Medical insurance field document intelligent question-answering system based on AI large model

Abstract

The invention discloses a medical insurance field document intelligent question-answering system and method based on an AI large model, and belongs to the technical field of artificial intelligence. The technical problems that an existing medical insurance question-answering system depends on manual configuration and is poor in flexibility and low in intelligent level are solved. According to the technical scheme, the system comprises a document processing module for analyzing and segmenting original documents such as medical insurance policy documents, vectorizing the original documents and storing the original documents into a vector database; and the knowledge question and answer module is used for analyzing user questions by using an AI large model, driving AIAgent to retrieve related knowledge in a vector database, performing reasoning in combination with injected professional knowledge such as business dictionaries and analysis logic, and generating accurate and traceable answers. Through context learning, logical reasoning and automatic generation ability of a large model, tedious intention model training and manual QA configuration are not needed, the intelligent level, adaptability and user experience of a medical insurance question-answering system are remarkably improved, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to an intelligent question-and-answer system for medical insurance documents based on a large AI model. Background Technology

[0002] With the development of information technology, artificial intelligence (AI) is being applied more and more widely in various industries. In the field of medical insurance, due to the complexity of policies, the abundance of documents, and the detailed regulations, insured individuals and related staff face many inconveniences when searching for information. Traditional document retrieval methods are inefficient and cannot guarantee the timeliness and accuracy of information. Therefore, AI-based intelligent question-answering systems have emerged.

[0003] Early medical insurance question-answering systems mostly used simple machine learning algorithms. These algorithms heavily relied on manual feature engineering, requiring domain experts to manually design features and label tens of thousands of intent data points to train the intent classification model. This process was time-consuming, labor-intensive, and costly, and limited the system's scalability and flexibility. More importantly, these models had poor generalization ability and were prone to misjudging intent when faced with new or complex questions outside the preset scope, leading to incorrect answers and failing to meet the needs of practical applications.

[0004] Existing mainstream medical insurance chatbots are usually built on a task-oriented basis, and their core technologies are intent management, slot management, and answer configuration.

[0005] In recent years, AI large-scale model technology has developed rapidly, and its powerful semantic understanding, logical reasoning, and text generation capabilities have provided new opportunities to solve the aforementioned problems. However, how to effectively transfer the general capabilities of large-scale models to the highly specialized and data-sensitive field of medical insurance, and build a truly intelligent, efficient, and easy-to-maintain question-answering system, is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0006] Purpose of the invention: To provide an intelligent question-and-answer system for medical insurance documents based on a large AI model, in order to solve the above-mentioned problems existing in the prior art.

[0007] Technical solution: A document intelligent question-answering system for the medical insurance field based on an AI large model, comprising two main parts: a document processing module and a knowledge question-answering module.

[0008] The document processing module is used to process the original medical insurance documents, and the document processing module includes:

[0009] The document parsing unit is used for text recognition, layout parsing, and knowledge extraction of medical insurance documents in various formats.

[0010] The document segmentation unit is used to segment the parsed document into multiple document blocks based on a strategy that combines semantics and structure;

[0011] The vectorized storage unit is used to convert the document block data into a vector representation and store it in a vector database, while also creating an index for the document.

[0012] The knowledge-based question-and-answer module is used to respond to user questions and generate answers. The knowledge-based question-and-answer module includes:

[0013] The problem analysis unit is used to perform semantic understanding, intent recognition, and logical reasoning on user questions using large AI models.

[0014] The AIAgent response unit is used to vectorize the user's question based on the question analysis results, perform similarity retrieval in the vector database to locate relevant knowledge fragments, and call preset tools and injected business knowledge to integrate and reason about the retrieval results to generate a preliminary answer.

[0015] The answer generation and tracing unit is used to convert the preliminary answer into natural language text and provide tracing information on the original document and paragraph position on which the answer is based.

[0016] In a further embodiment, the document parsing unit performs two parsing steps during layout parsing: physical layout parsing and logical layout parsing. Physical layout parsing is used to parse the document's page size, margins, headers, footers, and the position and size of text boxes to support accurate answer location and tracing. Logical layout parsing is used to parse the relationships and hierarchical structure between elements such as titles, paragraphs, lists, and tables to assist the system in understanding the document's content structure.

[0017] In a further embodiment, the document segmentation unit adopts a strategy that combines semantic segmentation and structural segmentation. First, it performs preliminary segmentation according to the hierarchical structure of the document, such as chapters and sections. Then, it performs secondary segmentation on the segmented document blocks according to the semantic coherence algorithm to ensure that each document block is relatively independent and complete in semantics.

[0018] In a further embodiment, the AlAgent response unit is further optimized through prompt word engineering, which includes constructing a prompt word framework based on medical insurance business scenarios and incorporating medical insurance policies, regulations and professional terminology into the prompt words.

[0019] In a further embodiment, the knowledge question-answering module also includes a knowledge injection unit. The knowledge injection unit injects professional knowledge from the medical insurance field into the system to improve the accuracy of question answering. The professional knowledge comprises three components: a business dictionary, complex problem analysis logic, and calculation formulas. The business dictionary covers professional terms, abbreviations, and synonyms. The complex problem analysis logic embeds the judgment logic for medical insurance reimbursement, qualification certification, and other business processes into the system in the form of rules. The calculation formulas integrate precise calculation models for medical insurance reimbursement and payment calculation.

[0020] In a further embodiment, the knowledge-based question-and-answer module also includes an answer quality assessment module. This module is used to batch upload frequently asked user questions and standard answers, and an AI model automatically scores the accuracy of the system-generated answers to assist the system in optimizing its selection process.

[0021] In a further embodiment, the question-answering method of the AI-based large-scale model-based intelligent question-answering system for medical insurance documents includes the following steps:

[0022] S1. Parse the original medical insurance documents and perform text recognition, layout parsing, and knowledge extraction.

[0023] S2. The parsed document is divided into multiple document blocks based on a strategy that combines semantics and structure;

[0024] S3. Convert the document block data into a vector representation, store it in a vector database, and create an index for the document;

[0025] S4. Receive user questions and use AI big data models to perform semantic understanding, intent recognition and logical reasoning on user questions;

[0026] S5. Start AIAgent, vectorize the user's question based on the question analysis results, and perform similarity retrieval in the vector database to locate relevant knowledge fragments;

[0027] S6. The AlAgent calls preset tools and injected business knowledge to integrate and reason about the search results to generate a preliminary answer;

[0028] S7. Convert the preliminary answer into natural language text, and provide the source information of the original document and paragraph position on which the answer is based, and return it to the user.

[0029] In a further embodiment, a knowledge injection step is included before the knowledge question-answering step. This knowledge injection step injects professional knowledge from the medical insurance field into the system by constructing a business dictionary, embedding complex problem analysis logic, and integrating calculation formulas.

[0030] In a further embodiment, the knowledge-based question-answering step further includes an answer quality assessment step. This step involves batch question-answering evaluation, where a large AI model scores the accuracy of the system-generated answers and adjusts the system based on the scoring results.

[0031] In a further embodiment, the method supports both text and voice question-and-answer modes. When receiving a voice question, it first converts the question into text using speech recognition technology before proceeding with subsequent steps.

[0032] Beneficial Effects: This invention, a medical insurance document intelligent question-answering system based on an AI large-scale model, has the following significant advantages:

[0033] 1. High degree of automation, reduced operation and maintenance costs: This invention utilizes the powerful capabilities of large models. It only requires injecting medical insurance professional knowledge documents into the system to realize end-to-end question and answer functions. It eliminates the need for time-consuming and labor-intensive manual intent annotation, model training and QA pair configuration in traditional methods, which greatly reduces manpower and time costs.

[0034] 2. High level of intelligence, enhancing user experience: This invention can deeply understand the context and complex intentions of user questions, perform logical reasoning, and generate natural, accurate, and coherent answers. The traceability of the answers also enhances users' trust in the system, significantly improving the user experience.

[0035] 3. High flexibility and adaptability: In the face of frequent changes in medical insurance policies, this invention only needs to update the documents in the knowledge base, and the system can learn new knowledge in real time and quickly adapt to changes without the need for large-scale adjustments to the system configuration, thus ensuring the timeliness and accuracy of information.

[0036] 4. Accurately inject professional knowledge to ensure the reliability of answers: By injecting business dictionaries, analytical logic and calculation formulas, the shortcomings of general large models in professional domain knowledge are effectively made up for, reducing the "illusion" phenomenon and ensuring the professionalism and reliability of Q&A in the medical insurance field. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the data flow structure of the present invention. Detailed Implementation

[0038] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0039] The intelligent question-and-answer system for medical insurance documents based on AI large-scale models involved in this invention mainly includes a document processing module and a knowledge question-and-answer module.

[0040] The document processing module is used to process raw medical insurance documents and automatically build a knowledge base. This module includes: a document parsing unit for text recognition, physical and logical layout parsing, and knowledge extraction of medical insurance documents in various formats such as JPG, PDF, PPT, and DOC; a document segmentation unit for segmenting the parsed documents into multiple independent, semantically complete document blocks based on a strategy combining semantics and structure; and a vectorized storage unit for converting document block data into high-dimensional vector representations using a deep learning model and storing them in a vector database (such as Milvus), while simultaneously creating an index for the documents (such as Elasticsearch).

[0041] The knowledge-based question-answering module is used to respond to user questions and generate answers. This module includes: a question analysis unit, which utilizes the natural language processing capabilities of the AI ​​model to perform deep semantic understanding, intent recognition, and logical reasoning on user questions; an AIAgent response unit, serving as the system's intelligent core, which, based on the question analysis results, vectorizes the user question and performs efficient similarity retrieval in a vector database to locate the most relevant knowledge fragments. Simultaneously, this unit invokes pre-set tools (such as computational tools) and business knowledge injected through prompt word engineering (such as business dictionaries and analytical logic) to integrate, filter, and reason about the retrieved information, generating a logically rigorous preliminary answer; and an answer generation and tracing unit, which transforms the preliminary answer into fluent, natural language text using the large model's generation capabilities, and provides tracing information such as the original document name and paragraph position on which the answer is based, to enhance the credibility of the answer.

[0042] Furthermore, the system may also include a knowledge injection unit, which injects professional knowledge in the medical insurance field, such as business dictionaries, complex problem analysis logic, calculation formulas, etc., into the system in a structured manner, effectively alleviating the "illusion" problem of large models and ensuring the professionalism and accuracy of the answers.

[0043] Furthermore, the system may also include an answer quality assessment module, which automatically scores the accuracy of system-generated answers by uploading "question-standard answer" pairs in batches and using a large model, thereby quickly identifying problems, guiding the correction of prompt word engineering and the supplementation of business knowledge, and forming an efficient closed-loop optimization mechanism.

[0044] The technical solution of this invention mainly includes two modules: a document processing process and a knowledge question-and-answer process, as detailed below:

[0045] I. The document processing process aims to transform the original document into a highly usable format, and mainly includes the following key steps:

[0046] Step 1: Manage medical insurance data, supporting document parsing engines.

[0047] This step involves comprehensive management of medical insurance data, encompassing the processing of numerous medical insurance policy documents, medical guidelines, and reimbursement details. Specifically, it supports uploading these documents in formats such as JPG, PNG, PDF, PPT, DOC, and DOCX. An OCR model is used to recognize the text, followed by a layout model to determine the positional coordinates of the recognized text, facilitating later tracing of the original document for answers. Finally, an extraction model automatically extracts knowledge.

[0048] In terms of parsing, there are mainly two methods: physical layout parsing and logical layout parsing. Among them, physical layout parsing focuses on the visual presentation elements of the document, such as the size of the document page, page margins, headers and footers, and the position and size of text boxes. By parsing these elements, it can provide effective support for the accurate location and tracing of document paragraphs.

[0049] Logical layout analysis focuses on the relationships and hierarchical structure between various elements within a document, involving the analysis of headings, paragraphs, lists, tables, etc. Its main purpose is to help the system deeply understand the content and structure of the document so that it can smoothly carry out a series of important tasks such as text parsing, data extraction, and semantic analysis.

[0050] Step 2: Perform document block segmentation

[0051] The process of dividing a complete document into multiple smaller parts or blocks. These blocks can be divided based on different criteria and strategies, aiming to break down large, complex documents into more manageable units to meet subsequent data processing, storage, and retrieval needs.

[0052] This invention employs a strategy combining semantic and structural segmentation. First, it segments the document according to its hierarchical structure, progressively refining it from high-level chapters to lower-level sections or subsections. For example, a book can be divided into different levels of document blocks based on its chapter, section, and subsection hierarchy. This allows for different levels of information management and processing, progressively refining information from the overall outline to detailed content, facilitating hierarchical storage and retrieval of documents. Second, considering semantic coherence and logical relationships, the document is divided into semantically coherent blocks. For instance, based on a semantic coherence algorithm, the document is segmented when semantic coherence falls below a certain threshold. In a product manual, if a section introduces a new function or feature but lacks semantic coherence with preceding sections, it is divided into a new document block, ensuring relative semantic coherence within each block, which aids in subsequent semantic analysis and processing.

[0053] Step 3: Perform vector database storage (Milvus) and document indexing (ES).

[0054] In the vector database storage (Milvus) stage, the first step is to extract features from the preprocessed document block data and transform it into a vector representation using a deep learning model. For example, word embedding technology can be used to convert text blocks into word vectors, or convolutional neural networks can be used to extract feature vectors from image document blocks. This accurately captures the semantic and feature information of the document blocks, laying a solid foundation for subsequent storage and retrieval.

[0055] Subsequently, using the Milvus Python SDK, a connection was established with the Milvus service. During this process, key parameters such as host address and port number were configured appropriately based on the actual deployment environment to ensure stable communication with the Milvus server. Simultaneously, necessary parameter settings were configured for Milvus based on data characteristics and application requirements, such as carefully selecting the storage engine (memory storage or disk storage) to optimize storage performance and resource utilization efficiency.

[0056] Next, within Milvus, collection attributes are defined according to business requirements, including name, vector dimensions, and data type (such as floating-point vectors), to create the collection. As the logical unit for data storage, the collection, once created, inserts the generated vector data into the corresponding collection in batches. Milvus possesses efficient batch vector insertion capabilities, quickly handling large-scale vector data storage tasks and effectively ensuring data integrity and consistency.

[0057] Finally, to improve vector retrieval efficiency, an index needs to be built on the collection after data insertion. During the index building process, index parameters, such as quantization factor and number of levels, need to be flexibly adjusted according to data characteristics to achieve optimal retrieval performance.

[0058] II. The knowledge-based question-and-answer process is a comprehensive information processing workflow designed to accurately understand user questions and provide precise answers based on given knowledge sources. It mainly includes the following key steps:

[0059] The first step is for the large model to analyze the user's question. This is a complex and meticulous process that integrates various natural language processing techniques and the model's internal mechanisms, mainly including text preprocessing, semantic understanding, intent recognition, and logical reasoning.

[0060] (1) Text preprocessing

[0061] The large model first performs text preprocessing on the input question. This includes lexical analysis, which segments the question text into individual words or phrases and identifies the part of speech of each phrase, such as nouns, verbs, and adjectives, providing basic lexical information for subsequent analysis. Next, syntactic analysis is performed to construct the grammatical structure tree of the question, clarifying the dependency relationships between various components in the sentence, such as subject-verb-object, attributive, adverbial, and complement relationships, helping to understand the basic structure of the sentence. Furthermore, normalization processing is performed, such as converting the text to a uniform capitalization, removing extra spaces, and handling special characters, ensuring that the input text is formatted correctly for easy model processing.

[0062] (2) Semantic understanding

[0063] Based on the preprocessed results, the large model leverages the linguistic knowledge and semantic representation capabilities learned during pre-training to understand the semantics of the question. The model captures the semantic information behind the words by matching and associating them with its vast knowledge system. For example, using word vector technology, each word is mapped to a low-dimensional vector space, and the semantic relevance of words is measured by the distance and similarity between vectors. Simultaneously, the model can combine contextual information to understand the meaning of words in specific contexts, eliminating ambiguity. For example, for polysemous words, the model can determine their accurate semantics based on the context of the question.

[0064] (3) Intent recognition

[0065] Building upon semantic understanding, the large model further identifies the user's intent in the question. This involves determining the question type, such as whether it is a factual question (asking for specific factual information), an explanatory question (seeking an explanation of a concept or phenomenon), or a suggestive question (expecting a certain suggestion or solution). The model analyzes the semantic features of the question, keywords, and a large number of text patterns learned during pre-training to determine the user's purpose in asking the question and the type of answer they expect. For example, for questions containing keywords such as "why" or "what is the reason," the model tends to identify them as explanatory questions.

[0066] (4) Logical reasoning

[0067] For some complex problems, large models may need to perform logical reasoning. The model utilizes the logical knowledge and reasoning rules learned during pre-training to logically analyze and deduce information from the problem. For example, when dealing with problems involving multiple conditions and logical relationships, the model can identify logical connectors between conditions (such as "if...then...", "and", "or", etc.) and apply corresponding logical reasoning methods to reach a conclusion. This process is similar to how humans think when solving logical problems, but the model performs automatic reasoning based on its learned patterns and parameters.

[0068] Step two, AIAgent obtains the optimal answer.

[0069] Subsequently, based on its understanding of the problem, the Agent initiates the knowledge retrieval module. According to the problem type and key information, the Agent transforms the problem into vector form. By calculating the similarity between vectors, it quickly locates relevant knowledge fragments in a vector database storing a large amount of medical insurance information (vector database storage (Milvus) and document indexing (ES) are performed in step three of the document processing process), and returns the retrieval results.

[0070] The connection between AIAgent and data sources and large models mainly includes the following steps:

[0071] (1) Select the preset Tool;

[0072] (2) Configure the values ​​of the configuration item variables, and perform the initial configuration and optimization settings for the prompt word project;

[0073] Based on business scenarios, a prompt word framework is constructed: A targeted prompt word framework is designed according to different medical insurance business scenarios and user intent. For example, for a scenario involving medical insurance reimbursement ratio inquiries, the prompt word framework could be designed as: "Suppose a user wants to know the medical insurance reimbursement ratio for [specific expense items] under [medical treatment type], please answer accurately based on the latest medical insurance policies." In this framework, by clearly indicating key information such as user needs, insurance type, medical treatment type, and expense items, the model is guided to generate an answer around these key points.

[0074] Incorporating policies, regulations, and professional terminology: The prompts appropriately incorporate relevant medical insurance policies, regulations, and professional terminology to ensure the model's generated answers are authoritative and professional. For example, prompts related to the scope of medical insurance reimbursement may mention, "According to the provisions of [specific medical insurance policy document name], the scope of medical insurance reimbursement covers items that conform to the 'National Basic Medical Insurance Drug Catalog,' 'Diagnosis and Treatment Item Catalog,' and 'Medical Service Facility Standards,'" enabling the model to provide accurate policy justification in its responses.

[0075] Analysis based on feedback data: Collect interaction data between users and the medical insurance chatbot, including user questions, chatbot answers, and user feedback on the answers (such as satisfaction ratings, follow-up questions, etc.). Conduct in-depth analysis of this data to identify instances where the chatbot's answers are inaccurate, incomplete, or difficult for users to understand. For example, if multiple users are found to be dissatisfied with the answers regarding the medical insurance out-of-town medical treatment registration process, further analysis should be conducted to determine whether the answers omit key steps or are not clearly and understandably worded.

[0076] Targeted Adjustment of Prompt Terms: Based on the analysis results, adjust the prompt terms accordingly. If it is found that the model does not consider the deduction rules for certain special expenses when answering questions about medical insurance reimbursement calculations, the prompt terms can be explicitly emphasized: "When calculating the medical insurance reimbursement amount, be sure to consider the deduction rules for [special expense items], and calculate according to the following formula...". At the same time, optimize the language of the prompt terms to make them more concise, clear, and easy to understand. For example, explain technical terms in a more straightforward way, or use examples to help the model generate answers that are easier for users to understand.

[0077] (3) Connect vector databases;

[0078] (4) Based on the semantic understanding and problem decomposition of the problem, select the corresponding tools, search for relevant information and data, use code execution tools to find the document paragraphs involved in the problem in the medical insurance document library, and finally integrate these information and results to output the answer.

[0079] While large-scale model-driven intelligent medical insurance dialogue assistants provide convenient services, illusion questions can lead to the transmission of incorrect information, impacting user rights and trust. Metadata definition management helps to organize the sources, characteristics, and relationships of data, providing more comprehensive and accurate background information for answers; a business dictionary standardizes and explains professional terminology in questions and answers, ensuring accurate and consistent terminology usage; and expert analysis logic is used to filter, integrate, and reason about search results, making answers more logically rigorous and insightful. This effectively improves the accuracy and reliability of responses.

[0080] Finally, after acquiring the relevant knowledge, the reasoning and integration module begins its work, organically integrating the scattered knowledge fragments to generate a preliminary answer. The natural language generation module then transforms the preliminary answer into a natural, fluent final answer that conforms to human language expression habits, presenting it to the user.

[0081] Step 3: Answer Quality Assessment Agent. This agent performs batch evaluations of question-and-answer accuracy by uploading frequently asked user questions and answers in batches. The large model automatically provides accuracy scores for the answers, quickly identifies problems, and corrects prompts and injects business knowledge.

[0082] Step four: The answer is returned in real time. The answer supports document tracing and paragraph location, meaning the system not only provides the answer but also clearly identifies the original document or information source upon which the answer is based, and accurately locates the specific position of that information within the document. This feature strongly supports the accuracy and reliability of the answer. The specific principle is as follows:

[0083] Data annotation and index construction: In the data preprocessing stage, each document in the knowledge base is annotated in detail, recording metadata information such as the document's source, creation time, and author. Simultaneously, text tagging technology is used to mark key information in the documents, such as paragraph topics and important concepts. Based on this, a comprehensive and refined index structure is constructed, including not only traditional keyword indexes but also semantic understanding-based indexes. For example, by converting document content into vector representations, vector similarity calculations are used to achieve more accurate semantic retrieval, enabling the system to quickly locate the original document based on key information in the answer.

[0084] Source information embedding and extraction: During the answer generation process, the system associates the answer with the document information it is based on, embedding document source information into the answer generation logic. Once the answer is generated, it can automatically extract the document source information related to that answer, including the document name, source path, and the answer's specific location within the document (such as page number, paragraph number, etc.). For example, if the answer is generated based on a specific clause in a medical insurance policy document, the system will accurately record the document's name, issuing agency, and the specific location of the clause for user reference.

[0085] User Interaction and Display: To facilitate users' access to the source information of answers, the system provides an intuitive interactive interface. When users have doubts about the reliability of an answer or wish to learn more, they can easily obtain complete document source information through simple operations (such as clicking on the source file in the answer). The system will display this information in a clear and easy-to-understand format, such as presenting the document source as a link. Users can click the link to directly jump to the location of the original document, achieving seamless connection between the answer and the original information and enhancing user trust in the answer.

[0086] This invention innovatively proposes an intelligent question-answering system for medical insurance documents based on a large AI model. This system has the following core features:

[0087] I. Deepening the Application of Medical Insurance Based on Large Model Technology

[0088] Leveraging the superior deep learning and natural language processing capabilities of large models, a series of sophisticated strategies significantly improve the system's question-answering performance in the medical insurance field.

[0089] Enhance semantic understanding: Utilize language patterns and semantic representations learned by large models from massive amounts of text data to deeply analyze user-inputted medical insurance-related questions, accurately capture the core intent and subtle semantic differences of the questions, and effectively avoid comprehension biases caused by semantic ambiguity.

[0090] Integrating expertise in the medical insurance field: This involves deeply integrating the professional knowledge system of the medical insurance field, including but not limited to various medical insurance policies, regulations, business processes, and reimbursement details, with the large-scale model. This approach enables the model to analyze and answer questions more effectively based on this professional knowledge background, ensuring that the generated answers are accurate and consistent with the actual business logic of medical insurance.

[0091] Context Awareness: The system possesses powerful context understanding and memory capabilities, enabling it to accurately grasp the evolution of user intent based on preceding information in multi-turn dialogue scenarios, thereby generating coherent, consistent, and practically tailored responses. This feature is particularly suitable for answering complex questions in the medical insurance field. For example, in multi-step consultations involving medical insurance reimbursement processes, it can provide precise follow-up guidance based on information such as the user's previously mentioned insurance type and medical history.

[0092] Supports multimodal information processing: In addition to traditional text information, the system can also process multimodal data such as images and audio. For example, for medical invoice images required for medical insurance reimbursement, the system can automatically identify key information and assist in answering reimbursement-related questions; for medical insurance inquiries made by users via voice, the system can convert them into text in real time and process them efficiently, providing users with a more convenient and diversified service experience.

[0093] Establish a continuous optimization and feedback loop: By collecting user feedback data on answers, including satisfaction ratings and error correction information, the system continuously optimizes its model parameters and question-answering strategies. Leveraging reinforcement learning algorithms in machine learning, the system adjusts its output based on feedback information, enabling it to continuously improve the accuracy of question understanding and the quality of answer generation over long-term operation, adapting to ever-changing user needs and the healthcare policy environment.

[0094] II. Multi-source data fusion ensures the integrity of medical insurance data.

[0095] This system is dedicated to integrating medical insurance data resources from diverse channels, aiming to ensure the comprehensiveness and timeliness of medical insurance data. Through efficient data collection and cleaning technologies, the system acquires data from multiple data sources, including official websites of medical insurance departments, policy document databases, medical institution information systems, and statistical reports, and then rigorously filters, deduplicates, and standardizes the format. This process not only guarantees the accuracy and consistency of the data but also enables the system to update medical insurance data in a timely manner, reflecting the latest policy developments, business changes, and actual operational status, providing users with intelligent question-and-answer services based on the latest and most comprehensive information.

[0096] III. Precise Injection of Professional Knowledge in the Medical Insurance Field

[0097] To quickly and effectively improve the accuracy of question and answer responses, this system supports the deep integration of professional knowledge in the field of medical insurance.

[0098] Business Dictionary Construction: A meticulously constructed business dictionary covers various professional terms, abbreviations, and synonyms in the medical insurance field. This dictionary not only provides standard definitions of terms but also includes their specific meanings and usage examples in different medical insurance business scenarios. When processing user questions and generating answers, the system standardizes terminology based on the business dictionary to ensure the accuracy and consistency of terminology usage and avoid erroneous answers caused by ambiguity in terminology understanding.

[0099] Embedded complex problem analysis logic: The system embeds the analytical logic of complex problems in the medical insurance field, such as the judgment logic of medical insurance reimbursement process and the reasoning rules for determining eligibility for insurance, into the system in the form of algorithmic models or rule engines. When faced with complex medical insurance problems, the system can break down and analyze the problem step by step according to the preset analytical logic, and generate logically rigorous, accurate and reliable answers by combining relevant data and knowledge.

[0100] Integrated Calculation Formulas: The system integrates precise calculation formulas for various calculation problems related to medical insurance, such as reimbursement amount calculation and payment amount accounting. These formulas strictly adhere to medical insurance policies and are flexibly adapted to different insurance types, medical treatment situations, and expense items. When answering related questions, the system can accurately call the corresponding formula, combine it with specific data to perform calculations, and provide users with accurate calculation results and detailed explanations of the calculation process.

[0101] Metadata Settings and Management: The system provides detailed settings and effective management of metadata for medical insurance data, including information such as data source, creation time, update frequency, and data owner. Through clear metadata annotation, the system can better track data reliability and timeliness, prioritizing authoritative and up-to-date data sources when processing issues, ensuring that generated answers are based on credible and valid information.

[0102] IV. The intelligent question-and-answer system achieves a seamless interactive experience.

[0103] Utilizing cutting-edge natural language processing technology, this system successfully establishes a seamless intelligent communication bridge between users and the system. The system fully supports various convenient question-and-answer modes, including voice and text, to fully respond to and meet the diverse needs of insured individuals.

[0104] Text-based Q&A: Users can input questions in text format, and the system can quickly understand the intent of the question and generate accurate, detailed, and easy-to-understand text answers in a short time. At the same time, the system provides rich text editing and interactive functions, such as question correction and follow-up questions, facilitating in-depth information exchange between users and the system.

[0105] Voice Q&A: Leveraging advanced speech recognition and synthesis technology, users can ask questions related to medical insurance via voice input. The system converts the speech into text in real time, processes it, and provides answers in a clear and natural voice format. This interaction method is particularly suitable for users in scenarios where typing is inconvenient, such as while driving or doing housework, allowing them to more conveniently access medical insurance information and greatly improving the efficiency and convenience of accessing services.

[0106] In a further preferred embodiment, traditional question-answering robots typically rely on intent classification models for training, requiring significant manpower to summarize knowledge question-and-answer pairs (QA pairs) related to medical insurance policies and to perform tedious custom configuration of intent-related answers. However, this invention, based on advanced large-scale model technology, fully utilizes its powerful context learning capabilities, excellent logical reasoning abilities, and smooth automatic generation capabilities, greatly simplifying the construction process of the question-answering system.

[0107] In this invention, by effectively integrating professional knowledge from the medical insurance field and optimizing the prompt word engineering based on real business scenarios, a highly efficient question-and-answer function for users can be achieved. This innovative model effectively overcomes the shortcomings of traditional question-and-answer robots in terms of flexibility and avoids a large amount of manual configuration work. In this way, this invention not only significantly improves the system's intelligence level but also brings users a more user-friendly and convenient experience, making it more adaptable and practical in real-world application scenarios.

[0108] In a further embodiment, a province had a medical insurance hotline with over twenty operators (working in four shifts, providing 24-hour human service), each making an average of 70 calls per day, totaling approximately 1400+ calls per day. In July of the same year, after embedding a large-scale model-based medical insurance document Q&A robot into the province's medical insurance WeChat official account, the call volume decreased from an average of 70 calls per operator per day to 30. The Q&A robot provides uninterrupted service 24 / 7, automating a large number of repetitive consultation tasks, quickly responding to user inquiries, effectively alleviating the problem of manpower shortages, reducing the pressure on human customer service, and saving human resource costs. Standardized consultation services reduce individualized interpretation differences, ensuring that all insured individuals receive consistent information. Transparent service processes enhance public trust in the medical insurance system. Simultaneously, collecting and analyzing user consultation data provides a scientific basis for the formulation and optimization of medical insurance policies.

[0109] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. An intelligent question-answering system for medical insurance documents based on an AI large-scale model, characterized by: include: A document processing module is used to process original medical insurance documents. The document processing module includes: The document parsing unit is used for text recognition, layout parsing, and knowledge extraction of medical insurance documents in various formats. The document segmentation unit is used to segment the parsed document into multiple document blocks based on a strategy that combines semantics and structure; The vectorized storage unit is used to convert the document block data into a vector representation and store it in a vector database, while also creating an index for the document. A knowledge-based question-and-answer module is used to respond to user questions and generate answers. This module includes: The problem analysis unit is used to perform semantic understanding, intent recognition, and logical reasoning on user questions using large AI models. The AIAgent response unit is used to vectorize the user's question based on the question analysis results, perform similarity retrieval in the vector database to locate relevant knowledge fragments, and call preset tools and injected business knowledge to integrate and reason about the retrieval results to generate a preliminary answer. The answer generation and tracing unit is used to convert the preliminary answer into natural language text and provide tracing information on the original document and paragraph position on which the answer is based.

2. The intelligent question-answering system for medical insurance documents based on an AI large-scale model as described in claim 1, characterized in that, The document parsing unit, when performing page layout parsing, includes: Physical layout analysis is used to analyze the page size, margins, headers, footers, and text box positions and sizes of a document to support accurate answer location and tracing. Logical layout parsing is used to analyze the relationships and hierarchical structure between elements such as headings, paragraphs, lists, and tables in a document, in order to help the system understand the content and structure of the document.

3. The intelligent question-answering system for medical insurance documents based on an AI large model as described in claim 1, characterized in that: The document segmentation unit adopts a strategy that combines semantic segmentation and structural segmentation. First, it performs preliminary segmentation according to the hierarchical structure of the document, such as chapters and sections. Then, it performs secondary segmentation on the segmented document blocks according to the semantic coherence algorithm to ensure that each document block is relatively independent and complete in semantics.

4. The intelligent question-answering system for medical insurance documents based on an AI large model as described in claim 1, characterized in that: The AlAgent response unit is further optimized through prompt word engineering, which includes building a prompt word framework based on medical insurance business scenarios and incorporating medical insurance policies, regulations and professional terminology into the prompt words.

5. The intelligent question-answering system for medical insurance documents based on an AI large-scale model according to claim 1, characterized in that, The knowledge-based question-and-answer module also includes: The knowledge injection unit is used to inject professional knowledge in the field of medical insurance into the system to improve the accuracy of question and answer. The expertise mentioned includes: A business dictionary covering technical terms, abbreviations, and synonyms; Complex problem analysis logic embeds the judgment logic of business such as medical insurance reimbursement and qualification certification into the system in the form of rules; The calculation formula integrates precise calculation models for medical insurance reimbursement and payment accounting.

6. The intelligent question-answering system for medical insurance documents based on an AI large-scale model according to claim 1, characterized in that, The knowledge-based question-and-answer module also includes: The answer quality assessment unit is used to upload common user questions and standard answers in batches. The AI ​​model automatically scores the accuracy of the system-generated answers to assist the system in optimizing the selection process.

7. The intelligent question-answering system for medical insurance documents based on an AI large model according to claim 1, characterized in that: The question-answering method of the AI-based large-scale model-based intelligent question-answering system for medical insurance documents includes the following steps: S1. Parse the original medical insurance documents and perform text recognition, layout parsing, and knowledge extraction. S2. The parsed document is divided into multiple document blocks based on a strategy that combines semantics and structure; S3. Convert the document block data into a vector representation, store it in a vector database, and create an index for the document; S4. Receive user questions and use AI big data models to perform semantic understanding, intent recognition and logical reasoning on user questions; S5. Start AIAgent, vectorize the user's question based on the question analysis results, and perform similarity retrieval in the vector database to locate relevant knowledge fragments; S6. The AlAgent calls preset tools and injected business knowledge to integrate and reason about the search results to generate a preliminary answer; S7. Convert the preliminary answer into natural language text, and provide the source information of the original document and paragraph position on which the answer is based, and return it to the user.

8. The intelligent question-answering system for medical insurance documents based on an AI large model according to claim 7, characterized in that, Before the knowledge-based question-and-answer step, it also includes: The knowledge injection step involves injecting professional knowledge from the medical insurance field into the system by constructing a business dictionary, embedding complex problem analysis logic, and integrating calculation formulas.

9. A medical insurance document intelligent question-answering system based on an AI large model according to claim 7, characterized in that, The knowledge-based question-and-answer process also includes: The answer quality assessment process involves batch question-and-answer evaluation, where a large AI model scores the accuracy of the system-generated answers and makes corrections based on the scoring results.

10. The intelligent question-answering system for medical insurance documents based on an AI large model as described in claim 7, characterized in that: The method supports both text and voice question-and-answer modes. When receiving a voice question, it first converts the question into text using speech recognition technology before proceeding with subsequent steps.

Images