update data; The first determining submodule is used to use the updated data as the semantic state vector.

[0117] In some optional implementations of this embodiment, the prediction module 304 includes: The first calling submodule is used to call a preset dual-channel network structure based on the entity prediction module; The prediction submodule is used to perform prediction processing on the knowledge package and the semantic state vector based on the dual-channel network structure to obtain the corresponding knowledge features and semantic features. The fusion submodule is used to fuse the knowledge features and the semantic features based on a preset gating mechanism to obtain the corresponding fused features; The classification submodule is used to classify the fused features based on a preset classifier to obtain the corresponding classification result; The second determining submodule is used to use the classification result as the entity prediction result.

[0118] In some optional implementations of this embodiment, the processing module 305 includes: The second calling submodule is used to call the adaptive generator based on the dynamic prompt adjustment module; The third submodule is used to retrieve preset prompt templates; The adjustment submodule is used to perform content adjustment processing on the prompt template based on the adaptive generator, according to the entity prediction result and the semantic state vector, to obtain the adjusted first prompt template. The optimization submodule is used to optimize the first prompt template based on a preset adaptation strategy to obtain the corresponding second prompt template. The third determining submodule is used to use the second prompt template as the target prompt data.

[0119] In some optional implementations of this embodiment, the generation module 306 includes: The fourth acquisition submodule is used to acquire the corresponding semantic confidence based on the semantic state vector; The fifth acquisition submodule is used to acquire the response generation strategy corresponding to the semantic confidence based on the strategy decision module; The decision submodule is used to perform decision processing on the question data, the knowledge package, and the business objective based on the response generation strategy and the target prompt data, and generate corresponding decision data. The fourth determining submodule is used to use the decision data as the response data.

[0120] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0121] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0122] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0123] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for problem-solving methods based on artificial intelligence. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0124] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions of the artificial intelligence-based problem-solving method.

[0125] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0126] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the artificial intelligence-based problem-solving method described above.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0128] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A problem-solving method based on artificial intelligence, characterized in that, Includes the following steps: Receive user input for questions; The problem data is processed by a preset knowledge refinement module to obtain a corresponding knowledge package. The question data is parsed and processed based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector; The knowledge package and the semantic state vector are predicted based on the preset entity prediction module to obtain the corresponding entity prediction results. The entity prediction results are processed by a preset dynamic prompt adjustment module to generate corresponding target prompt data. Based on the preset strategy decision-making module, the system performs decision processing on the question data, the knowledge package, and the preset business objectives according to the target prompt data, and generates corresponding response data. The response data is then processed for output.

2. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of performing knowledge retrieval processing on the problem data based on a preset knowledge refinement module to obtain a corresponding knowledge package specifically includes: Based on the knowledge refinement module, knowledge document fragments corresponding to the problem data are retrieved from a preset multi-source knowledge base; Obtain the dialogue context data corresponding to the question data; Based on the knowledge document fragments and the dialogue context data, a knowledge relevance assessment is performed to obtain a knowledge list sorted by relevance. The knowledge list is subjected to knowledge noise filtering and fusion processing to obtain processed knowledge data; The knowledge data is processed based on a preset knowledge package generation strategy to obtain the corresponding knowledge package.

3. The problem-solving method based on artificial intelligence according to claim 2, characterized in that, The step of generating and processing the knowledge data based on a preset knowledge package generation strategy to obtain a corresponding knowledge package specifically includes: Obtain the semantic context features and intent category features corresponding to the problem data; Obtain the user's profile features; The semantic context features, intent category features, profile features, and knowledge data are integrated and processed based on a preset dynamic knowledge fusion processor to obtain corresponding integrated data. The integrated data is used as the knowledge package.

4. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of parsing and processing the question data based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector specifically includes: Based on the dialogue semantic modeling module, semantic embedding and extraction are performed on the question data to obtain the corresponding current semantic vector; The current semantic vector and the corresponding historical semantic vector are processed based on a preset cross-wheel attention mechanism to obtain the corresponding fused semantic representation data; Obtain the preset semantic memory unit; The semantic memory unit is updated based on the fused semantic representation data to obtain the corresponding updated data; The updated data is used as the semantic state vector.

5. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of performing prediction processing on the knowledge package and the semantic state vector based on the preset entity prediction module to obtain the corresponding entity prediction result specifically includes: The entity prediction module invokes a preset dual-channel network structure. Based on the dual-channel network structure, the knowledge package and the semantic state vector are predicted and processed to obtain the corresponding knowledge features and semantic features. The knowledge features and semantic features are fused based on a preset gating mechanism to obtain the corresponding fused features; The fused features are classified based on a preset classifier to obtain the corresponding classification results; The classification result is used as the entity prediction result.

6. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of processing the entity prediction results based on the preset dynamic prompt adjustment module to generate corresponding target prompt data specifically includes: The adaptive generator is invoked based on the dynamic prompt adjustment module. Get the preset prompt template; Based on the adaptive generator, the prompt template is adjusted according to the entity prediction result and the semantic state vector to obtain the adjusted first prompt template. The first prompt template is optimized based on a preset adaptation strategy to obtain the corresponding second prompt template; Use the second prompt template as the target prompt data.

7. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The preset strategy decision-making module, in its steps of processing the question data, the knowledge package, and the preset business objectives based on the target prompt data to generate corresponding response data, specifically includes: Obtain the corresponding semantic confidence based on the semantic state vector; Based on the strategy decision module, a response generation strategy corresponding to the semantic confidence is obtained; Based on the response generation strategy, decision-making data is generated by processing the question data, the knowledge package, and the business objective according to the target prompt data. The decision data is used as the response data.

8. A problem-solving device based on artificial intelligence, characterized in that, include: The receiving module is used to receive user-input question data; The retrieval module is used to perform knowledge retrieval processing on the problem data based on a preset knowledge refinement module to obtain the corresponding knowledge package; The parsing module is used to parse and process the question data based on the preset dialogue semantic modeling module to obtain the corresponding semantic state vector; The prediction module is used to perform prediction processing on the knowledge package and the semantic state vector based on a preset entity prediction module to obtain the corresponding entity prediction result; The processing module is used to process the entity prediction results based on the preset dynamic prompt adjustment module to generate corresponding target prompt data; The generation module is used to perform decision processing on the question data, the knowledge package, and the preset business objectives based on the preset strategy decision module and the target prompt data, and generate corresponding response data. The output module is used to process the output of the response data.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the problem-solving method based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the problem-solving method based on artificial intelligence as described in any one of claims 1 to 7.

Images