Medical document retrieval enhancement generation method and system based on parent-child node structure

Abstract

The application provides a medical document retrieval enhancement generation method and system based on a parent-child node structure, relates to the technical field of medical document retrieval, and comprises the following steps: analyzing an input medical document to obtain a medical text unit sequence; constructing a three-layer medical semantic tree node; performing quality inspection on each paragraph node to obtain a retrieval abstract; storing the retrieval abstract into a vector database and storing atomic proposition quadruple content and paragraph keywords into a full-text retrieval library; routing a user query to the vector database to perform semantic retrieval, and simultaneously performing keyword retrieval in the full-text retrieval library in parallel to obtain a primary retrieval result set; and calculating the relevance score of the primary retrieval result set and the user query by using a reordering model, arranging the primary retrieval result set in descending order according to the score, and obtaining a final retrieval result list. The application can solve the problems of the existing RAG system in the medical document processing scene, such as insufficient document structure perception, lack of reliable guarantee for medical fact integrity, and lack of multi-granularity retrieval collaboration ability.

Description

Technical Field

[0001] This invention relates to the field of medical document retrieval technology, and in particular to a method and system for enhancing medical document retrieval based on a parent-child node structure. Background Technology

[0002] Retrieval Augmentation (RAG) technology provides an effective framework for knowledge-intensive natural language processing tasks by combining large language models with external knowledge base retrieval. In the medical field, documents such as clinical guidelines, research literature, and treatment reports are characterized by high density of technical terms, rigorous logical structures, and strict requirements for factual accuracy, making them important application scenarios for RAG technology. However, existing RAG systems generally employ a fixed character window-based segmentation strategy when processing medical documents. This strategy mechanically cuts documents according to a preset length threshold, without considering the semantic boundaries and chapter structure of the documents. This results in the same medical fact (such as the applicable conditions and conclusions of drug contraindications) being segmented into different text blocks, leading to semantic fragmentation. Simultaneously, the single-level parent-child node pair index structure lacks awareness of document chapter hierarchy, making it difficult for the retrieval system to flexibly switch between precise matching of atomic facts and comprehensive topic retrieval when faced with different granularity of query requirements, thus affecting the relevance and contextual completeness of the search results.

[0003] Chinese patent application CN118861088A discloses a short text query enhancement retrieval method based on a hierarchical knowledge base tree structure. This method first uses document structured information to segment documents into semantically independent text blocks, and then uses UMAP dimensionality reduction and Gaussian mixture model soft clustering to classify the text blocks into multiple topics, recursively constructing a hierarchical knowledge base tree from bottom to top, with parent nodes storing aggregated summaries of child nodes. In the retrieval stage, a two-stage query expansion strategy based on a large language model is adopted, combining a hybrid retrieval mechanism of BM25 keyword retrieval and vector retrieval, performing retrieval enhancement layer by layer from the coarse-grained summary layer to the fine-grained leaf nodes, ultimately generating the query answer. This scheme improves the granular coverage capability of multi-level information retrieval to some extent, but it is designed for a general knowledge base, document segmentation relies on unsupervised clustering, the aggregation boundaries of the hierarchical tree lack semantic constraints, and it does not yet address the structured extraction and targeted quality assurance mechanisms for key clinical facts (such as contraindications and dosage constraints) in medical documents. Summary of the Invention

[0004] In view of this, the present invention provides a medical document retrieval enhancement generation method and system based on a parent-child node structure. By constructing a three-layer medical semantic tree node structure with clear clinical semantic boundaries, and combining a proposition coverage quality inspection mechanism based on clinical priority weighting and an intent-aware cross-granularity fusion retrieval architecture, the present invention solves the problems of insufficient document structure awareness, lack of reliable guarantee of medical fact integrity, and lack of multi-granularity retrieval collaboration capability in existing RAG systems in medical document processing scenarios.

[0005] The technical solution of this invention is implemented as follows: On the one hand, this invention provides a medical document retrieval enhancement generation method based on a parent-child node structure, including: S1. Parse the input medical document, identify chapter boundary information, and segment the document text into a sequence of medical text units; S2. Using a large language model, each medical text unit is decomposed into a set of medical atomic propositions. Each atomic proposition satisfies a four-tuple structure including the applicable subject, intervention measures, conditional constraints, and conclusion and level of evidence, and is labeled with the proposition type. Using the original text of the medical text unit as paragraph nodes, each paragraph node is aggregated according to the chapter boundary information to generate chapter aggregation nodes. The atomic proposition set, paragraph nodes, and chapter aggregation nodes are respectively used as the first layer, second layer, and third layer, and interconnected through bidirectional pointers between layers to construct a three-layer medical semantic tree node. S3. The semantic summary of each paragraph node is checked by a weighted coverage quality check mechanism based on the atomic proposition set. The semantic summary that fails the quality check is triggered to be regenerated until the quality check is passed, and the retrieval summary of each paragraph node is obtained. S4. Encode the retrieval summary of the three-layer medical semantic tree node into a semantic vector and store it in the vector database according to the level. Store the atomic proposition quadruple content and paragraph keywords in the full-text retrieval database. Store the complete content of each layer node and the level attribution mapping in the content storage database. Among them, the vector database is divided into three-granularity hierarchical index partitions according to proposition level, paragraph level and chapter level. S5. Receive user queries, use a hybrid intent recognition mechanism to classify the query at a granular level, route the user query to the corresponding partition of the vector database to perform semantic retrieval based on the intent classification results, and simultaneously perform keyword retrieval in parallel in the full-text search library. Combine the two retrieval results to obtain a preliminary retrieval result set. S6. For each candidate paragraph node in the initial search result set, use the reordering model to calculate its relevance score to the user query, sort them in descending order of score, and obtain the final search result list.

[0006] Preferably, step S1 includes: Medical documents are converted into text using optical character recognition (OCR) methods. Regular expressions and rule engines are then used to identify chapter titles, paragraph separators, and metadata information, resulting in chapter boundary information represented by chapter numbers, title levels, and paragraph affiliation tags. The extracted continuous text is segmented into medical text unit sequences. ,in Let be the i-th medical text unit, and n be the total number of medical text units; Each medical text unit satisfies the length integrity constraint. , Medical text unit The number of characters, To preset the minimum character length threshold, This is the preset maximum character length threshold.

[0007] Preferably, in step S2, for the i-th medical text unit The extracted j-th atomic proposition is denoted as Each atomic proposition must satisfy the quadruple structure. ,in For the applicable subject, These are intervention measures or test items that must be mapped to a medical terminology ontology. Standard entries in For conditions or quantitative constraints, The conclusion and its corresponding level of evidence label; proposition type label. Taken from a predefined set of proposition types These correspond to contraindication propositions, dosage constraint propositions, indication propositions, and background information propositions, respectively; text fragments that do not satisfy the complete four-tuple structure are not included in the atomic proposition set. .

[0008] Preferably, in step S2, the chapter aggregation node Chapter-level aggregate summary function Structure-aware aggregation of the set of paragraph nodes belonging to the s-th chapter yields: ; in This refers to the r-th paragraph node within chapter s. This is the set of paragraph nodes contained in the s-th chapter of the document, where s is the chapter number; the aggregation boundary is determined by the chapter title and chapter number in the chapter boundary information.

[0009] Preferably, the weighted coverage quality inspection mechanism in step S3 is executed as follows: For the i-th paragraph node Corresponding semantic summary The large language model is invoked to perform proposition decomposition, resulting in a set of summary propositions. ,in From The kth atomic proposition extracted from it. p represents the total number of abstract propositions; A set of atomic propositions As a verification benchmark, a proposition verification model is used to verify... Each atomic proposition Determine whether each one is... Coverage, Coverage Determination Function The value is 1 when the proposition verification model determines that the coverage is true, and 0 otherwise. Each atomic proposition The clinical importance weight is ,in Based on the proposition type label Assigned type precedence coefficient, Based on the propositional quadruple The included evidence level labels The assigned level of evidence coefficient; The clinical priority-weighted proposition coverage rate is: ; Where m is the set of atomic propositions The total number of valid atomic propositions; when Temporal Semantic Summary Through quality inspection, Set the preset coverage threshold; If the quality check fails, a regeneration will be triggered, exceeding the preset maximum number of retries. Then, using the original medical text unit Used as a substitute for the search summary.

[0010] Preferably, in step S4, the vector coding model used to encode the retrieval summary into a semantic vector is denoted as... , Based on the Sentence-BERT architecture, multi-task contrastive learning training using medical text is employed, with the training objective function being: ,in To contrast the learning loss, we use the medical terminology triplet loss. Preset hyperparameters to balance the weights of the two types of losses; The contrastive learning loss function is: ; Where N is the number of training sample pairs, and n is the training sample index. Let be the vector representation of the nth training sample. To and A semantically relevant positive sample vector, where q is the negative sample index. For the qth pair A semantically unrelated vector of negative samples, where M is the total number of negative samples. The cosine similarity function is used. A preset temperature hyperparameter for adjusting the smoothness of the similarity distribution.

[0011] Preferably, the hybrid intent recognition mechanism in step S5 is divided into two stages: The first stage is rule triggering. If a user's query Q contains explicit numerical thresholds, precise dosages, or absolute operational semantic terms, it is determined to be a precise fact query, and the user's query is routed to the propositional vector partition. If the query contains comparative terms, it is classified as a review query and the user query is routed to the chapter-level vector partition. ; The second stage is the vector classifier. When the rule trigger in the first stage does not produce a deterministic decision, the user query Q is input into a three-class intent classifier fine-tuned using a medical query labeled dataset. Output the intent confidence vector: ; in This is a granular hierarchical index; prop, para, and chapter correspond to proposition-level precise fact queries, paragraph-level context queries, and chapter-level summary queries, respectively. For users to query Q corresponding to the first The intent confidence at the layer granularity satisfies ; When the maximum confidence level Not lower than the preset reliability threshold When a user query is in the highest confidence level, the corresponding partition with the highest confidence level is routed; otherwise, the propositional vector partitioning is activated simultaneously. Paragraph-level vector partitioning Chapter-level vector partitioning Parallel retrieval.

[0012] Preferably, in step S5, the method for obtaining the primary search result set is as follows: If the first-stage rule trigger produces a deterministic judgment, or if the highest intent confidence output by the second-stage vector classifier is not lower than the preset confidence threshold, then the user query will be routed to the corresponding single partition to perform semantic retrieval. The semantic retrieval results of that partition will be combined with the keyword retrieval results of the full-text retrieval library to form the primary retrieval result set. If the first-stage rule trigger does not produce a deterministic judgment and the highest intent confidence output by the second-stage vector classifier is lower than the preset confidence threshold, then the three partitions of proposition level, paragraph level and chapter level are activated simultaneously to perform semantic retrieval in parallel. After normalizing the retrieval scores in the candidate set of each partition, cross-granularity weighted fusion sorting is performed with the intent confidence of each level as the weight. The union of the candidate nodes after fusion sorting and the keyword retrieval results of the full-text retrieval library is taken to form the primary retrieval result set. For each level node hit in the initial search result set, the corresponding paragraph node is uniformly reduced through bidirectional pointer tracing between levels to obtain a candidate set of paragraph nodes for reordering.

[0013] Preferably, in step S6, the reordering model Based on the LightGBM architecture, the model uses the vector semantic similarity between each candidate paragraph node in the primary search result set and the user query, the keyword coverage ratio of the query medical terms in the candidate nodes, the evidence level weight of the candidate nodes, and the timeliness decay weight of the candidate nodes as basic features. The model also introduces polynomial extensions of the above features to generate higher-order interaction features, which together constitute the input feature vector. The model output is mapped to the relevance probability of each candidate paragraph node through the sigmoid function, and the final search result list is obtained by sorting the relevance probabilities in descending order.

[0014] In addition, the present invention also provides a medical document retrieval enhancement generation system based on a parent-child node structure to implement the above-described method, the system comprising: The document parsing module is used to parse the input medical documents, identify chapter boundary information, and segment the document text into a sequence of medical text units; The three-layer semantic tree node construction module is used to decompose each medical text unit into a set of medical atomic propositions using a large language model, aggregate paragraph nodes based on chapter boundary information to generate chapter aggregation nodes, and use the set of atomic propositions, paragraph nodes, and chapter aggregation nodes as the first, second, and third layers respectively and connect them with bidirectional pointers between layers to construct a three-layer medical semantic tree node. The quality inspection module is used to inspect the semantic summary of paragraph nodes based on the set of atomic propositions corresponding to each paragraph node and adopt a weighted coverage quality inspection mechanism. Semantic summaries that fail the quality inspection are regenerated until the quality inspection is passed, and the search summary of each paragraph node is obtained. The indexing and database module is used to encode the retrieval summary of the three-layer medical semantic tree node into a semantic vector and then store it into the vector database according to the level. It stores the content of atomic proposition quadruples and paragraph keywords into the full-text retrieval database and stores the complete content of each layer node and its level mapping into the content storage database. Among them, the vector database establishes a three-granularity hierarchical index partition according to the proposition level, paragraph level and chapter level. The retrieval execution module receives user queries, uses a hybrid intent recognition mechanism to classify the queries at a granular level, routes the user query to the corresponding partition of the vector database to perform semantic retrieval based on the intent classification results, and simultaneously performs keyword retrieval in parallel in the full-text retrieval database. The results from the two retrieval paths are combined to obtain a preliminary retrieval result set. The reordering module is used to calculate the relevance score of each candidate paragraph node in the primary search result set to the user query using a reordering model, and sort them in descending order of score to obtain the final search result list.

[0015] The present invention has the following advantages over the prior art: (1) This invention constructs medical documents into three layers: a proposition node layer, a paragraph node layer, and a chapter aggregation node layer, which are interconnected by bidirectional pointers. The chapter aggregation boundary is explicitly determined by the chapter title and chapter number identified during the document parsing stage, rather than relying on unsupervised clustering. This ensures that the aggregation result strictly corresponds to the clinical semantic chapters of the medical document itself, avoiding the cross-semantic topic aliasing problem introduced by unsupervised clustering.

[0016] (2) This invention requires that each atomic proposition must satisfy a four-tuple structure containing the applicable subject, intervention measure, conditional constraints, and conclusion and level of evidence, and standardizes the intervention measure field through a medical terminology ontology. Text fragments that do not satisfy the complete four-tuple structure are not included in the proposition set. This constraint transforms unstructured natural language medical text into structured propositional units with explicit semantic fields, enabling the proposition set to support precise Boolean filtering retrieval.

[0017] (3) This invention designs a clinical importance weight based on the product of the proposition type priority coefficient and the evidence level coefficient, and uses this weight to calculate the coverage of the original text's proposition set. Compared with equal-weighted coverage, this quality control mechanism makes the omission or distortion of high-risk proposition types have a more significant impact on the quality control indicators, and can identify factual distortions that are similar in semantic style but have different clinical meanings. Such distortions are difficult to be effectively captured by the overall semantic similarity index.

[0018] (4) This invention employs a two-stage intent recognition mechanism that combines rule triggers and vector classifiers. The rule triggers make deterministic judgments based on explicit linguistic features such as numerical thresholds, precise dosages, or absolute operational semantic words in the query; the vector classifier outputs a three-granularity intent confidence vector when a rule is not matched. When the intent confidence is lower than a preset threshold, the system activates all three index partitions in parallel to perform the retrieval, avoiding recall losses caused by partition misselection in scenarios where the query intent is ambiguous due to hard routing strategies.

[0019] (5) After performing intra-partition normalization on the retrieval scores in each partition candidate set, the present invention performs weighted fusion with intent confidence as the weight, so that the fusion ranking reflects the semantic relevance of the query to each granular node and the probability of intent attribution. When the intent is clear, it is equivalent to precise single partition routing, and when the intent is ambiguous, it realizes multi-granular soft parallel retrieval. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the three-layer medical semantic tree node structure of the present invention; Figure 3 This is a schematic diagram of the parallel data entry path of the present invention; Figure 4 This is a flowchart of the hybrid intent recognition and cross-granularity retrieval process of the present invention; Figure 5 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this invention provides a medical document retrieval enhancement generation method based on a parent-child node structure, comprising: S1. Parse the input medical document, identify chapter boundary information, and segment the document text into a sequence of medical text units; S2. Using a large language model, each medical text unit is decomposed into a set of medical atomic propositions. Each atomic proposition satisfies a four-tuple structure including the applicable subject, intervention measures, conditional constraints, and conclusion and level of evidence, and is labeled with the proposition type. Using the original text of the medical text unit as paragraph nodes, each paragraph node is aggregated according to the chapter boundary information to generate chapter aggregation nodes. The atomic proposition set, paragraph nodes, and chapter aggregation nodes are respectively used as the first layer, second layer, and third layer, and interconnected through bidirectional pointers between layers to construct a three-layer medical semantic tree node. S3. The semantic summary of each paragraph node is checked by a weighted coverage quality check mechanism based on the atomic proposition set. The semantic summary that fails the quality check is triggered to be regenerated until the quality check is passed, and the retrieval summary of each paragraph node is obtained. S4. Encode the retrieval summary of the three-layer medical semantic tree node into a semantic vector and store it in the vector database according to the level. Store the atomic proposition quadruple content and paragraph keywords in the full-text retrieval database. Store the complete content of each layer node and the level attribution mapping in the content storage database. Among them, the vector database is divided into three-granularity hierarchical index partitions according to proposition level, paragraph level and chapter level. S5. Receive user queries, use a hybrid intent recognition mechanism to classify the query at a granular level, route the user query to the corresponding partition of the vector database to perform semantic retrieval based on the intent classification results, and simultaneously perform keyword retrieval in parallel in the full-text search library. Combine the two retrieval results to obtain a preliminary retrieval result set. S6. For each candidate paragraph node in the initial search result set, use the reordering model to calculate its relevance score to the user query, sort them in descending order of score, and obtain the final search result list.

[0024] In one embodiment of the present invention, step S1 includes: The system receives raw medical documents from the healthcare information system, with input sources including clinical guidelines, research literature, and treatment reports. For scanned or image-formatted medical documents, the document parser first converts the unstructured image information into an editable text stream using optical character recognition (OCR) and performs format standardization. Subsequently, the parser uses regular expressions and a rule engine to identify chapter titles, paragraph separators, and metadata information within the document, obtaining chapter boundary information represented by chapter numbers, heading levels, and paragraph affixation tags.

[0025] After completing document structure recognition, the system segments the extracted continuous text into medical text unit sequences. ,in ( Let be the i-th medical text unit, and n be the total number of medical text units. Each medical text unit must satisfy the length integrity constraint: ; in Medical text unit The number of characters, A minimum character length threshold is preset, set to 50 characters, to remove invalid and fragmented content. A maximum character length threshold is preset, set to 2000 characters, to prevent excessively long units from causing difficulties in semantic compression. This constraint ensures the semantic integrity of each text unit while maintaining a balance in the granularity of segmentation.

[0026] like Figure 2 As shown, in one embodiment of the present invention, step S2 includes: Step S2 is divided into four parts: medical atomic proposition node construction, semantic summary generation, paragraph node definition, and chapter aggregation node construction. Finally, the three-layer nodes are integrated into a unified tree structure through bidirectional pointers between layers.

[0027] For the i-th medical text unit The system calls a large language model that has been fine-tuned from medical literature corpus. Decompose it into a set of medical atomic propositions : ; in Let m be the j-th atomic proposition extracted from the i-th text unit, and m be the total number of atomic propositions contained in that text unit. Model A large language model fine-tuned by instructions for medical proposition extraction can be used, such as a model based on LLaMA-3 or Qwen series models, to supervise and fine-tune the medical literature corpus so that it can identify independent medical facts in the text and output them in a four-tuple format.

[0028] For the i-th medical text unit The j-th extracted atomic proposition Each atomic proposition must satisfy the quadruple structure. ,in For the applicable subject, describe the characteristics of the patient population to which this proposition applies (e.g., "adult patients with eGFR < 30 mL / min"). For intervention measures or test items, they must be mapped to the medical terminology ontology. Standard entries in Covering a set of disease names based on ICD-10 A set of drug names based on ATC classification HGNC-based biomarker set Evidence level set based on Oxford evidence classification and clinical recommendation level set ; For conditions or quantitative constraints, containing explicit operational semantics (such as "prohibit use" or "half the dose"); This refers to the conclusion and its corresponding level of evidence label. Text fragments that do not satisfy the complete four-tuple structure are not included in the set of atomic propositions. This is to ensure that the granularity of the propositions is consistent with the completeness of the information required for clinical decision-making.

[0029] Each proposition At the same time Label the proposition type The type label is taken from a predefined set of proposition types. These correspond to the contraindication proposition, the dose constraint proposition, the indication proposition, and the background information proposition, respectively.

[0030] The second-level paragraph nodes directly use the original complete medical text units as storage content, that is... Paragraph node In the three-layer tree structure, it plays a bridging role: providing the proposition layer with a complete context for fact-checking and retrieval, providing the chapter layer with raw materials for topic aggregation, and serving as the context carrier for input to the large language model during the answer generation stage. Child nodes associated with paragraph nodes. For triples ,in For semantic summaries that have passed the quality check in step S3, vector-oriented semantic retrieval is performed. Based on medical terminology ontology A standardized set of structured keywords for precise terminology Boolean retrieval; It is a collection of medical atomic propositions, and also serves as a benchmark for abstract quality verification.

[0031] Third-level chapter aggregation node Chapter-level aggregate summary function Structure-aware aggregation of the set of paragraph nodes belonging to the s-th chapter yields: ; in This refers to the r-th paragraph node within chapter s. This is the set of paragraph nodes contained in the s-th chapter of the document, where s is the chapter number. Can be adopted with Large language models of the same series perform chapter-level summary generation, with the input being... The ordered concatenation of text from each paragraph node results in a chapter-specific thematic summary covering the core arguments, recommendation level distribution, and applicable conditions of the chapter, rather than a simple concatenation of paragraph summaries. The aggregation boundary is explicitly determined by the chapter titles and chapter numbers identified in step S1. The chapter structures of clinical guidelines and medical literature (such as "indications," "contraindications," and "medication for special populations") have clear clinical semantic boundaries. Therefore, the aggregation boundary determined in this way does not rely on unsupervised clustering, will not cross semantically unrelated topics, and the aggregation results have inherent consistency in clinical semantics.

[0032] After completing the above three-layer construction, the nodes in the three layers are interconnected through a bidirectional pointer system: ; Each proposition node Carrying a pointer to the paragraph node Reverse pointer Each paragraph node Carrying a pointer to the chapter's aggregation node Parent pointer Once a search at any level is successful, the pointer chain can be traced back to the complete parent context, ultimately ending at the paragraph node. The complete original text is used as the context carrier for input into the large language model.

[0033] By upgrading the original single-layer parent-child node pair to a three-layer medical semantic tree node structure, and using the chapter boundaries explicitly marked in the document format as the aggregation boundary, the index nodes at all three granularities strictly correspond to the clinical semantic chapters of the medical document itself, thus improving the problem of missing context anchors caused by cross-semantic boundary cutting under the fixed window segmentation method.

[0034] In one embodiment of the present invention, step S3 includes: The system employs a weighted coverage quality control mechanism based on atomic proposition sets to perform quality checks on the semantic summaries of each paragraph node. Before performing the quality check, the system first invokes the semantic summarization generation model. For text units Perform a summary generation process to obtain the summary text. , Generative large language models fine-tuned for medical text summarization tasks can be used, such as medical summarization fine-tuning models based on the Qwen or GLM series.

[0035] For the i-th paragraph node Corresponding semantic summary System calls to the large language model Perform proposition decomposition to obtain a set of summary propositions. ,in To extract from the abstract The kth atomic proposition extracted from it. p represents the total number of abstract propositions.

[0036] A set of atomic propositions As a verification benchmark, a proposition verification model is used to verify... Each atomic proposition Determine whether each one is... Coverage. The proposition verification model employs a Natural Language Inference (NLI) model fine-tuned for the medical proposition implication task, such as a three-class classification model based on the DeBERTa-v3 architecture and fine-tuned on the MedNLI dataset and a medical proposition pair annotation dataset, outputting three classification results: "Supported / Not Supported / Not Addressed". Coverage judgment function. The value is 1 when the proposition verification model determines that the coverage is true, and 0 otherwise; that is, it is only true when the determination result is "Supported". This judgment compares each statement at the atomic proposition level, and can identify factual distortions that are semantically similar but have different clinical meanings, such as "contraindication" being rewritten as "use with caution" or specific dosage values ​​being blurred. However, the overall cosine similarity is not sensitive to such distortions.

[0037] Each atomic proposition The clinical importance weights are: ; in Based on the proposition type label Assigned type precedence coefficient, Based on the propositional quadruple The included evidence level labels The assigned level of evidence coefficient. The type priority coefficient. The following rules apply to assigning values: contraindication is assigned a value of 4, dosage constraint a value of 3, indication a value of 2, and background information a value of 1. Priority coefficients are designed based on the severity of the consequences of clinical decisions. (Evidence level coefficient) The values ​​are based on the Oxford Grading System: The value is 1.0. Take 0.8 at that time. Take 0.6 at the time. Take 0.4 at this time.

[0038] The clinical priority-weighted proposition coverage rate is: ; Where m is the set of atomic propositions The total number of valid atomic propositions. When At that time, semantic summarization Through quality inspection, The preset coverage threshold is used; if the quality check fails, regeneration is triggered, and the maximum number of retries is exceeded. ( (Then followed by the original medical text unit) The alternative is used as the search summary to ensure that the data entry process is not interrupted.

[0039] Compared to equal-weighted coverage, The core characteristic is that if the abstract fully covers all background information propositions but omits high-level evidence contraindications, its weighted coverage may still be relatively high. Because contraindication propositions have the highest weight, this omission will significantly reduce the weighted coverage, thereby effectively intercepting omissions or distortions of high-risk clinical propositions during the quality control stage.

[0040] like Figure 3 As shown, in one embodiment of the present invention, step S4 includes: Step S4 employs a three-database, dual-engine storage architecture for the vector database. Introducing a three-granularity hierarchical index partitioning for the full-text search library Introducing propositional quadruple index partitioning and expanding the content repository. To support unified management of three-layer nodes, five parallel data entry paths are formed.

[0041] Path 1 involves storing the paragraph semantic summary vectors in the database: This involves storing the qualified semantic summaries corresponding to the second-level paragraph nodes. Through vector coding model Convert to semantic vector Store in a vector database Paragraph semantic index partition .

[0042] Path 2 involves storing the chapter aggregation node vectors in the database: This includes the third-level chapter aggregation nodes. through Encoded as chapter-level vectors Store in the chapter semantic index partition .

[0043] Path three involves the full-text storage of atomic proposition vectors and quadruples: This involves the first-level atomic proposition set. Each proposition On the one hand, the proposition text is... Encoded as a 768-dimensional propositional vector Store in the proposition-level index partition On the other hand, the content of propositional quadruples will be structured. Along with the proposition type label and clinical importance weight Stored together in the full-text search database Proposition Index Partition It supports precise Boolean filtering queries based on fields such as applicable subject, intervention measures, conditional constraints, and evidence level.

[0044] Path four involves storing the full text of paragraph node keywords in the database: This involves storing the structured keyword set of the second-level paragraph nodes. The corresponding parent node pointer is associated with and stored in the full-text search database paragraph keyword index partition. Constructing a medical terminology ontology Inverted index of standardized terms.

[0045] Path five involves storing the complete content of the three-tiered nodes into a unified content repository. Extended storage chapter aggregation node The aggregated summary text is used to establish a hierarchical mapping from paragraph nodes to chapter nodes. The three-level content index can be formally represented as follows: ; in A unique pointer to the paragraph node. This is a unique pointer to the chapter aggregation node. Each proposition record carries a reverse pointer. , pointing to the paragraph node to which it belongs. This ensures the traceability of the context after a proposition-level retrieval is successful. and , All indexes are built using a hierarchical navigable small-world graph (HNSW) structure, supporting efficient near nearest neighbor (ANN) retrieval. The three partitions are distinguished by partition labels.

[0046] The above-mentioned vector coding model is used to encode the retrieval summary into a semantic vector. Based on the Sentence-BERT architecture, multi-task contrastive learning training of medical text is used, and the training objective function is: ; in Cosine similarity contrastive learning loss is used to minimize the angular differences between semantically related texts in the embedding space; The medical term triplet loss is used to constrain the discriminative power between positive and negative samples in the medical concept dimension. Preset hyperparameters to balance the weights of the two types of losses. The specific form of the contrastive learning loss function is: ; Where N is the number of training sample pairs, and n is the training sample index. Let be the vector representation of the nth training sample. To and A semantically relevant positive sample vector, where q is the negative sample index. For the qth pair A semantically unrelated vector of negative samples, where M is the total number of negative samples. The cosine similarity function is used. A preset temperature hyperparameter for adjusting the smoothness of the similarity distribution.

[0047] like Figure 4 As shown, in one embodiment of the present invention, step S5 includes: After receiving a user query Q, the system first performs granular intent classification on Q using a hybrid intent recognition mechanism, and outputs an intent confidence vector. ,in For users to query Q corresponding to the first Layer size ( The intent confidence score, where prop, para, and chapter correspond to proposition-level exact fact queries, paragraph-level context queries, and chapter-level review queries, respectively, satisfying the following: .

[0048] The hybrid intent recognition mechanism consists of two stages; The first stage is a rule trigger: it checks whether query Q satisfies the following rules. If it does, it directly assigns full confidence to the corresponding type (the confidence of this type in the confidence vector is 1.0, and the rest are 0): If Q contains a specific numerical threshold (such as "eGFR<30"), a precise dose (such as "500 mg / d"), or absolute operational semantic words (such as "contraindicated" or "prohibited"), it is determined to be a precise fact query, and the user query is routed to the propositional vector partition. If Q contains comparative terms (such as "comparison", "difference", "different guidelines", "review"), it is determined to be a review query, and the user query is routed to the chapter-level vector partition. .

[0049] The second stage is the vector classifier: when the rule trigger in the first stage does not produce a deterministic decision, the user query Q is input into a three-class intent classifier fine-tuned using a medical query labeled dataset. , This is a fine-tuned version of a lightweight BERT model (such as BERT-base or its pre-trained medical variants, such as PubMedBERT) on a medical query intent annotation dataset, outputting an intent confidence vector. ; When the maximum confidence level Not lower than the preset reliability threshold When a user query is executed, semantic retrieval is performed on the partition with the highest confidence level; otherwise, propositional vector partitioning is activated simultaneously. Paragraph-level vector partitioning Chapter-level vector partitioning Parallel retrieval.

[0050] After completing the intent classification, the system performs two-way retrieval in parallel: vector retrieval and keyword retrieval.

[0051] Vector retrieval path: utilizing vector encoding model Convert query Q into a semantic vector In the target partitions routed by the intent recognition module, an approximate nearest neighbor search is performed, and the top k candidate nodes of each partition are selected to obtain the original similarity score set for each partition. k=50 is the size of the primary search candidate set for each partition, where query vector With the The vector of the d-th candidate node in the partition The cosine similarity.

[0052] Keyword retrieval path: Extract precise medical terms from Q using a query parser and map them to a set of keywords. ,exist and Constructing Boolean query expressions ,in This represents an exact match operation on a single medical term k, which, when executed, yields a set of keyword matching results. .

[0053] If the first-stage rule trigger produces a deterministic decision, or if the highest intent confidence output by the second-stage vector classifier is not lower than the preset confidence threshold. If the user query is routed to the corresponding single partition for semantic retrieval, the union of the semantic retrieval results of that partition and the keyword retrieval results of the full-text search library is taken to form the initial retrieval result set; if the first-stage rule trigger does not produce a deterministic decision and the highest intent confidence output by the second-stage vector classifier is lower than Then, the three partitions are activated simultaneously to perform semantic retrieval in parallel. After normalizing the retrieval scores in the candidate sets of each partition, cross-granularity weighted fusion ranking is performed with the intent confidence of each level as the weight.

[0054] Intent-aware cross-granularity fusion score The calculation method is as follows: ; in For candidate node d in the th The normalized similarity score within each partition is calculated as follows: ; in For candidate node d in the th The original cosine similarity score of the partition; For the first Partition the set of all candidate nodes returned by this retrieval; and These are the minimum and maximum scores within the candidate set of this partition, respectively; To prevent numerical stability constants with denominators of zero, after in-partition min-max normalization, This eliminates the problem of incomparable scores across partitions caused by differences in the distribution of node vectors at different granularities.

[0055] The design ensures that: when the intent recognition is highly certain (a certain...) When the intent is ambiguous (three), the fusion score degenerates into the individual normalized score of that partition, equivalent to precise single-partition routing; when the intent is ambiguous (three), the fusion score degenerates into the individual normalized score of that partition, equivalent to precise single-partition routing. When the distribution is close to uniform, the fusion score combines the contributions of the three partitions to achieve soft, multi-granular parallel retrieval. This mechanism directly links the confidence output of query intent recognition with the retrieval ranking, avoiding recall losses caused by partition misselection in scenarios with ambiguous intent due to hard routing.

[0056] Vector retrieval candidate nodes After sorting, select the top 50 results and then compare them with the keyword search results set. Take the union of the results to form the initial search result set. : ; right All nodes hit in the hit list, whether they are proposition nodes or chapter aggregation nodes, are accessed via bidirectional pointers between levels. or The reverse tracing reduces the results to the corresponding paragraph nodes. This yields a candidate set of paragraph nodes with consistent granularity, which is then used for reordering in step S6.

[0057] By concatenating rule triggers with a lightweight vector classifier, normalizing them, and then weighting the cross-granularity retrieval scores with intent confidence, step S5 maintains the efficiency of precise single-partition routing when the intent is clear, and avoids the omission of relevant candidate nodes due to incorrect partition selection caused by hard routing when the intent is ambiguous. At the same time, it incorporates the problem of the distribution difference of node vectors of different granularities into the sorting design to resolve it, thereby improving the comparability and sorting consistency of cross-partition candidate sets under the multi-granularity indexing system.

[0058] In one embodiment of the present invention, step S6 includes: After receiving the candidate set of paragraph nodes reduced in step S5, the system processes each candidate paragraph node... By paragraph node pointer From content repository Retrieve the complete original text to ensure that subsequent relevance assessments are conducted within the full context, rather than relying solely on the search summary. .

[0059] The system provides each candidate paragraph node Calculate the feature vector of multiple correlations The four types of features are defined as follows.

[0060] Vector similarity score Defined as the cosine similarity between the query vector and the semantic summary vector of the candidate paragraph nodes: ; in For users to query Q-vector encoding model The resulting query vector, This is the semantic summary vector corresponding to the candidate paragraph node. Describes the L2 norm of a vector. Measure the semantic proximity of the query and candidate results in the embedding space.

[0061] Keyword matching degree Defined as the coverage ratio between the set of medical terms in the query and the set of keywords in the candidate results: ; in To extract medical terminology ontology from user query Q using a query parser A standardized set of medical terms for mapping. Child nodes of candidate paragraph nodes The structured keyword set stored in the middle, Measure the coverage of the query intent by candidate results at the level of medical terminology.

[0062] Weight of Evidence Level Based on the evidence level labels marked in the atomic propositions associated with the candidate paragraph nodes Assignment: ; ; in For text from parent node The evidence level label parsed from the function The pre-defined weight mapping rules are used to reflect the differences in clinical credibility among different levels of evidence, so that high-quality clinical evidence can obtain a relatively high ranking in the re-ranking.

[0063] Timeliness weight Exponential decay processing is applied to the publication time of candidate results: ; in The release time of the candidate results. The time is the current time, and the difference between the two is in years. This is a time decay coefficient used to appropriately reduce the weight of content published earlier, prioritizing the presentation of more timely medical knowledge at the ranking level.

[0064] The above four basic characteristics The system comprises four dimensions, which are then expanded using a second-order polynomial to generate higher-order interactive features. Specifically, the second-order expansion consists of two parts: one is the second-order self-multiplication term of each basic feature. The first dimension is 4, used to capture the nonlinear contribution of each feature itself; the second is the cross-product term between each pair of basic features. ,common Dimensions are used to explicitly model the joint effects between features, such as This also reflects the synergistic effect of semantic relevance and evidence quality on the ranking contribution. This also examines the combined effect of terminology coverage and timeliness. The 4-dimensional basic features, together with the aforementioned 4-dimensional self-multiplication terms and 6-dimensional cross-product terms, constitute a 12-dimensional input feature vector. : ; The first four dimensions are the original basic features, the middle four dimensions are second-order self-multiplication terms, and the last six dimensions are pairwise cross-product terms, totaling 14 dimensions. Considering that the model filters redundant terms with contributions below a threshold during the training phase based on feature importance, 12 effective feature inputs are ultimately retained. This polynomial extension method is a standard feature engineering technique that allows LightGBM to directly model nonlinear interactions using the product relationships between features during decision tree splitting without increasing model complexity, thus overcoming the limitations of tree models in perceiving explicit feature interactions.

[0065] Input: Reordering model based on LightGBM architecture The model output is mapped to the relevance probability of each candidate paragraph node via the sigmoid function: ; in The sigmoid function maps the model's original output to the interval (0, 1). The closer the value is to 1, the higher the clinical relevance of the candidate paragraph node to the user's query. All candidate results are based on... Sort the results in descending order and output the final list of search results. The complete parent node text for each candidate option is provided for use in subsequent answer generation steps.

[0066] By introducing multinomial extended features, the re-ranking model can explicitly express the joint contribution of multidimensional relevance factors through feature product terms without increasing network depth. This allows the ranking results to simultaneously take into account the synergistic effect between semantic matching degree, term coverage, evidence quality and timeliness. Compared with simply using 4-dimensional independent features, the modeling capability in the feature interaction dimension is improved.

[0067] After the reordering is completed in step S6, the system starts from... Extract the complete paragraph node text of the Top-K (K is 5) results. As evidence context, a prompt template was constructed and input into a large language model fine-tuned with medical dialogue data. Generate the final answer: ; The prompt template is constructed in the following format: "Based on the following clinical evidence context:" Please answer the inquiry as a clinical expert: Requirements: Strictly base answers on provided evidence, clearly indicate the level of evidence, and avoid inferences. Generative large language models fine-tuned from medical dialogue datasets (such as MedDialog, CMedQA, etc.) and clinical guideline question-and-answer data can be used, for example, fine-tuned versions of medical dialogues based on the Qwen or ChatGLM series. Final answer It also includes evidence citation and confidence assessment. The answer content is strictly derived from the provided parent node text and is labeled with a confidence level based on the evidence level, so that the retrieval and generation form a complete traceable closed loop.

[0068] In addition, such as Figure 5 As shown, the present invention also provides a medical document retrieval enhancement generation system based on a parent-child node structure, including a document parsing module, a three-layer semantic tree node construction module, a quality inspection module, an index storage module, a retrieval execution module, and a reordering module.

[0069] The document parsing module receives the original medical document, performs document structure parsing through OCR recognition and a regular expression rule engine, and outputs chapter boundary information and a sequence of medical text units segmented according to length integrity constraints. .

[0070] The three-layer semantic tree node construction module integrates the proposition decomposition submodule, paragraph aggregation submodule, and chapter aggregation submodule, which are responsible for calling the large language model respectively. Complete the extraction and type labeling of atomic proposition quadruples, construct paragraph nodes from raw text units, and call them based on chapter boundary information. Generate chapter aggregation nodes, maintain bidirectional pointer relationships between the three-level nodes, and output a complete three-level medical semantic tree node structure.

[0071] The quality control module uses the set of atomic propositions corresponding to each paragraph node. As the verification benchmark, the summary generation model is called sequentially. Proposition decomposition model And a proposition validation model based on the DeBERTa-v3 architecture to calculate clinical priority-weighted proposition coverage. Semantic summaries that fail quality checks will be regenerated if the maximum number of retries is exceeded. The original text units are then replaced, and a quality-tested search summary is output.

[0072] The indexing and database insertion module writes the semantic vectors of the three-layer nodes into the vector database using five parallel paths. Three-granularity hierarchical index partitioning ( , , The structured content of atomic proposition quadruples and paragraph keywords will be written into the full-text search database. Write the complete content of the three-layer nodes and their hierarchical affiliation mappings into the content repository. Build a multi-engine collaborative indexing system.

[0073] The retrieval execution module integrates an intent recognition submodule, a vector retrieval submodule, and a keyword retrieval submodule. It sequentially executes rule triggers and a three-class vector classifier on user queries. Two-stage intent recognition, based on intent confidence vector The query is routed to the corresponding vector partition. When the intent is uncertain, the three partitions are activated in parallel and the cross-granularity search scores are weighted and fused with intent confidence. At the same time, keyword Boolean search is performed in the full-text search library. The hit nodes at each level are uniformly reduced to paragraph nodes through bidirectional pointer tracing and the primary search result set is output.

[0074] The reordering module uses a reordering model based on the LightGBM architecture. Calculate a 12-dimensional feature vector for each candidate paragraph node. The correlation probability is output after sigmoid mapping. The final search results list is obtained by sorting the results in descending order of probability and then passing them through a content repository. Retrieve the complete original text of each paragraph node for subsequent use in a large language model. Generate the final answer with evidence level labels.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing medical document retrieval and generation based on a parent-child node structure, characterized in that, include: S1. Parse the input medical document, identify chapter boundary information, and segment the document text into a sequence of medical text units; Step S1 includes: Medical documents are converted into text using optical character recognition (OCR) methods. Regular expressions and rule engines are then used to identify chapter titles, paragraph separators, and metadata information, resulting in chapter boundary information represented by chapter numbers, title levels, and paragraph affiliation tags. S2. Using a large language model, each medical text unit is decomposed into a set of medical atomic propositions. Each atomic proposition satisfies a four-tuple structure including the applicable subject, intervention measures, conditional constraints, and conclusion and level of evidence, and is labeled with the proposition type. Using the original text of the medical text unit as paragraph nodes, each paragraph node is aggregated according to the chapter boundary information to generate chapter aggregation nodes. The atomic proposition set, paragraph nodes, and chapter aggregation nodes are respectively used as the first layer, second layer, and third layer, and interconnected through bidirectional pointers between layers to construct a three-layer medical semantic tree node. S3. The semantic summary of each paragraph node is checked by a weighted coverage quality check mechanism based on the atomic proposition set. The semantic summary that fails the quality check is triggered to be regenerated until the quality check is passed, and the retrieval summary of each paragraph node is obtained. The weighted coverage quality control mechanism is based on the set of atomic propositions. As a verification benchmark, a proposition verification model is used to verify... Each atomic proposition Determine whether each proposition is included in the set of abstract propositions. Covered; among which, Using a large language model to analyze the i-th paragraph node Corresponding semantic summary The result obtained by performing propositional decomposition Represents the i-th medical text unit The j-th atomic proposition extracted; S4. Encode the retrieval summary of the three-layer medical semantic tree node into a semantic vector and store it in the vector database according to the level. Store the atomic proposition quadruple content and paragraph keywords in the full-text retrieval database. Store the complete content of each layer node and the level attribution mapping in the content storage database. Among them, the vector database is divided into three-granularity hierarchical index partitions according to proposition level, paragraph level and chapter level. S5. Receive user queries, use a hybrid intent recognition mechanism to classify the query at a granular level, route the user query to the corresponding partition of the vector database to perform semantic retrieval based on the intent classification results, and simultaneously perform keyword retrieval in parallel in the full-text search library. Combine the two retrieval results to obtain a preliminary retrieval result set. S6. For each candidate paragraph node in the initial search result set, use the reordering model to calculate its relevance score to the user query, sort them in descending order of score, and obtain the final search result list.

2. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 1, characterized in that, Step S1 also includes: The extracted continuous text is segmented into medical text unit sequences. ,in For the i-th medical text unit, , where n is the total number of medical text units; Each medical text unit satisfies the length integrity constraint. , Medical text unit The number of characters, To preset the minimum character length threshold, This is the preset maximum character length threshold.

3. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 2, characterized in that, In step S2, for the i-th medical text unit The extracted j-th atomic proposition is denoted as Each atomic proposition must satisfy the quadruple structure. ,in For the applicable subject, These are intervention measures or test items that must be mapped to a medical terminology ontology. Standard entries in For conditions or quantitative constraints, The conclusion and its corresponding level of evidence label; proposition type label. Taken from a predefined set of proposition types These correspond to contraindication propositions, dosage constraint propositions, indication propositions, and background information propositions, respectively; text fragments that do not satisfy the complete four-tuple structure are not included in the atomic proposition set. .

4. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 2, characterized in that, In step S2, the chapter aggregation node Chapter-level aggregate summary function Structure-aware aggregation of the set of paragraph nodes belonging to the s-th chapter yields: ； in This refers to the r-th paragraph node within chapter s. This is the set of paragraph nodes contained in the s-th chapter of the document, where s is the chapter number; the aggregation boundary is determined by the chapter title and chapter number in the chapter boundary information.

5. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 3, characterized in that, The weighted coverage quality inspection mechanism in step S3 is executed as follows: For the i-th paragraph node Corresponding semantic summary The large language model is invoked to perform proposition decomposition, resulting in a set of summary propositions. ,in From The kth atomic proposition extracted from it. p represents the total number of abstract propositions; A set of atomic propositions As a verification benchmark, a proposition verification model is used to verify... Each atomic proposition Determine whether each one is... Coverage, Coverage Determination Function The value is 1 when the proposition verification model determines that the coverage is true, and 0 otherwise. Each atomic proposition The clinical importance weight is ,in Based on the proposition type label Assigned type precedence coefficient, Based on the propositional quadruple The included evidence level labels The assigned level of evidence coefficient; The clinical priority-weighted proposition coverage rate is: ； Where m is the set of atomic propositions The total number of valid atomic propositions; when Temporal Semantic Summary Through quality inspection, Set the preset coverage threshold; If the quality check fails, a regeneration will be triggered, exceeding the preset maximum number of retries. Then, using the original medical text unit Used as a substitute for the search summary.

6. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 1, characterized in that, In step S4, the vector coding model used to encode the retrieval summary into a semantic vector is denoted as... , Based on the Sentence-BERT architecture, multi-task contrastive learning training using medical text is employed, with the training objective function being: ,in To compare learning loss, For medical terminology triplet loss, Preset hyperparameters to balance the weights of the two types of losses; The contrastive learning loss function is: ； Where N is the number of training samples, and n is the training sample index. Let be the vector representation of the nth training sample. To and A semantically relevant positive sample vector, where q is the negative sample index. For the qth pair A semantically unrelated vector of negative samples, where M is the total number of negative samples. The cosine similarity function is used. A preset temperature hyperparameter for adjusting the smoothness of the similarity distribution.

7. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 1, characterized in that, The hybrid intent recognition mechanism in step S5 consists of two stages: The first stage is rule triggering. If a user query Q contains explicit numerical thresholds, precise dosages, or absolute operational semantic terms, it is determined to be a precise fact query and the user query is routed to the propositional vector partition. If the query contains comparative terms, it is classified as a review query and the user query is routed to the chapter-level vector partition. ; The second stage is the vector classifier. When the rule trigger in the first stage does not produce a deterministic decision, the user query Q is input into a three-class intent classifier fine-tuned using a medical query labeled dataset. Output the intent confidence vector: ； in This is a granular hierarchical index; prop, para, and chapter correspond to proposition-level precise fact queries, paragraph-level context queries, and chapter-level summary queries, respectively. For users to query Q corresponding to the first The intent confidence at the layer granularity satisfies ; When the maximum confidence level Not lower than the preset reliability threshold When a user query is active, route it to the partition with the highest confidence level; otherwise, activate the propositional vector partitioning. Paragraph-level vector partitioning Chapter-level vector partitioning Parallel retrieval.

8. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 7, characterized in that, In step S5, the initial search result set is obtained as follows: If the first-stage rule trigger produces a deterministic judgment, or if the highest intent confidence output by the second-stage vector classifier is not lower than the preset confidence threshold, then the user query will be routed to the corresponding single partition to perform semantic retrieval. The semantic retrieval results of that partition will be combined with the keyword retrieval results of the full-text retrieval library to form the primary retrieval result set. If the first-stage rule trigger does not produce a deterministic judgment and the highest intent confidence output by the second-stage vector classifier is lower than the preset confidence threshold, then the three partitions of proposition level, paragraph level and chapter level are activated simultaneously to perform semantic retrieval in parallel. After normalizing the retrieval scores in the candidate set of each partition, cross-granularity weighted fusion sorting is performed with the intent confidence of each level as the weight. The union of the candidate nodes after fusion sorting and the keyword retrieval results of the full-text retrieval library is taken to form the primary retrieval result set. For each level node hit in the initial search result set, the corresponding paragraph node is uniformly reduced through bidirectional pointer tracing between levels to obtain a candidate set of paragraph nodes for reordering.

9. The medical document retrieval enhancement generation method based on a parent-child node structure according to claim 8, characterized in that, In step S6, the reordering model Based on the LightGBM architecture, the model uses the vector semantic similarity between each candidate paragraph node in the primary search result set and the user query, the keyword coverage ratio of the query medical terms in the candidate nodes, the evidence level weight of the candidate nodes, and the timeliness decay weight of the candidate nodes as basic features. The model also introduces polynomial extensions of the above features to generate higher-order interaction features, which together constitute the input feature vector. The model output is mapped to the relevance probability of each candidate paragraph node through the sigmoid function, and the final search result list is obtained by sorting the relevance probabilities in descending order.

10. A medical document retrieval enhancement and generation system based on a parent-child node structure, characterized in that, The system is used to implement the method according to any one of claims 1-9, the system comprising: The document parsing module is used to parse the input medical documents, identify chapter boundary information, and segment the document text into a sequence of medical text units; The three-layer semantic tree node construction module is used to decompose each medical text unit into a set of medical atomic propositions using a large language model, aggregate paragraph nodes based on chapter boundary information to generate chapter aggregation nodes, and use the set of atomic propositions, paragraph nodes, and chapter aggregation nodes as the first, second, and third layers respectively and connect them with bidirectional pointers between layers to construct a three-layer medical semantic tree node. The quality inspection module is used to inspect the semantic summary of paragraph nodes based on the set of atomic propositions corresponding to each paragraph node and adopt a weighted coverage quality inspection mechanism. Semantic summaries that fail the quality inspection are regenerated until the quality inspection is passed, and the search summary of each paragraph node is obtained. The indexing and database module is used to encode the retrieval summary of the three-layer medical semantic tree node into a semantic vector and then store it into the vector database according to the level. It stores the content of atomic proposition quadruples and paragraph keywords into the full-text retrieval database and stores the complete content of each layer node and its level mapping into the content storage database. Among them, the vector database establishes a three-granularity hierarchical index partition according to the proposition level, paragraph level and chapter level. The retrieval execution module receives user queries, uses a hybrid intent recognition mechanism to classify the queries at a granular level, routes the user query to the corresponding partition of the vector database to perform semantic retrieval based on the intent classification results, and simultaneously performs keyword retrieval in parallel in the full-text retrieval database. The results from the two retrieval paths are combined to obtain a preliminary retrieval result set. The reordering module is used to calculate the relevance score of each candidate paragraph node in the primary search result set to the user query using a reordering model, and sort them in descending order of score to obtain the final search result list.

Images