Large model-based data retrieval enhancement generation method in medical insurance field

Abstract

The invention discloses a data retrieval enhancement generation method in the medical insurance field based on a large model, and aims to improve the accuracy and efficiency of medical insurance data retrieval. Comprising the following steps: cleaning, standardizing and marking medical insurance data; training a large model by using the preprocessed data, analyzing the natural language query of the user, and generating a corresponding problem vector; according to the generated question vector, obtaining candidate texts highly related to the question from a knowledge base by adopting a dynamic hybrid retrieval method, and performing fine rearrangement and reverse rearrangement on the candidate texts; the rearranged text is sent into the large model, and a final output result is generated in combination with generation rule and output rule constraint conditions; optimizing a retrieval result by combining an interactive attention mechanism based on # imgabs0 # with a self-adaptive weight allocation method; according to the method, the query intention of the user can be reflected more accurately, the correlation and accuracy of the retrieval result are improved, and various data retrieval requirements in the medical insurance field are met.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and natural language processing technology, and in particular to a data retrieval enhancement generation method in the field of medical insurance based on a large model. Background Art

[0002] In the field of modern medicine and medical insurance, accurate data retrieval and processing are important foundations for achieving efficient medical services and precision medicine. However, with the development of medical informatization and the explosive growth of medical data, how to effectively retrieve accurate and relevant information from massive medical insurance data has become a technical problem that needs to be solved urgently. Existing medical insurance data retrieval methods mainly rely on keyword matching and traditional information retrieval technology. These methods perform well when processing simple queries, but they are often unable to cope with complex natural language queries and medical data that requires deep understanding.

[0003] Traditional keyword matching methods cannot understand the semantics of user queries and can only perform simple matching based on keywords, resulting in low accuracy of retrieval results. Especially when it comes to professional medical terms and complex sentences, the retrieval results are often unsatisfactory. Keyword matching methods cannot distinguish between synonyms and polysemous words, nor can they handle complex query intent. For example, when a user enters "the latest treatment for diabetes", traditional methods may only return text containing "diabetes" or "treatment", but cannot understand the user's real needs, resulting in inaccurate results. Most existing information retrieval methods lack the ability to deeply understand natural language semantics and cannot effectively parse the true intent of user queries. The processing capabilities for synonyms, polysemous words and contextual associations are weak, resulting in poor relevance of retrieval results. For example, "heart disease" and "coronary heart disease" are closely related in medicine, but traditional retrieval may not be able to associate the two, thus missing important related information.

[0004] As medical data continues to grow, traditional information retrieval methods are inefficient when processing large-scale data and cannot quickly respond to user query needs. This is particularly unfavorable for medical scenarios with high real-time requirements. When processing millions or even hundreds of millions of data, the response speed of existing technologies is significantly reduced, and it cannot meet the needs of clinical decision-making and real-time consultation. Existing retrieval methods generally lack recommendation functions and cannot provide customized retrieval results based on users' historical queries and preferences, resulting in poor user experience. Medical information retrieval requires results based on the specific needs of doctors or patients. For example, doctors may focus on the latest research progress on a specific disease, while patients may be more concerned about treatment plans and rehabilitation advice. Existing methods cannot effectively make recommendations based on users' historical behaviors and preferences, causing users to spend a lot of time screening information.

[0005] Medical information is distributed in different databases and lacks a unified means of retrieval and integration, resulting in a serious problem of information islands. When doctors and patients retrieve information, they often need to query across multiple platforms, which is not only time-consuming and labor-intensive, but also easy to miss important information. Existing technologies cannot effectively integrate and unify access to data from different sources, limiting the comprehensiveness and availability of information. The inconsistent formats and standards of medical data from different sources increase the complexity of data processing. When performing data preprocessing, existing technologies face problems such as diverse data formats and inconsistent standards, which makes the data cleaning and standardization process time-consuming and labor-intensive, affecting the accuracy and efficiency of retrieval.

[0006] Therefore, how to provide a data retrieval enhancement generation method in the medical insurance field based on a large model is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] One purpose of the present invention is to propose a data retrieval enhancement generation method in the field of medical insurance based on a large model. The present invention makes full use of advanced artificial intelligence technology and natural language processing technology, and describes in detail the whole process from data preprocessing, large model training, dynamic hybrid retrieval, fine rearrangement and reverse rearrangement, rule constraint generation and output, and interactive attention mechanism. The method has the advantages of improving the accuracy and relevance of medical insurance data retrieval, being able to effectively parse and understand the user's natural language query, ensuring that the results comply with medical semantics and policy specifications, and improving the real-time performance of data processing and retrieval.

[0008] According to an embodiment of the present invention, a data retrieval enhancement generation method in the field of medical insurance based on a large model includes the following steps: S1. Preprocess the medical insurance data, including data cleaning, data standardization and data labeling; S2. Train the big model by using the pre-processed medical insurance data to train the big model; S3, parse user queries, parse the natural language queries input by users through a large model, extract query intent and key information, and pre-process user queries; S4. Generate a corresponding question vector based on the user query; S5. Based on the generated question vector and question text, the vector similarity and the weight of BM25+ are dynamically adjusted through a dynamic hybrid retrieval method to obtain a candidate text sequence that is highly relevant to the question from the knowledge base; S6, rearrange the candidate text sequence in reverse order; S7, sending the rearranged candidate text sequence into the large model, and applying rule constraints to obtain the final output; S8, based on The interactive attention mechanism establishes associations between user queries and knowledge base texts, and combines it with an adaptive weight allocation method to dynamically adjust the weights of different attention heads to optimize the accuracy and relevance of retrieval results.

[0009] Optionally, the S3 specifically includes: S31. Receive natural language query input by user ; S32. Using large models for natural language queries Perform word segmentation and grammatical analysis to generate word segmentation sequences and dependency trees; S33. Apply knowledge graph-enhanced multimodal embedding technology to extract medical-related entities from word sequences , represents the total number of entities, where Indicates the extracted Medical-related entities, combining structured medical data and insurance claims data: ; in, represents the optimal subset of medical-related entities, represents the weight, Represents the entity matching degree in the knowledge graph, Represents the matching degree of multimodal data, Representing Entities With structured medical data and insurance claims data The correlation degree, Indicates that obtaining Get the maximum value of the medical-related entity set; S34. Using deep intent classifiers for natural language queries Perform intent classification and generate intent labels ; S35. Extract the optimal subset of medical-related entities and intent tags Encode, where Represents the optimal number of entities after screening, satisfying , generating a query vector representation : ; in, and Represent the weights of medical-related entities and multimodal data, respectively. represents the embedding representation of the entity, represents the embedding representation of multimodal data, The embedding representation of the intent label, is the weight of the intent label; S36, the query vector is represented Perform standardization to obtain a standardized query vector .

[0010] Optionally, the S4 specifically includes: S41, preprocessing the standardized user query vector Input to Embedding is performed in the embedding model to generate an initial vector representation ; S42, apply multi-layer bidirectional encoder to represent the initial vector Processing to generate an intermediate vector representation : ; in, represents a bidirectional long short-term memory network, represents a bidirectional gated recurrent unit network, represents the interaction of the attention mechanism, Represents a parallel combination of vectors.

[0011] S43, the intermediate vector represents Input into the multi-head self-attention mechanism, calculate the attention weight of each vector element, and generate a multi-head weighted vector representation : ; in, Indicates The weight of the attention head, Indicates The weight matrix of the attention head, Indicates The bias term of the attention head, is the number of attention heads.

[0012] S44, multi-head weighted vector representation Perform fusion and dimensionality reduction to generate the final problem vector representation : ; in, represents the activation function, represents the fusion weight.

[0013] S45. Represent the generated final question vector Perform standardization to obtain the standardized problem vector .

[0014] Optionally, the S5 specifically includes: S51. Split the files in the knowledge base, vectorize the segmented text blocks, and use the pre-trained embedding model Convert the text block into a vector and store the converted knowledge base vector into the vector library In order to perform vector retrieval; S52. Apply the BM25 algorithm to the segmented text blocks, calculate the matching score between the user query and each text block, obtain the original BM25 score, and normalize the original BM25 score so that the original BM25 score can be weighted in the same range as the vector similarity: ; in, Be original Score, and All The maximum and minimum values ​​in the score; S53, receiving standardized question vector , and calculate the cosine similarity between it and the vector in the knowledge base: ; in, and denote the standardized question vector and the vector in the knowledge base respectively, represents the dot product of vectors, and Represents the Euclidean norm of a vector; S54. Based on the feature analysis of the problem, the dynamic hybrid retrieval method is used to dynamically adjust the vector similarity and Weight: ; in, and Represent vector similarity and The weight of the search, their sum is ; S55. The weight adjustment function adjusts the vector similarity and The weights of , balance the importance of each feature in the question text: ; ; in, , and are adjustment coefficients, and the sum of all adjustment coefficients is equal to , Indicates the query word length, is the preset maximum question word length, It is the ratio of the frequency of keywords appearing in the question to the total number of words. is a binary indicator, equal to if the query contains professional terms , otherwise ; S56. Based on the comprehensive correlation score For candidate text Sort and output the sorted candidate text sequence.

[0015] Optionally, the S6 specifically includes: S61, sorting the candidate text sequences according to the comprehensive relevance scores and the normalized problem vector Input to Perform fine rearrangement in the model and generate fine rearrangement scores : in, The model uses a normalized question vector and candidate text Score the content; In the present invention, the The model is used to finely rearrange the candidate texts initially screened by the vector retrieval and BM25 hybrid strategy. Bidirectional Generative Embedding Ranker is a re-ranking model that combines a large language model structure. It can accurately score the matching degree between the query and the candidate paragraphs based on semantic integrity, contextual consistency and knowledge structure. Its main function is to make up for the shortcomings of traditional vector similarity at the semantic granularity level and improve the matching relevance and accuracy between the final output text and the user query.

[0016] S62, score based on detailed rearrangement For candidate text Reorder and generate new candidate text sequences ; S63. New candidate text sequence Perform reverse processing to generate a text sequence after reverse order ; S64. Calculate the comprehensive correlation score after reverse order rearrangement : ; in, and represents the weight coefficient of dynamic allocation, Represents the normalized problem vector The text vector after reverse order The cosine similarity between ; S65, comprehensive relevance score after rearrangement in reverse order Rearrange the text sequence in reverse order Verify and output the final candidate text sequence; S66, the final candidate text sequence As input, combined with the normalized question vector Tips for generating results for large models : ; in, represents the weight coefficient, Represents the function used to generate prompts; S67. Input the generated prompts into the large model to generate the final output results.

[0017] Optionally, the S7 specifically includes: S71. Obtaining a candidate text sequence after reverse order rearrangement and the normalized problem vector , Generate rule constraints, define rule constraints for text generation , including grammatical rules, medical terminology constraints, and medical insurance policy rules: ; in, Indicates Rule constraints; S72, generated by adjustment Constraints, in Directly add rule prompts to generate constraint prompts : ; in, Represents the function that generates prompts; S73. Standardize the problem vector and the candidate text vector after reverse order Combined with generation rule constraints Input into the large model to generate intermediate results : ; in, represents the rule constraint weight, represents the generating function combining rule constraints, is the number of rules used in the generation process; S74. Define output rule constraints , including output format, information integrity and logical consistency, through traditional filtering algorithm constraints, using keyword filtering, establishing a keyword blacklist, and automatically detecting whether the generated text contains words that should not appear; S75, the intermediate results Perform output rule constraint processing to generate the final output result : ; in, represents the output rule constraint weight, represents the processing function combined with the output rule constraints, is the number of rules used in the output process, Indicates output rule constraints The rules.

[0018] Optionally, the S8 specifically includes: S81. Define the rule constraints for the interactive attention mechanism , including semantic relevance, contextual consistency, and information importance; S82. Vectorize the problem and candidate text vector Input to In the model, the attention score is calculated : ; in, Indicates The weight of the attention head, , and The weight matrices representing queries, keys, and values, respectively, Represents the dimension of the key; S83. Calculate the comprehensive attention score , combined with the rule constraints of the interactive attention mechanism ; S84, using adaptive weight allocation method to dynamically adjust weight coefficients based on user feedback data and , optimize the comprehensive attention score: ; ; in, Indicates candidate text Relevant feedback, Representing the rule constraints of the interactive attention mechanism The Rules; S85. Output the candidate text sequence sorted based on the optimized comprehensive attention score as the final retrieval result.

[0019] The beneficial effects of the present invention are: The present invention utilizes a dynamic hybrid retrieval method, combines the BM25+ algorithm and vector similarity calculation, and dynamically adjusts the weights of vector similarity and BM25+ to adapt to query requirements of different features. This method can not only capture the semantic similarity of complex queries, but also cover a wide range of keyword matching requirements, improve the depth and breadth of retrieval, and thus ensure that the retrieval results are more relevant and accurate. In the process of rearranging the candidate texts, the sorting of the candidate texts is further optimized by introducing the BGE-Ranker model for fine rearrangement and reverse order processing, ensuring that the most relevant texts can be displayed first. Combined with the generation rule constraints and output rule constraints, when generating and outputting prompts, the predefined rules are strictly followed, and the keyword filtering and blacklist detection mechanisms are used to ensure that the generated texts comply with medical semantics and policy specifications, avoid words that should not appear, and improve the security and accuracy of the generated results.

[0020] The present invention is based on the interactive attention mechanism of Transformer. By establishing a deep semantic association between user queries and knowledge base texts, combined with an adaptive weight allocation method, the weights of different features can be dynamically adjusted, thereby improving the semantic understanding of complex queries and making the retrieval results closer to the user's true intentions. In addition, by analyzing the user's historical queries and feedback, personalized information recommendations are achieved. Both doctors and patients can obtain customized retrieval results, which significantly improves user experience and satisfaction. In terms of processing large-scale medical data, the present invention realizes fast vector retrieval by splitting and vectorizing the knowledge base files and storing them in the vector library Qdrant, significantly improving the real-time response capability. Combined with the fine rearrangement and reverse rearrangement mechanisms, when processing millions or even hundreds of millions of data, it can still quickly respond to user query needs, meeting the high requirements of clinical decision-making and real-time consultation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0022] Figure 1 A flowchart of a data retrieval enhancement generation method in the medical insurance field based on a large model proposed by the present invention; Figure 2A detailed flow chart of the dynamic hybrid retrieval steps; Figure 3 Detailed flowchart for fine rearrangement and reverse order rearrangement of candidate texts. DETAILED DESCRIPTION

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0024] refer to Figure 1-3 , a data retrieval enhancement generation method in the medical insurance field based on a large model, comprising the following steps: S1. Preprocess the medical insurance data, including data cleaning, data standardization and data labeling; S2. Train the big model by using the pre-processed medical insurance data to train the big model; S3, parse user queries, parse the natural language queries input by users through a large model, extract query intent and key information, and pre-process user queries; S4. Generate a corresponding question vector based on the user query; S5. Based on the generated question vector and question text, the vector similarity and the weight of BM25+ are dynamically adjusted through a dynamic hybrid retrieval method to obtain a candidate text sequence that is highly relevant to the question from the knowledge base; S6, rearrange the candidate text sequence in reverse order; S7, sending the rearranged candidate text sequence into the large model, and applying rule constraints to obtain the final output; S8, based on The interactive attention mechanism establishes associations between user queries and knowledge base texts, and combines it with an adaptive weight allocation method to dynamically adjust the weights of different attention heads to optimize the accuracy and relevance of retrieval results.

[0025] In this implementation, S3 specifically includes: S31. Receive natural language query input by user ; S32. Using large models for natural language queries Perform word segmentation and grammatical analysis to generate word segmentation sequences and dependency trees; S33. Apply knowledge graph-enhanced multimodal embedding technology to extract medical-related entities from word sequences , represents the total number of entities, where Indicates the extracted Medical-related entities, combining structured medical data and insurance claims data: ; in, represents the optimal subset of medical-related entities, represents the weight, Represents the entity matching degree in the knowledge graph, Represents the matching degree of multimodal data, Representing Entities With structured medical data and insurance claims data The correlation degree, Indicates that obtaining Get the maximum value of the medical-related entity set; S34. Using deep intent classifiers for natural language queries Perform intent classification and generate intent labels ; S35. Extract the optimal subset of medical-related entities and intent tags Encode, where Represents the optimal number of entities after screening, satisfying , generating a query vector representation : ; in, and Represent the weights of medical-related entities and multimodal data, respectively. represents the embedding representation of the entity, represents the embedding representation of multimodal data, The embedding representation of the intent label, is the weight of the intent label; S36, the query vector is represented Perform standardization to obtain a standardized query vector .

[0026] In this implementation manner, the S4 specifically includes: S41, preprocessing the standardized user query vector Input to Embedding is performed in the embedding model to generate an initial vector representation ; S42, apply multi-layer bidirectional encoder to represent the initial vector Processing to generate an intermediate vector representation : ; in, represents a bidirectional long short-term memory network, represents a bidirectional gated recurrent unit network, represents the interaction of the attention mechanism, Represents a parallel combination of vectors.

[0027] S43, the intermediate vector represents Input into the multi-head self-attention mechanism, calculate the attention weight of each vector element, and generate a multi-head weighted vector representation : ; in, Indicates The weight of the attention head, Indicates The weight matrix of the attention head, Indicates The bias term of the attention head, is the number of attention heads.

[0028] S44, multi-head weighted vector representation Perform fusion and dimensionality reduction to generate the final problem vector representation : ; in, represents the activation function, represents the fusion weight.

[0029] S45. Represent the generated final question vector Perform standardization to obtain the standardized problem vector .

[0030] In this implementation manner, S5 specifically includes: S51. Split the files in the knowledge base, vectorize the segmented text blocks, and use the pre-trained embedding model Convert the text block into a vector and store the converted knowledge base vector into the vector library In order to perform vector retrieval; S52. Apply the BM25 algorithm to the segmented text blocks, calculate the matching score between the user query and each text block, obtain the original BM25 score, and normalize the original BM25 score so that the original BM25 score can be weighted in the same range as the vector similarity: ; in, Be original Score, and All The maximum and minimum values ​​in the score; S53, receiving standardized question vector , and calculate the cosine similarity between it and the vector in the knowledge base: ; in, and denote the standardized question vector and the vector in the knowledge base respectively, represents the dot product of vectors, and Represents the Euclidean norm of a vector; S54. Based on the feature analysis of the problem, the dynamic hybrid retrieval method is used to dynamically adjust the vector similarity and Weight: ; in, and Represent vector similarity and The weight of the search, their sum is ; S55. The weight adjustment function adjusts the vector similarity and The weights of , balance the importance of each feature in the question text: ; ; in, , and are adjustment coefficients, and the sum of all adjustment coefficients is equal to , Indicates the query word length, is the preset maximum question word length, It is the ratio of the frequency of keywords appearing in the question to the total number of words. is a binary indicator, equal to if the query contains professional terms , otherwise ; S56. Based on the comprehensive correlation score For candidate text Sort and output the sorted candidate text sequence.

[0031] In this implementation manner, S6 specifically includes: S61, sorting the candidate text sequences according to the comprehensive relevance scores and the normalized problem vector Input to Perform fine rearrangement in the model and generate fine rearrangement scores : in, The model uses a normalized question vector and candidate text Score the content; S62, score based on detailed rearrangement For candidate text Reorder and generate new candidate text sequences ; S63. New candidate text sequence Perform reverse processing to generate a text sequence after reverse order ; S64. Calculate the comprehensive correlation score after reverse order rearrangement : ; in, and represents the weight coefficient of dynamic allocation, Represents the normalized problem vector The text vector after reverse order The cosine similarity between ; S65, comprehensive relevance score after rearrangement in reverse order Rearrange the text sequence in reverse order Verify and output the final candidate text sequence; S66, the final candidate text sequence As input, combined with the normalized question vector Tips for generating results for large models : ; in, represents the weight coefficient, Represents the function used to generate prompts; S67. Input the generated prompts into the large model to generate the final output results.

[0032] In this implementation manner, the S7 specifically includes: S71. Obtaining a candidate text sequence after reverse order rearrangement and the normalized problem vector , Generate rule constraints, define rule constraints for text generation , including grammatical rules, medical terminology constraints, and medical insurance policy rules: ; in, Indicates Rule constraints; S72, generated by adjustment Constraints, in Directly add rule prompts to generate constraint prompts : ; in, Represents the function that generates prompts; S73. Standardize the problem vector and the candidate text vector after reverse order Combined with generation rule constraints Input into the large model to generate intermediate results : ; in, represents the rule constraint weight, represents the generating function combining rule constraints, is the number of rules used in the generation process; S74. Define output rule constraints , including output format, information integrity and logical consistency, through traditional filtering algorithm constraints, using keyword filtering, establishing a keyword blacklist, and automatically detecting whether the generated text contains words that should not appear; S75, the intermediate results Perform output rule constraint processing to generate the final output result : ; in, represents the output rule constraint weight, represents the processing function combined with the output rule constraints, is the number of rules used in the output process, Indicates output rule constraints The rules.

[0033] In this implementation manner, S8 specifically includes: S81. Define the rule constraints for the interactive attention mechanism , including semantic relevance, contextual consistency, and information importance; S82. Vectorize the problem and candidate text vector Input to In the model, the attention score is calculated : ; in, Indicates The weight of the attention head, , and The weight matrices representing queries, keys, and values, respectively, Represents the dimension of the key; S83. Calculate the comprehensive attention score , combined with the rule constraints of the interactive attention mechanism ; S84, using adaptive weight allocation method to dynamically adjust weight coefficients based on user feedback data and , optimize the comprehensive attention score: ; ; in, Indicates candidate text Relevant feedback, Representing the rule constraints of the interactive attention mechanism The Rules; S85. Output the candidate text sequence sorted based on the optimized comprehensive attention score as the final retrieval result.

[0034] Embodiment 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to a medical insurance management center. The user Mr. Zhang is an employed employee. He needs to understand the specific process and policy of hospitalization reimbursement for urban employee medical insurance. At present, Mr. Zhang raised the question through the intelligent consultation of the management center: "How to reimburse the hospitalization of employed urban employee medical insurance?" Based on the large model-based medical insurance field data retrieval enhancement generation method of the present invention, an efficient and accurate answer is given.

[0035] First, we preprocess and parse Mr. Zhang’s natural language query. Through preprocessing, we segment and standardize the question. Then, we use a large model (such as BERT or GPT-4) to parse Mr. Zhang’s query and generate a question vector. Then, we enter the dynamic hybrid retrieval stage, combining BM25+ and vector similarity calculation to retrieve candidate text sequences that are highly relevant to the question from the knowledge base.

[0036] The specific process is as follows: Mr. Zhang inputs the question: "How can the urban employee medical insurance reimburse hospitalization?" The BCE embedding model is used to vectorize the question and generate the question vector First, calculate the BM25+ score between the question text and the text in the knowledge base, and get the BM25+ score sequence Then calculate the cosine similarity score between the question text vector and the text vector in the knowledge base , and obtain a similarity score sequence.

[0037] According to the characteristics of the current question, the weights of BM25+ and vector similarity are dynamically assigned. For example, for the question "How to claim reimbursement for hospitalization of urban employees' medical insurance", the word sequence obtained after word segmentation is "employee", "urban", "employee medical insurance", "hospitalization", "reimbursement". Calculate the vector similarity weight, use the adjustment coefficient and the default weight ratio, and the word length of the question , the preset maximum question word length , the ratio of the frequency of keywords appearing in the question to the total number of words . Because the question contains technical terms, So the weight is calculated , The final weighted score of each knowledge base text is calculated according to the weighted weight, and the top N knowledge base texts with the highest weighted scores are selected as candidate text sequences.

[0038] Use the BGE-Ranker model to finely rearrange the candidate text sequence. Input the candidate text sequence into the model and calculate the rearrangement score , sort in descending order according to the reordering score, and select the top N knowledge base texts with the highest scores as the reordered text sequence. In order to improve the output accuracy, the position of the text with high reordering scores in the model is closer to the question part, that is, the text sequence after reordering is obtained by reverse reordering. The reordered text sequence and the standardized question vector are compared. Generate prompts and input them into the big model. Generate rule constraints, including adding rule prompts directly into the prompts, such as "Answer user questions based on the medical insurance regulations and policies in the knowledge base, and the answers must be based on the content of the knowledge base".

[0039] To ensure that the output content is compliant, a keyword blacklist is established using a traditional filtering algorithm to automatically detect whether the generated text contains words that should not appear. When a word in the blacklist is detected, the large model will regenerate the answer; otherwise, the answer text generated by the large model will be used as the final output.

[0040] In July 2023, the Medical Insurance Management Center introduced the method of the present invention and conducted actual testing and application. The following is the actual test data: In the test, a total of 1,000 user query cases were selected, of which 500 were simple queries (such as "What should I do if my medical insurance card is lost") and 500 were complex queries (such as "How to reimburse the hospitalization of employed urban employees' medical insurance"). Retrieve and generate answers for each query, and record the retrieval and generation time, accuracy, and user satisfaction.

[0041] In simple queries, the average retrieval and generation time was 1.2 seconds, the accuracy was 95%, and the user satisfaction was 93%. In complex queries, the average retrieval and generation time was 2.8 seconds, the accuracy was 90%, and the user satisfaction was 88%. Compared with before the introduction of the present invention, the accuracy of complex queries increased by 20%, the retrieval time decreased by 30%, and the user satisfaction increased by 25%.

[0042] Table 1 Test data Table 2 Detailed data of complex query Table 3 Overall performance data From the above table, we can see that in simple queries (such as "What should I do if I lose my medical insurance card", "How to change the medical insurance payment base", "How to bind family members' medical insurance"), the average retrieval and generation time is 1.2 seconds, the retrieval accuracy rate reaches 95%, and the user satisfaction rate is 93%. These results show that it performs well in processing simple queries and can quickly and accurately provide the information required by users. For complex queries (such as "How to reimburse the hospitalization of employed urban employees' medical insurance", "Reimbursement policy for medical treatment in other places", "How to apply for major disease medical insurance"), the average retrieval and generation time is 2.8 seconds, the retrieval accuracy rate reaches 90%, and the user satisfaction rate is 88%. Although it takes a little longer to process complex queries, it can still provide high-quality answers, and user satisfaction is significantly improved.

[0043] In the process of processing complex queries, the BM25+ score between the question text and the text in the knowledge base is first calculated, and then the cosine similarity score between the question text vector and the text vector in the knowledge base is calculated. By dynamically adjusting the weights of BM25+ and vector similarity, more accurate search results can be provided based on the characteristics of the question. For example, for the query "How to reimburse the hospitalization of employed urban workers for medical insurance", the word sequence after word segmentation is "employed workers", "urban", "employee medical insurance", "hospitalization", "reimbursement", and the final calculated weight is , , ensuring the accuracy of the query results. The candidate texts are finely rearranged using the BGE-Ranker model, and are processed in reverse order according to the rearrangement scores, thereby ensuring that the most relevant texts are displayed first. For example, the rearrangement score sequence is 0.90, 0.85, 0.87, 0.84, 0.83, and the reverse rearrangement results are 0.83, 0.84, 0.87, 0.85, 0.90, which further optimizes the sorting of the text. By generating rule constraints and output rule constraints, keyword filtering and blacklist detection mechanisms are combined when generating prompts and final outputs to ensure that the generated text complies with medical semantics and policy specifications. For example, for generated results containing sensitive words, answers are automatically regenerated to ensure that the output content is healthy and compliant. Test data show that after introducing the method of the present invention, the accuracy of processing complex queries is increased from 70% to 90%, the average retrieval time is reduced from 4 seconds to 2.8 seconds, and user satisfaction is increased from 63% to 88%. These significant improvements prove the effectiveness and superiority of the present invention in practical applications.

[0044] The method of the present invention significantly improves the accuracy, relevance and real-time performance of medical insurance data retrieval and generation in practical applications. Whether it is a simple query or a complex query, it can efficiently solve the medical insurance problem of users, and provide strong technical support for medical consultation, clinical decision support, medical insurance claims, etc. The present invention not only overcomes the defects of the prior art such as low retrieval accuracy, weak semantic understanding ability, inability to process large-scale data, lack of personalized recommendations, information islands and data standardization problems.

[0045] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A data retrieval enhancement generation method in the field of medical insurance based on a large model, characterized in that: The steps include: S1. Preprocess the medical insurance data, including data cleaning, data standardization and data labeling; S2. Train the big model by using the pre-processed medical insurance data to train the big model; S3, parse user queries, parse the natural language queries input by users through the big model, extract the query intent and key information, and pre-process the user queries; S4. Generate a corresponding question vector based on the user query; S5. Based on the generated question vector and question text, the vector similarity and the weight of BM25+ are dynamically adjusted through a dynamic hybrid retrieval method to obtain a candidate text sequence that is highly relevant to the question from the knowledge base; S6, rearrange the candidate text sequence in reverse order; S7, sending the rearranged candidate text sequence into the large model, and applying rule constraints to obtain the final output; S8, based on The interactive attention mechanism establishes associations between user queries and knowledge base texts, and combines it with an adaptive weight allocation method to dynamically adjust the weights of different attention heads to optimize the accuracy and relevance of retrieval results.

2. According to claim 1, a data retrieval enhancement generation method in the field of medical insurance based on a large model is characterized in that: The S3 specifically includes: S31. Receive natural language query input by user ; S32. Using large models for natural language queries Perform word segmentation and grammatical analysis to generate word segmentation sequences and dependency trees; S33. Apply knowledge graph-enhanced multimodal embedding technology to extract medical-related entities from word sequences , represents the total number of entities, where Indicates the extracted Medical-related entities, combining structured medical data and insurance claims data: ； in, represents the optimal subset of medical-related entities, represents the weight, Represents the entity matching degree in the knowledge graph, Represents the matching degree of multimodal data, Representing Entities With structured medical data and insurance claims data The correlation degree, Indicates that obtaining Get the maximum value of the medical-related entity set; S34. Using deep intent classifiers for natural language queries Classify intent and generate intent labels ; S35. Extract the optimal subset of medical-related entities and intent tags Encode, where Represents the optimal number of entities after screening, satisfying , generating a query vector representation : ； in, and Represent the weights of medical-related entities and multimodal data, respectively. represents the embedding representation of the entity, represents the embedding representation of multimodal data, The embedding representation of the intent label, is the weight of the intent label; S36, the query vector is represented Perform standardization to obtain a standardized query vector .

3. According to the method of claim 1, the data retrieval enhancement generation method in the field of medical insurance based on a large model is characterized in that: The S4 specifically includes: S41, preprocessing the standardized user query vector Input to Embedding is performed in the embedding model to generate an initial vector representation ; S42, apply multi-layer bidirectional encoder to represent the initial vector Processing to generate an intermediate vector representation : ； in, represents a bidirectional long short-term memory network, represents a bidirectional gated recurrent unit network, represents the interaction of the attention mechanism, Represents parallel combinations of vectors; S43, the intermediate vector represents Input into the multi-head self-attention mechanism, calculate the attention weight of each vector element, and generate a multi-head weighted vector representation : ； in, Indicates The weight of the attention head, Indicates The weight matrix of the attention head, Indicates The bias term of the attention head, is the number of attention heads; S44, multi-head weighted vector representation Perform fusion and dimensionality reduction to generate the final problem vector representation : ； in, represents the activation function, represents the fusion weight; S45. Represent the generated final question vector Perform standardization to obtain the standardized problem vector .

4. According to the method of claim 1, the data retrieval enhancement generation method in the field of medical insurance based on a large model is characterized in that: The S5 specifically includes: S51. Split the files in the knowledge base, vectorize the segmented text blocks, and use the pre-trained embedding model Convert the text block into a vector and store the converted knowledge base vector into the vector library In order to perform vector retrieval; S52. Apply the BM25 algorithm to the segmented text blocks, calculate the matching score between the user query and each text block, obtain the original BM25 score, and normalize the original BM25 score so that the original BM25 score can be weighted in the same range as the vector similarity: ； in, Be original Score, and All The maximum and minimum values ​​in the score; S53, receiving standardized question vector , and calculate the cosine similarity between it and the vector in the knowledge base: ； in, and denote the standardized question vector and the vector in the knowledge base respectively, represents the dot product of vectors, and Represents the Euclidean norm of a vector; S54. Based on the feature analysis of the problem, the dynamic hybrid retrieval method is used to dynamically adjust the vector similarity and Weight: ； in, and Represent vector similarity and The weight of the search, their sum is ; S55. The weight adjustment function adjusts the vector similarity and The weights of , balance the importance of each feature in the question text: ； ； in, , and are adjustment coefficients, and the sum of all adjustment coefficients is equal to , Indicates the query word length, is the preset maximum question word length, It is the ratio of the frequency of keywords appearing in the question to the total number of words. is a binary indicator, equal to if the query contains professional terms , otherwise ; S56. Based on the comprehensive correlation score For candidate text Sort and output the sorted candidate text sequence.

5. According to the method of claim 1, the data retrieval enhancement generation method in the medical insurance field based on a large model is characterized in that: The S6 specifically includes: S61, sorting the candidate text sequences according to the comprehensive relevance scores and the normalized problem vector Input to Perform fine rearrangement in the model and generate fine rearrangement scores : in, The model uses a normalized question vector and candidate text Score the content; S62, score based on detailed rearrangement For candidate text Reorder and generate new candidate text sequences ; S63. New candidate text sequence Perform reverse processing to generate a text sequence after reverse order ; S64. Calculate the comprehensive relevance score after reverse order rearrangement : ； in, and represents the weight coefficient of dynamic allocation, Represents the normalized problem vector The text vector after reverse order The cosine similarity between ; S65, comprehensive relevance score after rearrangement in reverse order Rearrange the text sequence in reverse order Verify and output the final candidate text sequence; S66, the final candidate text sequence As input, combined with the normalized question vector Tips for generating results for large models : ； in, represents the weight coefficient, represents the function used to generate prompts; S67. Input the generated prompts into the large model to generate the final output results.

6. According to the method of claim 1, the data retrieval enhancement generation method in the field of medical insurance based on a large model is characterized in that: The S7 specifically includes: S71. Obtaining a candidate text sequence after reverse order rearrangement and the normalized problem vector , Generate rule constraints, define rule constraints for text generation , including grammatical rules, medical terminology constraints, and medical insurance policy rules: ； in, Indicates Rule constraints; S72, generated by adjustment Constraints, in Add rule prompts directly to generate constraint prompts : ； in, Represents the function that generates prompts; S73. Standardize the problem vector and the candidate text vector after reverse order Combined with generation rule constraints Input into the large model to generate intermediate results : ； in, represents the rule constraint weight, represents the generating function combining rule constraints, is the number of rules used in the generation process; S74. Define output rule constraints , including output format, information integrity and logical consistency, through traditional filtering algorithm constraints, using keyword filtering, establishing a keyword blacklist, and automatically detecting whether the generated text contains words that should not appear; S75, the intermediate results Perform output rule constraint processing to generate the final output result : ； in, represents the output rule constraint weight, represents the processing function combined with the output rule constraints, is the number of rules used in the output process, Indicates output rule constraints The rules.

7. According to claim 1, a data retrieval enhancement generation method in the field of medical insurance based on a large model is characterized in that: The S8 specifically includes: S81. Define the rules and constraints for the interactive attention mechanism , including semantic relevance, contextual consistency, and information importance; S82. Vectorize the problem and candidate text vector Input to In the model, the attention score is calculated : ； in, Indicates The weight of the attention head, , and The weight matrices representing queries, keys, and values, respectively, Represents the dimension of the key; S83. Calculate the comprehensive attention score , combined with the rule constraints of the interactive attention mechanism ; S84, using adaptive weight allocation method to dynamically adjust weight coefficients based on user feedback data and , optimize the comprehensive attention score: ； ； in, Indicates candidate text Relevant feedback, Representing the rule constraints of the interactive attention mechanism The Rules; S85. Output the candidate text sequence sorted based on the optimized comprehensive attention score as the final retrieval result.

Images