Scientific and technical literature intelligent retrieval method based on generative artificial intelligence and related equipment

Abstract

The invention provides a scientific and technological literature intelligent retrieval method and related equipment based on generative artificial intelligence, and the method comprises the steps: carrying out the multi-layer semantic annotation of medical scientific and technological literatures, constructing a symptom-disease dynamic association map, and building a medical scientific and technological literature knowledge base; performing medical context analysis and multi-modal feature extraction on user query, and generating a unified retrieval vector in combination with Boolean operation nested analysis and semantic alignment processing; performing evidence grading retrieval and clinical scene matching based on the unified retrieval vector to obtain a preliminary candidate literature set, and optimizing the preliminary candidate literature set into a target candidate literature set through fine-grained semantic recalculation; and calculating a retrieval prior probability according to the evaluation dimension, performing knowledge weighted fusion on the candidate literature, and generating a medical science and technology literature recommendation report. According to the method, the medical term association relationship is deeply understood through the association map, the result is ensured to be matched with the patient characteristics through evidence grading retrieval and clinical scene matching and screening, and the accuracy of document retrieval is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence and related equipment. Background Art

[0002] With the rapid development of medical research and the explosive growth of medical scientific literature, intelligent retrieval of scientific literature based on generative AI has become an important tool for clinical decision-making and medical research. Traditional methods for intelligent retrieval of scientific literature based on generative AI rely primarily on keyword matching and simple semantic similarity calculations, which have significant limitations when processing complex medical queries.

[0003] Existing intelligent retrieval systems for scientific literature based on generative AI generally suffer from insufficient semantic understanding, failing to accurately grasp the complex relationships between medical terms and clinical context. Existing retrieval methods struggle to accurately match a patient's specific clinical characteristics with research scenarios in the literature, resulting in significant discrepancies between retrieval results and actual clinical needs. Medical staff must spend considerable time sifting through the vast amount of search results to identify truly valuable literature information. Summary of the Invention

[0004] The main purpose of this invention is to solve the technical problems of the existing scientific literature intelligent retrieval methods based on generative artificial intelligence, which have insufficient semantic understanding depth and cannot achieve accurate clinical scenario matching; A first aspect of the present invention provides a method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence, the method comprising: Perform multi-layer semantic annotation on medical scientific literature, build a symptom-disease dynamic association map based on the annotated medical terms, and perform semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints; The user query is parsed in medical context and multimodal features are extracted through a dual encoder architecture. The logical relationship in the user query is parsed through Boolean nested operations and standardized and semantically aligned using the symptom-disease dynamic association map and semantic fingerprints to obtain a unified retrieval vector. Performing evidence hierarchical retrieval and clinical scenario matching on the unified retrieval vector and the medical science and technology literature knowledge base to obtain a preliminary candidate document set, and performing fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set; The retrieval prior probability is calculated according to the evaluation dimension of the target candidate document set, and the candidate document set is subjected to knowledge weighted fusion according to the retrieval prior probability to obtain a medical science and technology document recommendation report.

[0005] Optionally, in a first implementation of the first aspect of the present invention, performing multi-layer semantic annotation on medical scientific literature, constructing a symptom-disease dynamic association map based on the annotated medical terms, and performing semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints includes: Identify medical terminology in medical scientific literature and use the medical ontology knowledge base to annotate medical terms to obtain medical term annotation results, where the medical terms include disease names, symptom descriptions, drug names, and treatment plans; calculating association strength weights based on co-occurrence patterns, causal relationship descriptions, and time series information of symptom terms and disease terms in the medical term annotation results; The graph structure is constructed with symptom terms and disease terms as graph nodes and association strength weights as edge weights to obtain a symptom-disease dynamic association graph; Performing semantic mapping processing on the medical term annotation results, and obtaining semantic mapping results by establishing a three-layer correspondence relationship between a professional term layer, a clinical description layer, and a patient expression layer; According to the symptom-disease dynamic association map and semantic mapping results, feature extraction is performed on medical scientific literature, and the main disease categories, core symptom groups, treatment plan types and patient population characteristics are extracted to generate semantic fingerprints, thereby obtaining a medical scientific literature knowledge base containing semantic fingerprints.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the user query is subjected to medical context parsing and multimodal feature extraction using a dual encoder architecture, and the logical relationship in the user query is subjected to Boolean nested parsing, and the symptom-disease dynamic association graph and semantic fingerprint are used for standardized conversion and semantic alignment to obtain a unified retrieval vector, which includes: The ViT-B / 16 encoder in the dual encoder architecture extracts patch-level fine-grained features from the image input in the user query, and the BERT encoder in the dual encoder architecture extracts token-level semantic features from the text query in the user query to obtain multimodal query features. Determining the query type of the user query based on the multimodal query features to obtain query context features, and performing Boolean nested parsing processing on the logical relationships in the user query to obtain a structured query logical representation; The symptom-disease dynamic association graph is used to perform standardized conversion on the patient symptom description queried by the user, and non-standard symptom expressions are converted into standard medical terms to obtain standardized symptom features; According to the semantic fingerprint, semantic alignment processing is performed on multimodal query features, standardized symptom features and structured query logic representation, and query context features are integrated to generate a unified retrieval vector.

[0007] Optionally, in a third implementation of the first aspect of the present invention, performing evidence hierarchical retrieval and clinical scenario matching on the unified retrieval vector and the medical science and technology literature knowledge base respectively to obtain a preliminary candidate document set, and performing fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set including: Performing an evidence-level search based on the unified search vector and the medical science and technology literature knowledge base, and obtaining candidate documents of evidence level by matching query complexity with document evidence level; Perform clinical scenario matching with the medical science and technology literature knowledge base based on the unified search vector, and obtain candidate clinical scenario literature by matching patient characteristics with literature clinical scenarios; Merge the evidence classification candidate documents and clinical scenario candidate documents, remove duplicate documents and sort them by search score to obtain a preliminary candidate document set; The ColBERTv2 model is used to calculate the semantic matching score between each candidate document in the preliminary candidate document set and the unified search vector. The top M documents are selected by score ranking to obtain the roughly ranked candidate documents. By extracting the patch-level and token-level features of the coarse-ranked candidate documents, calculating the maximum similarity matching score between the features, re-ranking and screening the Top-K documents, the target candidate document set is obtained.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the evidence-graded search is performed based on the unified search vector and the medical science and technology literature knowledge base, and the evidence-graded candidate documents obtained by matching the query complexity with the document evidence level include: The query complexity score is obtained by calculating the semantic depth index, symptom combination complexity index and diagnostic reasoning level index of the unified retrieval vector; Based on the query complexity score, the literature in the medical science and technology literature knowledge base is pre-labeled according to the evidence level of randomized controlled trials, systematic reviews, cohort studies, and case reports to obtain a graded labeled literature library; According to the preset mapping relationship between complexity threshold and evidence level, the query complexity score is matched with the documents of each level in the graded annotated document library, and the selection priority score of each document is calculated to obtain a priority score list; Literature with a matching degree exceeding the preset threshold is screened according to the priority score list, and then sorted according to the citation frequency of the literature to obtain candidate literature for evidence classification.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the clinical scenario matching is performed with the medical science and technology literature knowledge base based on the unified search vector, and the clinical scenario candidate documents obtained by matching patient characteristics with the clinical scenarios of the documents include: The age range, gender information, main symptoms, medical history, and medication status are extracted from the unified retrieval vector using a named entity recognition algorithm to obtain a structured patient feature vector. By analyzing the research subject description, inclusion and exclusion criteria, and clinical trial design of each article in the medical science and technology literature knowledge base, a clinical scenario feature matrix including applicable population, disease stage, and treatment plan is constructed for the article; By calculating the cosine similarity between the structured patient feature vector and the literature clinical scenario feature matrix, a list of patient-literature similarity scores was obtained; The clinical applicability credibility of each document is calculated based on the similarity score list and combined with the Bayesian confidence model. The documents with clinical applicability credibility exceeding the preset threshold are screened to obtain candidate clinical scenario documents.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the calculating of the retrieval prior probability based on the evaluation dimension of the target candidate document set and performing knowledge-weighted fusion on the candidate document set based on the retrieval prior probability to obtain a medical science and technology literature recommendation report includes: The prior probability of each document in the target candidate document set is calculated by multi-dimensional evaluation vector and softmax normalization of the evaluation dimension to obtain the prior probability distribution; Performing knowledge weighted fusion on the target candidate document set according to the prior probability distribution, and obtaining a weighted fusion knowledge representation through prior probability weighting and feature splicing operations; Generate recommended content based on the weighted fusion knowledge representation, convert it into structured recommendation text through a decoder, and obtain preliminary recommended content; The preliminary recommended content is annotated with literature citations, and citation marks and citation lists are added according to the source of the content to obtain a medical science and technology literature recommendation report.

[0011] A second aspect of the present invention provides a scientific and technological literature intelligent retrieval device based on generative artificial intelligence, the scientific and technological literature intelligent retrieval device based on generative artificial intelligence comprising: A semantic annotation module is used to perform multi-layer semantic annotation processing on medical scientific literature, construct a symptom-disease dynamic association map based on the annotated medical terms, and perform semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints; A query parsing module is used to perform medical context analysis and multimodal feature extraction on user queries using a dual encoder architecture, perform Boolean nested parsing on logical relationships in user queries, and perform standardized conversion and semantic alignment using the symptom-disease dynamic association graph and semantic fingerprints to obtain a unified retrieval vector; an intelligent retrieval module, configured to perform evidence hierarchical retrieval and clinical scenario matching respectively based on the unified retrieval vector and the medical science and technology literature knowledge base to obtain a preliminary candidate document set, and perform fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set; The fusion recommendation module is used to calculate the retrieval prior probability based on the evaluation dimension of the target candidate document set and perform knowledge weighted fusion on the candidate document set based on the retrieval prior probability to obtain a medical science and technology document recommendation report.

[0012] The third aspect of the present invention provides a scientific and technological literature intelligent retrieval device based on generative artificial intelligence, comprising: a memory and at least one processor, wherein instructions are stored in the memory, and the memory and the at least one processor are interconnected through lines; the at least one processor calls the instructions in the memory so that the scientific and technological literature intelligent retrieval device based on generative artificial intelligence performs the steps of the above-mentioned scientific and technological literature intelligent retrieval method based on generative artificial intelligence.

[0013] The fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the steps of the above-mentioned scientific and technological literature intelligent retrieval method based on generative artificial intelligence.

[0014] The above-mentioned intelligent retrieval method and related equipment for scientific and technological literature based on generative artificial intelligence establish a medical and technological literature knowledge base containing semantic fingerprints by performing multi-layer semantic annotation on medical and technological literature and constructing a symptom-disease dynamic association map; performing medical context analysis and multimodal feature extraction on user queries, combining Boolean nested analysis and semantic alignment processing to generate a unified retrieval vector; performing evidence hierarchical retrieval and clinical scenario matching based on the unified retrieval vector to obtain a preliminary candidate document set, and optimizing it to a target candidate document set through fine-grained semantic recalculation; calculating the retrieval prior probability according to the evaluation dimension, performing knowledge weighted fusion on the candidate documents, and generating a medical and technological literature recommendation report. The present invention deeply understands the association relationship between medical terms through the symptom-disease dynamic association map, ensures that the results match the patient's clinical characteristics through evidence hierarchical retrieval and clinical scenario matching screening, and improves the accuracy of literature retrieval.

[0015] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0016] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of a first embodiment of a method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence in an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of an intelligent retrieval device for scientific and technological literature based on generative artificial intelligence in an embodiment of the present invention; Figure 3 Schematic diagram of an embodiment of an intelligent retrieval device for scientific and technological literature based on generative artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] The terms "including," "having," and any variations thereof, as used in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to the process, method, product, or device.

[0020] To facilitate understanding of this embodiment, first, a method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence disclosed in an embodiment of the present invention is described in detail. Figure 1 As shown, this method includes the following steps: 101. Perform multi-layer semantic annotation on medical scientific literature, construct a symptom-disease dynamic association map based on the annotated medical terms, and perform semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints; In one embodiment of the present invention, the multi-layer semantic annotation processing of medical scientific literature, constructing a symptom-disease dynamic association graph based on the annotated medical terms, and performing semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints includes: performing medical professional terminology recognition processing on the medical scientific literature, performing entity annotation on the medical terms through a medical ontology knowledge base to obtain medical terminology annotation results, wherein the medical terms include disease names, symptom descriptions, drug names, and treatment plans; calculating association strength weights based on co-occurrence patterns, causal relationship descriptions, and time series information of symptom terms and disease terms in the medical terminology annotation results; constructing a graph structure using symptom terms and disease terms as graph nodes and using association strength weights as edge weights to obtain a symptom-disease dynamic association graph; performing semantic mapping on the medical terminology annotation results to obtain a semantic mapping result by establishing a three-layer correspondence between a professional terminology layer, a clinical description layer, and a patient expression layer; performing feature extraction on the medical scientific literature based on the symptom-disease dynamic association graph and the semantic mapping results, extracting major disease categories, core symptom groups, treatment plan types, and patient population characteristics to generate semantic fingerprints, and obtaining a medical scientific literature knowledge base containing semantic fingerprints.

[0021] Specifically, the multi-layer semantic annotation process for medical scientific literature utilizes a BiLSTM-CRF (Bidirectional Long Short-Term Memory Network-Conditional Random Field) model combined with a medical ontology knowledge base to extract and annotate terms. BiLSTM-CRF is a sequence annotation model in which the BiLSTM layer captures contextual information in the text, while the CRF layer ensures global consistency in the annotation sequence. This combination is particularly well-suited for addressing the ambiguous terminology boundaries and strong contextual dependencies found in medical text. The medical ontology knowledge base, centered around the internationally standardized Medical Subject Headings (MeSH), covers standard terminology systems for diseases, drugs, and genes, and also integrates authoritative medical terminology libraries such as UMLS and SNOMED CT. During term recognition, the system first uses the pre-trained BioBERT model to generate word vector representations for medical scientific literature. The BiLSTM-CRF model then identifies key medical entities such as disease names, symptom descriptions, drug names, and treatment regimens. Identified terms are mapped to standard concepts in the medical ontology through string matching and semantic similarity calculation. Each term is assigned a corresponding CUI code (Concept Unique Identifier) ​​and semantic type label, resulting in structured medical term annotation results.

[0022] Specifically, based on the medical terminology annotation results, the system uses the Pointwise Mutual Information (PMI) algorithm to calculate the strength of association between symptom terms and disease terms. The PMI algorithm measures the degree of association by calculating the ratio of the probability of two terms appearing together to the product of their independent probabilities. This algorithm effectively identifies term pairs with true semantic connections while filtering out accidental co-occurrences. Co-occurrence pattern analysis is performed at the sentence, paragraph, and document granularity levels. The system uses a sliding window mechanism to calculate co-occurrence frequencies within different distance ranges, assigning higher weight to co-occurrences with closer distances. Causal relationship descriptions are extracted using a dependency parsing approach. The system uses the Stanford CoreNLP toolkit to construct a sentence dependency syntactic tree. It then matches and identifies predefined causal relationship patterns, including explicit causal markers such as "lead to," "cause," and "due to," as well as implicit causal relationships within subject-verb-object structures. Time series information processing uses time expression recognition and standardization technology. The system first uses regular expressions and dictionary matching to identify time expressions, and then converts relative time into time points on the absolute timeline through the TimeML standard to build a temporal model of symptom development. This processing method can capture the order and time interval information of symptoms during disease progression.

[0023] Specifically, the symptom-disease dynamic association graph is constructed using the NetworkX graph computing framework. Each node in the graph contains attribute information such as term ID, standard name, semantic type, and clinical frequency. During the graph construction process, the system uses the Louvain community discovery algorithm to identify clustering patterns of symptom-disease associations. The Louvain algorithm is a community detection method based on modularity optimization that can automatically discover closely related groups of nodes in the graph and combine closely related terms to form disease syndromes. The dynamic update mechanism of the graph is based on an incremental learning algorithm. When new documents are added, the system only needs to recalculate the local graph structure involving the new terms, without having to reconstruct the entire graph. This design significantly improves the scalability and real-time performance of the system. The semantic mapping process uses a multi-level semantic equivalence network. The professional terminology layer uses MeSH descriptors as the standard representation. The clinical description layer is constructed by mining physician description text in electronic medical records. The patient expression layer is established based on user expression data from online health forums and question-and-answer platforms. The mapping relationship between the three layers is established through a combination of Word2Vec word vector similarity calculation and manual verification. Word2Vec is a word embedding technology that can map words into a high-dimensional vector space. Words with similar semantics are closer in the vector space. The system establishes cross-layer mapping relationships by calculating the cosine similarity between terms at different levels.

[0024] Specifically, semantic fingerprint generation utilizes the Graph Convolutional Network (GCN) technology, which takes the symptom-disease association graph as input and learns low-dimensional representation vectors for nodes through multi-layer graph convolution operations. GCN is a deep learning model specifically designed for processing graph-structured data. It effectively aggregates node neighborhood information, bringing semantically related medical concepts closer together in vector space. This property is particularly well-suited for processing complex inter-conceptual relationships in the medical field. The system performs graph convolution on the disease, symptom, treatment, and patient feature nodes mentioned in each article, generating feature vectors of four different dimensions. Major disease categories are determined by scoring the importance of disease nodes in the article. This score is calculated based on node degree centrality and the PageRank algorithm. Degree centrality reflects the degree of connectivity of nodes in the graph, while the PageRank algorithm considers the propagation of node importance. Core symptom clusters are identified using the density-based clustering algorithm DBSCAN. This algorithm automatically identifies typical symptom combination patterns based on the distribution density of symptom nodes in the graph, eliminating the need to pre-specify the number of clusters. It is particularly suitable for detecting irregularly shaped symptom clusters. Treatment plan type extraction maps specific treatment measures to upper-level treatment categories by constructing a treatment ontology hierarchy. The system uses a hierarchical clustering method to construct a treatment classification system based on the similarity of treatment effects and mechanisms of action. Patient population feature extraction uses named entity recognition technology to extract demographic information from the methodology section of the literature, and then obtains feature distributions such as age distribution, gender ratio, and disease severity through statistical analysis. The four-dimensional feature vectors are concatenated and processed through principal component analysis (PCA) to generate a 512-dimensional semantic fingerprint vector. PCA is a linear dimensionality reduction technology that can reduce vector dimensions while retaining the main information and improve computational efficiency. The semantic fingerprint not only retains the core medical semantics of the literature, but also has good computational performance.

[0025] 102. Perform medical context analysis and multimodal feature extraction on user queries using a dual encoder architecture, perform Boolean nested analysis on logical relationships in user queries, and utilize the symptom-disease dynamic association graph and semantic fingerprints for standardized conversion and semantic alignment to obtain a unified retrieval vector. In one embodiment of the present invention, the method of performing medical context parsing and multimodal feature extraction on user queries through a dual encoder architecture, performing Boolean nested parsing on logical relationships in user queries, and performing standardized conversion and semantic alignment using the symptom-disease dynamic association map and semantic fingerprint to obtain a unified retrieval vector includes: extracting patch-level fine-grained features from image inputs in user queries through the ViT-B / 16 encoder in the dual encoder architecture, and extracting token-level semantic features from text queries in user queries through the BERT encoder in the dual encoder architecture to obtain multimodal query features; determining the query type of the user query based on the multimodal query features to obtain query context features, and performing Boolean nested parsing on logical relationships in the user query to obtain a structured query logical representation; performing standardized conversion on patient symptom descriptions in user queries using the symptom-disease dynamic association map, converting non-standard symptom expressions into standard medical terms to obtain standardized symptom features; performing semantic alignment on multimodal query features, standardized symptom features, and structured query logical representations based on the semantic fingerprint, and fusing query context features to generate a unified retrieval vector.

[0026] Specifically, the medical context parsing and multimodal feature extraction for user queries utilizes a dual-encoder architecture consisting of a ViT-B / 16 (Vision Transformer Base / 16) encoder and a BERT (Bidirectional Encoder Representations from Transformers) encoder, with each encoder responsible for processing visual and textual information, respectively. The ViT-B / 16 encoder specifically processes image inputs in user queries, including visual information such as medical images, photos of skin lesions, and images of medical examination reports. It segments the input image into 16×16 pixel patches, each of which is considered an independent visual unit. This patch segmentation technique draws on the token concept in the Transformer architecture, transforming image processing into a sequence processing problem. Each patch is converted into a fixed-dimensional vector representation through a linear projection layer, and positional encoding information is then added to preserve spatial relationships. Through the self-attention mechanism of the multi-layer Transformer encoder, the ViT-B / 16 encoder captures long-range dependencies between different patches and extracts high-level semantic information that reflects the characteristics of the medical image. At the same time, the BERT encoder processes the text portion of the user query, including natural language expressions such as symptom descriptions, medical history information, and treatment needs. BERT uses WordPiece word segmentation technology to break down the text into sub-word units (tokens), and each token is converted into a vector representation through a pre-trained word embedding layer. BERT's bidirectional encoding mechanism enables it to simultaneously consider the contextual information of a token, a feature that is particularly important for understanding complex semantic relationships in medical text. After processing by the multi-layer Transformer encoder, BERT outputs token-level feature vectors containing rich semantic information. These vectors integrate multiple levels of information such as lexical semantics, syntactic structure, and context.

[0027] Specifically, query type determination is based on multimodal query features through classification and identification. This process utilizes a multi-head classifier to accurately identify query intent. Based on the patterns of feature vectors, the classifier identifies different query types, including diagnostic consultation, treatment plan query, drug information retrieval, and prognosis assessment. Each type corresponds to a different contextual processing strategy. Diagnostic consultation queries focus on the association between symptoms and diseases, treatment plan queries emphasize the effectiveness of interventions, drug information retrieval focuses on drug mechanisms of action and side effects, and prognosis assessment emphasizes disease progression and risk factors. Query context features include structured information such as query type labels, urgency ratings, and professional field classifications. These features provide important insights for subsequent search strategy adjustments. Nested Boolean operations are parsed using a recursive descent parsing algorithm, which is capable of handling complex logical expression structures. The parsing process first identifies logical connectives in the query, including basic logical operators such as AND, OR, and NOT, as well as the precedence relationships indicated by parentheses. The recursive descent parser constructs an abstract syntax tree based on operator precedence. Each node in the tree represents a logical operation or operand, and leaf nodes correspond to specific medical concepts or conditions. The parser supports complex, nested logical expressions, accurately understanding complex queries such as "(diabetes AND hypertension) OR (heart disease AND NOT kidney disease)." The parsing results form a structured logical representation of the query, preserving the logical semantics of the original query while providing an actionable data structure for subsequent search execution.

[0028] Specifically, the standardization conversion process utilizes the symptom-disease dynamic association graph to normalize patient symptom descriptions. This process addresses the retrieval challenges presented by the diverse terminology in the medical field. Patients often use everyday terms such as "headache," "stomach ache," and "short breath" when describing their symptoms, which can differ significantly from standard medical terminology. The standardization conversion employs a graph matching algorithm to identify the standard terminology node within the symptom-disease dynamic association graph that best matches the patient's description. The algorithm first calculates the semantic similarity between the patient's symptom description and each symptom node in the graph, using a weighted combination of the cosine distance and edit distance of word vectors. The algorithm then identifies the optimal terminology mapping path by combining edge weights in the graph and considering the strength of associations between symptoms. For example, a patient's description of "chest tightness and pain" would be mapped to a combination of the standard terms "chest pain" and "dyspnea," while "dizziness and lightheadedness" would be mapped to standard descriptors such as "vertigo" and "dizziness." This conversion not only standardizes terminology but also supplements relevant symptom information that was not explicitly expressed by the patient through the association information in the graph. Standardized symptom features are represented in vector form, where each dimension corresponds to a standard medical term, and the vector value reflects the importance and confidence of the symptom in the user query.

[0029] Specifically, semantic alignment utilizes a multimodal fusion transformer architecture, specifically designed to handle the alignment and fusion of heterogeneous features. The alignment process first projects multimodal query features, standardized symptom features, and structured query logical representations into a unified semantic space. A learned transformation matrix then maps features from different modalities into vector representations of the same dimension. Semantic fingerprints play a key role in this process, serving as prior knowledge in the medical field to guide the direction and objectives of feature alignment. Specifically, semantic fingerprints provide the standard positions of medical concepts in the semantic space. The alignment process aligns query features toward these standard positions, ensuring that semantically similar concepts are closely spaced in the vector space. The alignment algorithm employs an attention mechanism to calculate the association weights between different feature components. High-weighted associations indicate that feature components are semantically highly correlated and warrant greater attention during the fusion process. Query contextual features participate in the fusion process as conditional information, and different fusion strategies are employed for different query types. Diagnosis queries prioritize the contribution of symptom features, treatment queries emphasize the importance of logical relationships, and medication queries emphasize the synergistic effect of multimodal information. The fused transformer learns the complex interactions between features through a multi-head self-attention mechanism, while the cross-attention layer integrates information from different modalities. The resulting unified retrieval vector is a high-dimensional, dense vector that encodes the complete semantic information of the user query, including textual semantics, visual features, logical structure, standardized symptoms, and contextual information. This unified retrieval vector boasts excellent semantic expression and computational efficiency, preserving the rich information of the original query while providing an optimized data representation for subsequent similarity calculations and retrieval matching.

[0030] 103. Performing evidence hierarchical retrieval and clinical scenario matching on the unified retrieval vector and the medical science and technology literature knowledge base to obtain a preliminary candidate document set, and performing fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set; In one embodiment of the present invention, the method of performing evidence hierarchical retrieval and clinical scenario matching respectively on the unified retrieval vector and the medical science and technology literature knowledge base to obtain a preliminary candidate document set, and performing fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set includes: performing evidence hierarchical retrieval on the unified retrieval vector and the medical science and technology literature knowledge base to obtain evidence hierarchical candidate documents by matching query complexity with document evidence level; performing clinical scenario matching on the unified retrieval vector and the medical science and technology literature knowledge base to obtain clinical scenario candidate documents by matching patient characteristics with document clinical scenarios; merging the evidence hierarchical candidate documents and clinical scenario candidate documents, removing duplicate documents and sorting them by retrieval score to obtain a preliminary candidate document set; calculating the semantic matching score of each candidate document in the preliminary candidate document set with the unified retrieval vector by the ColBERTv2 model, screening the top M documents by score sorting, and obtaining coarse-ranked candidate documents; extracting patch-level and token-level features of the coarse-ranked candidate documents, calculating the maximum similarity matching score between the features, re-sorting and screening the Top-K documents, and obtaining a target candidate document set.

[0031] Specifically, evidence-based retrieval employs a dual mechanism of query complexity assessment and literature evidence level matching. This process first performs a complexity analysis on a unified search vector, assessing query complexity by calculating the distribution pattern and association density of different semantic components in the vector. The query complexity assessment algorithm analyzes dimensions such as the number of symptom combinations in the vector, the hierarchical depth of disease associations, the level of nested logical operations, and the degree of integration of multimodal information. Simple queries involve a single symptom or a clear diagnosis, while complex queries involve multiple symptom combinations, rare diseases, multi-system diseases, or complex treatment options. Based on a mapping between a preset complexity threshold and evidence level, the algorithm matches the query complexity score with the literature evidence level in the medical and scientific literature knowledge base. Literature evidence levels are categorized according to evidence-based medicine standards into randomized controlled trials, systematic reviews, cohort studies, case-control studies, and case reports, each corresponding to different levels of evidence strength and clinical credibility. High-complexity queries prioritize systematic reviews and high-quality randomized controlled trials, as these studies have stronger evidence strength and wider applicability, providing reliable evidence-based support for complex clinical problems. Moderate-complexity queries match cohort and case-control studies, which can provide valuable clinical evidence under specific conditions. Low-complexity queries can obtain useful information from case reports and clinical guidelines. The matching process uses a weighted scoring mechanism, taking into account factors such as query complexity matching, literature quality rating, publication time, and citation frequency. This generates an evidence grading score for each candidate article, ultimately forming a collection of candidate articles with a graded evidence score.

[0032] Specifically, clinical scenario matching achieves precise screening by deeply analyzing the similarities between patient characteristics and clinical scenarios in literature. This process extracts key clinical characteristics of patients from a unified search vector, including age range, gender, primary symptoms, medical history, medication use, disease severity, and other multi-dimensional information. Feature extraction utilizes a multi-label classifier to parse the different dimensions of the vector to identify the patient's specific clinical profile. Furthermore, a clinical scenario feature matrix is ​​pre-constructed for each article in the medical and scientific literature knowledge base. This matrix describes key information such as the study population characteristics, inclusion and exclusion criteria, disease stage, and treatment regimen. This clinical scenario feature matrix is ​​automatically extracted from the methodology section of the article using natural language processing techniques and combined with manual annotation for quality control. The matching algorithm utilizes multi-dimensional similarity calculations to assess the degree of match between patient characteristics and clinical scenarios in the literature across demographic, disease, and treatment context dimensions. Demographic matching considers overlap in age distribution and consistency in sex ratio. Disease feature matching analyzes similarity in symptom combinations, disease severity, and complications. Treatment context matching assesses the consistency of factors such as previous treatment history, medication use, and treatment response. The algorithm also incorporates a Bayesian inference mechanism to calculate the confidence level of a patient's characteristics belonging to a specific clinical scenario based on their probability distribution. The matching results form a patient-document similarity matrix, where each element represents the strength of the match between the corresponding patient characteristics and the clinical scenario in the literature. Candidate clinical scenario documents are identified through threshold screening and sorting.

[0033] Specifically, the candidate literature merging process utilizes a combined deduplication and ranking strategy. First, duplicates are jointly deduplicated for both evidence grading and clinical scenario candidate literature. The deduplication algorithm uses unique document identifiers, such as DOIs and PubMedIDs, for precise matching. Furthermore, document fingerprinting technology is employed to identify duplicates with identical content but different identifiers. Document fingerprinting uses a hashing algorithm to encode key information, such as the title, abstract, and keywords, to generate a unique digital fingerprint. Identical or highly similar documents have identical or similar fingerprint values. The deduplicated literature collection is then ranked according to a comprehensive search score, which incorporates multiple evaluation dimensions, including evidence grading scores, clinical scenario matching scores, literature quality indicators, and timeliness weighting. This score fusion utilizes a weighted linear combination model, with weighting parameters dynamically adjusted based on query type and user preferences. Diagnostic queries prioritize clinical scenario matching, review queries emphasize the importance of evidence grade, and therapeutic queries balance the contributions of both dimensions. The ranked collection of literature forms a preliminary candidate literature collection, ensuring both the quality of the evidence and its high relevance to the patient's clinical situation.

[0034] Specifically, fine-grained semantic recalculation employs a two-stage ranking optimization strategy. The first stage uses the ColBERTv2 model for coarse ranking. ColBERTv2 is an efficient retrieval model based on BERT. This model encodes the query and document separately as multiple vector sequences, rather than a single dense vector. Specifically, ColBERTv2 processes the unified query vector and the text content of the candidate document separately through the BERT encoder, generating a sequence of token-level vector representations. Each token corresponds to a vector, and the vector sequence preserves the fine-grained semantic information and positional relationships of the text. Similarity calculation uses a maximum similarity matching strategy. For each token vector in the query, the vector with the highest similarity is searched among all token vectors in the document for a match. The maximum similarity values ​​of all tokens are then summed to obtain the overall matching score. This calculation method captures the precise semantic correspondence between the query and the document, identifying semantic relevance even when the lexical expressions are not completely identical. The advantage of ColBERTv2 lies in its excellent balance between computational efficiency and retrieval accuracy. By pre-computing document vectors and using an efficient vector retrieval algorithm, it can quickly perform similarity calculations on large-scale document repositories. The coarse ranking results are sorted in descending order by matching score, and the top M documents are selected for the refined ranking stage. The second stage of refined ranking utilizes deep matching technology using patch- and token-level features to extract more detailed semantic features from the coarse ranking candidate documents. Patch-level feature extraction is applied to non-textual content within the document, such as charts, formulas, and structured data, converting these contents into vector representations through image segmentation and feature encoding techniques. Token-level features perform a deeper semantic analysis of the document's textual content, including entity relationship extraction, semantic role labeling, and sentiment analysis. The maximum similarity matching between features utilizes a bidirectional attention mechanism, calculating not only the matching degree between query features and document features, but also the matching degree between document features and query features. This bidirectional matching ensures the completeness and accuracy of semantic correspondences. The refined ranking algorithm comprehensively considers multiple factors, including patch matching scores, token matching scores, and global semantic consistency, to generate the final ranking score. Based on a preset threshold, the top-K documents are selected to form a target candidate document set, which represents the most relevant and highest-quality medical and scientific literature recommendations for the user's query.

[0035] Furthermore, the evidence graded retrieval is performed based on the unified retrieval vector and the medical science and technology literature knowledge base, and the evidence graded candidate documents are obtained by matching the query complexity with the document evidence level, including: obtaining the query complexity score by calculating the semantic depth index, symptom combination complexity index and diagnostic reasoning level index of the unified retrieval vector; pre-labeling the documents in the medical science and technology literature knowledge base according to the evidence level of randomized controlled trials, systematic reviews, cohort studies, and case reports based on the query complexity score to obtain a graded labeled document library; matching the query complexity score with the documents of each level in the graded labeled document library based on the preset complexity threshold and evidence level mapping relationship, calculating the selection priority score of each document, and obtaining a priority score list; screening the documents with a matching degree exceeding the preset threshold according to the priority score list, and sorting them according to the document citation frequency to obtain the evidence graded candidate documents.

[0036] Specifically, the query complexity score is calculated by extracting complexity metrics across three core dimensions from the unified search vector. This multi-dimensional assessment method comprehensively reflects the complexity of medical queries and the strength of evidence required. The semantic depth metric is calculated by analyzing the abstraction level and semantic association depth of medical concepts in the unified search vector. This metric utilizes the medical ontology hierarchy to assess the average and maximum depth of the involved concepts within the MeSH tree. Specifically, queries involving superficial symptoms such as "fever" and "headache" have relatively shallow semantic depth, while queries involving complex pathophysiological mechanisms, molecular biological processes, or refined diagnostic classifications experience significantly higher semantic depth. The algorithm traverses the activated medical concept nodes in the vector, calculates the path length from each concept to the root node in the ontology hierarchy, and then uses a weighted average method to derive an overall semantic depth value. Weights are assigned based on the activation strength of the concepts in the vector, with concepts with higher activation strength contributing more to the final depth calculation. The symptom combination complexity metric assesses the association patterns and combinatorial complexity between symptoms in a query. This metric is calculated by analyzing the connectivity and clustering patterns within the symptom-disease dynamic association graph. The algorithm first identifies all symptom concepts contained in the vector and then constructs a subgraph structure of these symptoms within the association graph. The complexity of the subgraph is measured using graph-theoretic metrics such as connectivity, clustering coefficient, and shortest path length. Highly interconnected symptom combinations indicate complex clinical syndromes that require a higher level of medical evidence, while isolated or weakly connected symptom combinations are relatively simple. The diagnostic reasoning level metric reflects the complexity of the medical reasoning involved in the query. This metric is calculated by analyzing the nesting depth of logical operations, the complexity of conditional dependencies, and the length of the reasoning chain. Simple, direct symptom-disease correspondences have a lower reasoning level, while queries involving differential diagnosis, multi-step reasoning, and probabilistic inference have a higher reasoning level. The algorithm parses the logical structure encoded in the vector, calculating the nesting level of Boolean operations, the number of conditional branches, and the complexity of the reasoning steps. These three metrics are normalized and weighted to produce a comprehensive query complexity score.

[0037] Specifically, the pre-annotation of the evidence level in the medical and scientific literature knowledge base utilizes an automated classification technique based on document features. This process assigns a grade of evidence label to each article in the knowledge base, providing a standardized grading basis for the subsequent matching process. The pre-annotation algorithm first extracts key features of the article, including study design type, sample size, research methods, statistical analysis methods, and the impact factor of the published journal. Study design type identification is achieved through a combination of keyword matching and a machine learning classifier. The algorithm searches for specific study design identifiers in the title, abstract, and methodology section of the article, such as "randomized controlled trial," "systematic review," "cohort study," and "case report." Simultaneously, a multi-class classifier is trained to automatically categorize the article content. The classifier learns text feature patterns for different study types based on pre-annotated training data. Sample size is extracted from the article using digital entity recognition technology. The algorithm identifies numerical information describing the number of study subjects, such as numbers near words like "participants," "patients," and "subjects." Statistical analysis methods are identified through the construction of a statistical method dictionary and pattern matching. Common statistical methods include t-tests, chi-square tests, regression analysis, and survival analysis. The impact factor information of the published journals is obtained by matching with the journal database. Literature published in journals with high impact factors receives a higher quality rating. These features are processed through feature engineering and input into a multi-layer perceptron classifier. The classifier outputs the probability distribution of each document belonging to different levels of evidence. The document's evidence level labeling is determined based on the highest probability. The hierarchically labeled literature library is organized according to the evidence pyramid structure of evidence-based medicine. The highest level is systematic reviews and meta-analyses, followed by randomized controlled trials, then cohort studies and case-control studies, and the lowest level is case reports and expert opinions.

[0038] Specifically, the mapping between complexity thresholds and evidence levels is pre-established based on evidence-based medicine principles and clinical practice experience. This mapping reflects the varying strength of evidence requirements for queries of varying complexity. Low-complexity queries involve typical symptoms of common diseases or standard treatment regimens. These queries can be adequately supported by case reports, clinical guidelines, or low-level studies. Moderate-complexity queries involve multiple symptom combinations, differential diagnosis of diseases, or comparative treatment options, requiring moderate-strength evidence from cohort or case-control studies. High-complexity queries involve rare diseases, complex syndromes, innovative treatments, or multifactorial interactions, requiring high-strength evidence from randomized controlled trials or systematic reviews. The mapping algorithm compares the query complexity score with a pre-set threshold range to determine the complexity category to which the query belongs. Based on this correspondence, a matching evidence level range is then selected. The priority score is calculated by comprehensively considering factors such as the degree of complexity match, evidence level weighting, and literature quality indicators. The degree of complexity match is measured by calculating the overlap between the query complexity score and the applicable complexity range of the literature; a higher overlap indicates a better match. Evidence level weighting is assigned based on the evidence strength scale of evidence-based medicine, with higher-level evidence receiving a higher base weight. Literature quality indicators include journal impact factor, number of citations, and publication time. Recently published, high-quality literature receives additional quality points. Priority scores are calculated using a weighted linear combination model. The score list is sorted from high to low priority, providing a quantitative basis for the subsequent screening process.

[0039] Specifically, candidate literature for evidence grading is screened using a strategy combining threshold filtering and multidimensional ranking. This process ensures that selected literature meets the query complexity requirements while possessing high academic and clinical value. Threshold screening, based on a preset minimum priority score, eliminates literature with a low match to the query. The threshold setting considers a balance between query type, user needs, and search accuracy. Diagnostic queries use a higher matching threshold to ensure search accuracy, while exploratory queries use a lower threshold to improve search coverage. Citation frequency ranking prioritizes the academic impact of the literature as a key ranking factor. High citation frequency reflects a literature's recognition and practical application value within the academic community. The algorithm obtains citation data for each literature, including total citations, average annual citations, and the h-index. These metrics are normalized and then fused with the priority score. The fusion algorithm employs a dynamic weighting mechanism: newly published literature is given a higher weight than citation frequency, while citation frequency is given a lower weight. Citation frequency is gradually weighted more for older literature. The ranking also considers the timeliness of the literature. Medical knowledge is rapidly updated, and recently published literature has greater clinical guidance. The algorithm uses a time decay function to adjust the weight of older publications. The final set of candidate literature for evidence grading includes high-quality literature that highly matches the query complexity, has an appropriate level of evidence, and has significant academic influence. These literatures provide a reliable evidence base for subsequent clinical scenario matching.

[0040] Furthermore, the clinical scenario matching is performed based on the unified retrieval vector and the medical science and technology literature knowledge base, and clinical scenario candidate documents are obtained by matching patient characteristics with the clinical scenarios of the documents, including: extracting age range, gender information, main symptoms, past medical history and medication status from the unified retrieval vector through a named entity recognition algorithm to obtain a structured patient feature vector; by analyzing the research object description, inclusion and exclusion criteria and clinical trial design of each document in the medical science and technology literature knowledge base, a clinical scenario feature matrix containing applicable population, disease stage and treatment plan is constructed for the document; by calculating the cosine similarity between the structured patient feature vector and the clinical scenario feature matrix of the document, a patient-document similarity score list is obtained; according to the similarity score list and in combination with the Bayesian confidence model, the clinical applicability credibility of each document is calculated, and the documents with clinical applicability credibility exceeding a preset threshold are screened to obtain clinical scenario candidate documents.

[0041] Specifically, the construction of structured patient feature vectors utilizes deep learning-based named entity recognition technology, which is specifically optimized for entity extraction in the medical field. The named entity recognition algorithm first decodes the unified search vector, converting the high-dimensional, dense vector back into a parsable semantic representation. This process utilizes a reverse mapping technique, using a trained decoder to convert the numerical representations in the vector space into medical concepts and descriptive text. Age range extraction is achieved by identifying the temporal and numerical information encoded in the vector. The algorithm searches for age-related semantic patterns, including explicit numeric expressions such as "65 years old" and "45-50 years old," as well as descriptive age expressions such as "middle-aged," "elderly," and "adolescent." These descriptive expressions are converted into specific numerical intervals using a pre-built age mapping dictionary. The algorithm also considers the ambiguity and cultural differences in age expressions, using probability distributions to represent age uncertainty. Gender information extraction utilizes gender-related linguistic markers and medical terminology. The algorithm identifies explicit gender-descriptive terms and simultaneously analyzes gender-related medical symptoms and disease patterns, such as gynecological diseases and prostate problems, for gender-specific health issues. The extraction of main symptoms uses a medical symptom ontology library for semantic matching. The algorithm compares the semantic activation pattern of symptoms in the vector with the standard symptom term library to identify the core symptom manifestations involved in the patient's description. Symptom extraction not only includes the symptoms explicitly expressed by the patient, but also infers related implicit symptoms through the symptom association map. The extraction of past medical history information focuses on historical information that has a significant impact on the current health status, such as chronic diseases, major surgical history, and genetic diseases. The algorithm uses time series analysis to identify the chronological order and duration of disease occurrence. Medication extraction includes information such as current medication, past medication, and drug allergy history. The algorithm obtains a complete medication portrait through techniques such as drug name recognition, dosage extraction, and medication time analysis. These extracted information are standardized and vectorized to form a multi-dimensional structured patient feature vector. Each dimension of the vector corresponds to a specific patient characteristic attribute, and the vector value reflects the presence and importance of the feature.

[0042] Specifically, the construction of a clinical scenario feature matrix is ​​achieved through in-depth analysis of the methodology and study design information in medical and scientific literature. This process requires extracting structured clinical scenario descriptions from unstructured literature text. Analysis of study subject descriptions utilizes information extraction techniques, focusing on the demographic, clinical, and inclusion criteria of study participants. The algorithm identifies key descriptive phrases and statistical data, such as "mean age," "male-female ratio," "disease duration," and "disease severity," to extract a basic profile of the study population. Analysis of inclusion and exclusion criteria utilizes a combination of rule extraction and semantic parsing. The algorithm identifies explicit inclusion and exclusion criteria in the literature, which describe the patient populations for which the study is applicable and the circumstances in which it is not applicable. Inclusion criteria analysis encompasses multiple dimensions, including disease diagnostic criteria, symptom severity, age range, and gender requirements. Exclusion criteria include constraints such as comorbidities, medication conflicts, and special physiological conditions. Clinical trial design analysis identifies information such as study type, intervention, control setting, and follow-up duration to understand the applicable scenario of the study. Different study designs correspond to different clinical application conditions and evidence strength. The characteristics of the applicable population are determined by comprehensively considering the description of the research subjects and the inclusion and exclusion criteria. The algorithm calculates the distribution of the research population in various characteristic dimensions, such as the mean and standard deviation of the age distribution, the gender ratio, and the graded distribution of disease severity. Disease stage information is obtained by analyzing the descriptions of disease progression, staging, prognosis, etc. in the literature. Different disease stages correspond to different treatment strategies and prognostic expectations. Treatment plan characteristics include detailed information such as the type of intervention, dosage, course of treatment, and combination medication. This information determines the scope of application and clinical translation value of the research results. The clinical scenario feature matrix organizes this information in the form of a multidimensional array. The rows of the matrix correspond to different characteristic dimensions, and the columns correspond to different characteristic values ​​or characteristic intervals. The matrix elements represent the values ​​or probability distribution of the corresponding characteristics of the literature.

[0043] Specifically, the patient-literature similarity calculation uses the cosine similarity algorithm to measure the degree of match between the structured patient feature vector and the clinical scenario feature matrix of the literature. This calculation method is effective for comparing similarity between high-dimensional sparse vectors. Cosine similarity measures the directional similarity between two vectors by calculating the cosine of the angle between them. This method is unaffected by vector length and is particularly suitable for comparing features of different dimensions and value ranges. During the calculation process, the algorithm first converts the clinical scenario feature matrix of the literature into a vector representation of the same dimension as the patient feature vector. This conversion process utilizes feature alignment and interpolation techniques. For continuous features such as age and disease duration, the algorithm calculates the degree of overlap between the patient feature values ​​and the literature feature distribution. A higher degree of overlap indicates greater similarity. For discrete features such as gender and disease type, the algorithm uses either exact matching or fuzzy matching to calculate similarity. Feature weights are assigned based on their importance to clinical decision making, with core diagnostic features and key risk factors receiving higher weights, while secondary demographic features receive relatively lower weights. The similarity calculation also considers the hierarchical structure of features. For example, symptoms can be organized by organ system, severity, and other levels. The algorithm calculates similarity at different levels and performs a weighted fusion. The results form a list of patient-document similarity scores. Each element in the list corresponds to the degree of match between a document and the current patient. The scores range from 0 to 1, with higher scores indicating a better match.

[0044] Specifically, the application of the Bayesian confidence model further enhances the accuracy and reliability of clinical applicability assessment. This model integrates prior knowledge and observational evidence to calculate the confidence level of clinical applicability of each article for a specific patient. The prior probability of the Bayesian model is derived from statistical analysis of large-scale clinical data and reflects the general distribution of applicability across different types of articles. The model considers the impact of factors such as study type, sample size, study quality, and publication time on applicability. Randomized controlled trials generally have higher prior applicability probabilities, while case reports have relatively lower prior probabilities. Observational evidence is derived from patient-article similarity scores, which serve as input to the likelihood function to update the prior probability. The Bayesian update process utilizes Monte Carlo sampling, a method capable of handling complex probability distributions and high-dimensional parameter spaces. The model also incorporates an uncertainty quantification mechanism, calculating not only the confidence level of the point estimate but also providing confidence intervals to indicate the degree of uncertainty in the estimate. The calculation of clinical applicability confidence comprehensively considers factors such as patient profile matching, study quality assessment, and the difficulty of clinical translation. A high confidence level indicates that the research results of the article have strong clinical guidance value for current patients. Threshold screening is based on a balance between clinical safety and effectiveness. Setting a threshold too low will include too many irrelevant studies, while setting a threshold too high will miss valuable research. The final set of candidate clinical scenario literature selected represents research literature that is highly compatible with the patient's clinical situation and has strong clinical application value. These literature provides personalized evidence-based support for clinical decision-making.

[0045] 104. Calculate the retrieval prior probability based on the evaluation dimension of the target candidate document set and perform knowledge weighted fusion on the candidate document set based on the retrieval prior probability to obtain a medical science and technology document recommendation report.

[0046] In one embodiment of the present invention, the retrieval prior probability is calculated according to the evaluation dimension of the target candidate document set, and the candidate document set is subjected to knowledge weighted fusion according to the retrieval prior probability to obtain a medical science and technology literature recommendation report, including: calculating the prior probability of each document in the target candidate document set through the multi-dimensional evaluation vector of the evaluation dimension and softmax normalization to obtain a prior probability distribution; performing knowledge weighted fusion on the target candidate document set according to the prior probability distribution, and obtaining a weighted fusion knowledge representation through prior probability weighting and feature splicing operations; generating recommended content for the weighted fusion knowledge representation, and converting it into a structured recommendation text through a decoder to obtain preliminary recommended content; annotating the preliminary recommended content with document citations, adding reference tags and reference lists according to the content source, and obtaining a medical science and technology literature recommendation report.

[0047] Specifically, the prior probability of the target candidate literature set is calculated based on a multidimensional evaluation system that comprehensively assesses each literature from multiple perspectives, including evidence quality, clinical relevance, timeliness, and impact. The construction of the multidimensional evaluation vector first extracts key quality indicators from each literature, including evidence quality dimensions such as study design type, sample size, statistical rigor, journal impact factor, and peer review quality. Evidence quality is assessed using the criteria of evidence-based medicine, with randomized controlled trials and systematic reviews receiving the highest quality scores, cohort studies and case-control studies receiving intermediate scores, and case reports and expert opinions receiving lower scores. The clinical relevance dimension is calculated by analyzing the degree of match between the literature content and the user query, taking into account factors such as disease relevance, symptom matching, treatment applicability, and patient population similarity. The timeliness dimension assesses the guiding value of the publication date for current clinical practice. Medical knowledge is rapidly updated, and more recently published literature has greater timeliness. The algorithm uses a time decay function to weight older publications. The impact dimension measures the impact of a document in academia and clinical settings through metrics such as citations, downloads, and social media dissemination. The evaluation values ​​of these dimensions are standardized and combined into a multidimensional evaluation vector, with each component of the vector representing the performance level of the document in the corresponding dimension. The Softmax normalization function converts the multidimensional evaluation vector into a probability distribution. This function ensures that the sum of the prior probabilities of all documents is equal to 1 through exponential transformation and normalization operations, while maintaining the relative size relationship of the probability values. The temperature parameter of the Softmax function controls the smoothness of the probability distribution. Higher temperature values ​​produce smoother probability distributions, while lower temperature values ​​produce sharper distributions. The temperature parameter is adjusted based on the query type and application scenario. The prior probability distribution reflects the degree to which each document is selected and valued. Documents with high prior probabilities receive greater weight and attention in the subsequent knowledge fusion process.

[0048] Specifically, weighted knowledge fusion employs a feature integration technique guided by prior probabilities. This technique weights and combines the knowledge content in the target candidate document set according to prior probabilities to form a unified knowledge representation. The weighted fusion process first performs deep feature extraction on each document. A pretrained medical BERT model is used to encode core content, such as the document's title, abstract, and key conclusions, to generate high-dimensional semantic feature vectors. These feature vectors capture key information, including the document's primary medical concepts, research findings, and clinical significance. The prior probability weighting operation multiplies each document's feature vector by its corresponding prior probability value. The feature contribution of high-probability documents is amplified, while that of low-probability documents is minimized. This weighting mechanism ensures that high-quality, highly relevant documents dominate the fusion results. The feature concatenation operation concatenates all weighted document feature vectors in a predetermined order to form a comprehensive feature representation encompassing all candidate document information. The concatenation process utilizes an attention mechanism to address interactions between features. Attention weights are dynamically calculated based on feature importance and relevance, with more attention allocated to highly relevant feature segments. The fusion algorithm also incorporates a redundancy elimination mechanism, identifying and merging semantically similar feature fragments through similarity calculations to prevent duplication of information from interfering with the fusion results. The weighted fusion knowledge representation is a high-dimensional, dense vector that encodes the comprehensive knowledge content of the candidate document set, preserving the unique contributions of each document while highlighting the core ideas and important findings of high-quality documents.

[0049] Specifically, recommendation content generation utilizes a sequence-to-sequence generative model. This model, based on a Transformer-based decoder, converts the weighted fusion knowledge representation into human-readable structured recommendation text. The decoder employs an autoregressive generative approach to generate the vocabulary and sentences for the recommendation report, guided and constrained by the weighted fusion knowledge representation. The generative model was specifically trained on the task of generating abstracts and reports from medical scientific literature, learning the language patterns, terminology conventions, and report structure standards of the medical field. The structured recommendation text follows the standard medical report format, including standard sections such as Background Introduction, Main Findings, Clinical Significance, Treatment Recommendations, and Precautions. The Background Introduction summarizes basic information about the relevant disease and current research status. The Main Findings section summarizes key research results and clinical evidence from the candidate literature. The Clinical Significance section explains the guiding value of the research findings for clinical practice. The Treatment Recommendations section provides specific diagnostic and treatment recommendations based on evidence-based evidence. During the generation process, the decoder dynamically focuses on different parts of the weighted fusion knowledge representation using an attention mechanism to ensure that the generated content remains consistent and accurate with the source literature. The model also incorporates medical knowledge constraints to avoid generating content that conflicts with medical common sense. The preliminary recommendations are highly readable and professional, and can provide medical staff with clear and accurate literature reviews and clinical guidance.

[0050] Specifically, citation annotation utilizes automated citation management and annotation technology to ensure that every point and conclusion in the recommendation report is clearly supported by a source. The citation annotation process first establishes a mapping between the recommended content and the source literature. Using text similarity calculations and semantic matching techniques, the algorithm identifies the corresponding source literature fragment for each statement in the recommendation report. An algorithm analyzes the semantic structure of the recommended content and identifies key statements requiring citation support, such as research data, clinical findings, treatment efficacy, and side effect information. For each statement requiring citation, the algorithm searches the candidate literature collection for the most relevant supporting evidence, taking into account factors such as semantic similarity, content consistency, and strength of evidence. Citation tags are added to the recommendation content using standard academic citation format, using either numeric superscripts or parenthetical citations. Citation tags not only identify the source of information but also provide a level of confidence in the evidence. Citations of high-quality research are distinguished by special tags. The citation list is organized according to standard bibliographic format and includes complete bibliographic information, including author information, article title, journal name, year of publication, volume, issue, and page numbers. The citation list also provides a brief summary and key findings of each article, helping readers quickly understand the main content of the cited article. The algorithm also checks the completeness and accuracy of citations, ensuring that each citation tag has a corresponding literature entry and that the information in each literature entry is accurate. The resulting medical and scientific literature recommendation report features a complete citation system, ensuring the academic rigor of the content while providing readers with literature clues for further research. The recommendation report is presented in a standardized medical document format, with a clear chapter structure, standard citation format, and a complete reference list, providing a reliable literature review and evidence-based support for clinical decision-making and academic research.

[0051] In this embodiment, by performing multi-layer semantic annotation on medical scientific literature and constructing a symptom-disease dynamic association map, a medical scientific literature knowledge base containing semantic fingerprints is established; medical context analysis and multimodal feature extraction are performed on user queries, and a unified retrieval vector is generated by combining Boolean nested analysis and semantic alignment processing; based on the unified retrieval vector, evidence hierarchical retrieval and clinical scenario matching are performed to obtain a preliminary candidate document set, and the target candidate document set is optimized through fine-grained semantic recalculation; the retrieval prior probability is calculated according to the evaluation dimension, and the candidate documents are subjected to knowledge weighted fusion to generate a medical scientific literature recommendation report. The present invention deeply understands the association relationship between medical terms through the symptom-disease dynamic association map, ensures that the results match the patient's clinical characteristics through evidence hierarchical retrieval and clinical scenario matching screening, and improves the accuracy of document retrieval.

[0052] The above describes the scientific literature intelligent retrieval method based on generative artificial intelligence in the embodiment of the present invention. The following describes the scientific literature intelligent retrieval device based on generative artificial intelligence in the embodiment of the present invention. The scientific literature intelligent retrieval device based on generative artificial intelligence is described in detail. Figure 2 In one embodiment of the present invention, an intelligent retrieval device for scientific and technological literature based on generative artificial intelligence includes: Semantic annotation module 201 is used to perform multi-layer semantic annotation processing on medical scientific literature, construct a symptom-disease dynamic association map based on the annotated medical terms, and perform semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints; The query parsing module 202 is configured to perform medical context parsing and multimodal feature extraction on the user query using a dual encoder architecture, perform Boolean nested parsing on the logical relationships in the user query, and perform standardized conversion and semantic alignment using the symptom-disease dynamic association map and semantic fingerprints to obtain a unified retrieval vector. Intelligent retrieval module 203, configured to perform evidence hierarchical retrieval and clinical scenario matching respectively based on the unified retrieval vector and the medical scientific literature knowledge base to obtain a preliminary candidate document set, and perform fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set; The fusion recommendation module 204 is used to calculate the retrieval prior probability according to the evaluation dimension of the target candidate document set and perform knowledge weighted fusion on the candidate document set according to the retrieval prior probability to obtain a medical science and technology document recommendation report.

[0053] In an embodiment of the present invention, the scientific and technological literature intelligent retrieval device based on generative artificial intelligence runs the above-mentioned scientific and technological literature intelligent retrieval method based on generative artificial intelligence. The scientific and technological literature intelligent retrieval device based on generative artificial intelligence establishes a medical and technological literature knowledge base containing semantic fingerprints by performing multi-layer semantic annotation on medical and technological literature and constructing a symptom-disease dynamic association map; performs medical context analysis and multimodal feature extraction on user queries, combines Boolean operation nested analysis and semantic alignment processing to generate a unified retrieval vector; performs evidence hierarchical retrieval and clinical scenario matching based on the unified retrieval vector to obtain a preliminary candidate document set, and optimizes it to a target candidate document set through fine-grained semantic recalculation; calculates the retrieval prior probability according to the evaluation dimension, performs knowledge weighted fusion on the candidate documents, and generates a medical and technological literature recommendation report. The present invention deeply understands the association relationship between medical terms through the symptom-disease dynamic association map, ensures that the results match the patient's clinical characteristics through evidence hierarchical retrieval and clinical scenario matching screening, and improves the accuracy of document retrieval.

[0054] above Figure 2The intelligent retrieval device for scientific and technological literature based on generative artificial intelligence in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The intelligent retrieval device for scientific and technological literature based on generative artificial intelligence in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0055] Figure 3 This is a schematic diagram of the structure of a generative AI-based intelligent retrieval device for scientific and technological literature, provided by an embodiment of the present invention. This generative AI-based intelligent retrieval device for scientific and technological literature 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage medium 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions and operations within the generative AI-based intelligent retrieval device for scientific and technological literature 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and the generative AI-based intelligent retrieval device for scientific and technological literature 300 may execute the series of instructions and operations stored in the storage medium 330 to implement the steps of the generative AI-based intelligent retrieval method for scientific and technological literature.

[0056] The scientific literature intelligent retrieval device 300 based on generative artificial intelligence may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the scientific and technological literature intelligent retrieval device based on generative artificial intelligence shown does not constitute a limitation on the scientific and technological literature intelligent retrieval device based on generative artificial intelligence provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0057] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence.

[0058] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0059] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0060] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence, characterized in that: The scientific and technological literature intelligent retrieval method based on generative artificial intelligence includes: Perform multi-layer semantic annotation on medical scientific literature, build a symptom-disease dynamic association map based on the annotated medical terms, and perform semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints; The user query is parsed in medical context and multimodal features are extracted through a dual encoder architecture. The logical relationship in the user query is parsed through Boolean nested operations and standardized and semantically aligned using the symptom-disease dynamic association map and semantic fingerprints to obtain a unified retrieval vector. Performing evidence hierarchical retrieval and clinical scenario matching on the unified retrieval vector and the medical science and technology literature knowledge base to obtain a preliminary candidate document set, and performing fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set; The retrieval prior probability is calculated according to the evaluation dimension of the target candidate document set, and the candidate document set is subjected to knowledge weighted fusion according to the retrieval prior probability to obtain a medical science and technology document recommendation report.

2. The method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence according to claim 1, characterized in that: The multi-layer semantic annotation processing of medical scientific and technological literature, the construction of a symptom-disease dynamic association map based on the annotated medical terms, and the semantic mapping of the medical terms to obtain a medical scientific and technological literature knowledge base containing semantic fingerprints include: Identify medical terminology in medical scientific literature and use the medical ontology knowledge base to annotate medical terms to obtain medical term annotation results, where the medical terms include disease names, symptom descriptions, drug names, and treatment plans; calculating association strength weights based on co-occurrence patterns, causal relationship descriptions, and time series information of symptom terms and disease terms in the medical term annotation results; The graph structure is constructed with symptom terms and disease terms as graph nodes and association strength weights as edge weights to obtain a symptom-disease dynamic association graph; Performing semantic mapping processing on the medical term annotation results, and obtaining semantic mapping results by establishing a three-layer correspondence relationship between a professional term layer, a clinical description layer, and a patient expression layer; According to the symptom-disease dynamic association map and semantic mapping results, feature extraction is performed on medical scientific literature, and the main disease categories, core symptom groups, treatment plan types and patient population characteristics are extracted to generate semantic fingerprints, thereby obtaining a medical scientific literature knowledge base containing semantic fingerprints.

3. The method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence according to claim 1, characterized in that: The dual encoder architecture is used to perform medical context analysis and multimodal feature extraction on user queries, and the logical relationships in the user queries are subjected to Boolean nested analysis. The symptom-disease dynamic association graph and semantic fingerprint are used for standardization conversion and semantic alignment to obtain a unified retrieval vector including: The ViT-B / 16 encoder in the dual encoder architecture extracts patch-level fine-grained features from the image input in the user query, and the BERT encoder in the dual encoder architecture extracts token-level semantic features from the text query in the user query to obtain multimodal query features. Determining the query type of the user query based on the multimodal query features to obtain query context features, and performing Boolean nested parsing processing on the logical relationships in the user query to obtain a structured query logical representation; The symptom-disease dynamic association graph is used to perform standardized conversion on the patient symptom description queried by the user, and non-standard symptom expressions are converted into standard medical terms to obtain standardized symptom features; According to the semantic fingerprint, semantic alignment processing is performed on multimodal query features, standardized symptom features and structured query logic representation, and query context features are integrated to generate a unified retrieval vector.

4. The method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence according to claim 1, characterized in that: The unified search vector and the medical science and technology literature knowledge base are used to perform evidence hierarchical retrieval and clinical scenario matching respectively to obtain a preliminary candidate document set, and the preliminary candidate document set is subjected to fine-grained semantic recalculation to obtain a target candidate document set including: Performing an evidence-level search based on the unified search vector and the medical science and technology literature knowledge base, and obtaining candidate documents of evidence level by matching query complexity with document evidence level; Perform clinical scenario matching with the medical science and technology literature knowledge base based on the unified search vector, and obtain candidate clinical scenario literature by matching patient characteristics with literature clinical scenarios; Merge the evidence classification candidate documents and clinical scenario candidate documents, remove duplicate documents and sort them by search score to obtain a preliminary candidate document set; The ColBERTv2 model is used to calculate the semantic matching score between each candidate document in the preliminary candidate document set and the unified search vector. The top M documents are selected by score ranking to obtain the roughly ranked candidate documents. By extracting the patch-level and token-level features of the coarse-ranked candidate documents, calculating the maximum similarity matching score between the features, re-ranking and screening the Top-K documents, the target candidate document set is obtained.

5. The method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence according to claim 4, characterized in that: The evidence graded search is performed based on the unified search vector and the medical science and technology literature knowledge base, and the candidate documents for evidence grade are obtained by matching the query complexity with the document evidence grade, including: The query complexity score is obtained by calculating the semantic depth index, symptom combination complexity index and diagnostic reasoning level index of the unified retrieval vector; Based on the query complexity score, the literature in the medical science and technology literature knowledge base is pre-labeled according to the evidence level of randomized controlled trials, systematic reviews, cohort studies, and case reports to obtain a graded labeled literature library; According to the preset mapping relationship between complexity threshold and evidence level, the query complexity score is matched with the documents of each level in the graded annotated document library, and the selection priority score of each document is calculated to obtain a priority score list; Literature with a matching degree exceeding the preset threshold is screened according to the priority score list, and then sorted according to the citation frequency of the literature to obtain candidate literature for evidence classification.

6. The method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence according to claim 4, characterized in that: The clinical scenario matching is performed based on the unified search vector and the medical science and technology literature knowledge base, and the clinical scenario candidate documents obtained by matching patient characteristics with the clinical scenarios of the documents include: The age range, gender information, main symptoms, medical history, and medication status are extracted from the unified retrieval vector using a named entity recognition algorithm to obtain a structured patient feature vector. By analyzing the research subject description, inclusion and exclusion criteria, and clinical trial design of each article in the medical science and technology literature knowledge base, a clinical scenario feature matrix including applicable population, disease stage, and treatment plan is constructed for the article; By calculating the cosine similarity between the structured patient feature vector and the literature clinical scenario feature matrix, a list of patient-literature similarity scores was obtained; The clinical applicability credibility of each document is calculated based on the similarity score list and combined with the Bayesian confidence model. The documents with clinical applicability credibility exceeding the preset threshold are screened to obtain candidate clinical scenario documents.

7. The method for intelligent retrieval of scientific and technological literature based on generative artificial intelligence according to claim 1, characterized in that: The calculation of the retrieval prior probability based on the evaluation dimension of the target candidate document set and the knowledge weighted fusion of the candidate document set based on the retrieval prior probability to obtain a medical science and technology literature recommendation report includes: The prior probability of each document in the target candidate document set is calculated by multi-dimensional evaluation vector and softmax normalization of the evaluation dimension to obtain the prior probability distribution; Performing knowledge weighted fusion on the target candidate document set according to the prior probability distribution, and obtaining a weighted fusion knowledge representation through prior probability weighting and feature splicing operations; Generate recommended content based on the weighted fusion knowledge representation, convert it into structured recommendation text through a decoder, and obtain preliminary recommended content; The preliminary recommended content is annotated with literature citations, and citation marks and citation lists are added according to the source of the content to obtain a medical science and technology literature recommendation report.

8. A scientific literature intelligent retrieval device based on generative artificial intelligence, characterized in that: The scientific and technological literature intelligent retrieval device based on generative artificial intelligence includes: A semantic annotation module is used to perform multi-layer semantic annotation processing on medical scientific literature, construct a symptom-disease dynamic association map based on the annotated medical terms, and perform semantic mapping on the medical terms to obtain a medical scientific literature knowledge base containing semantic fingerprints; A query parsing module is used to perform medical context analysis and multimodal feature extraction on user queries using a dual encoder architecture, perform Boolean nested parsing on logical relationships in user queries, and perform standardized conversion and semantic alignment using the symptom-disease dynamic association graph and semantic fingerprints to obtain a unified retrieval vector; an intelligent retrieval module, configured to perform evidence hierarchical retrieval and clinical scenario matching respectively based on the unified retrieval vector and the medical science and technology literature knowledge base to obtain a preliminary candidate document set, and perform fine-grained semantic recalculation on the preliminary candidate document set to obtain a target candidate document set; The fusion recommendation module is used to calculate the retrieval prior probability based on the evaluation dimension of the target candidate document set and perform knowledge weighted fusion on the candidate document set based on the retrieval prior probability to obtain a medical science and technology document recommendation report.

9. A scientific literature intelligent retrieval device based on generative artificial intelligence, characterized in that: The scientific and technological literature intelligent retrieval device based on generative artificial intelligence includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory so that the scientific and technological literature intelligent retrieval device based on generative artificial intelligence performs the steps of the scientific and technological literature intelligent retrieval method based on generative artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the steps of the intelligent retrieval method for scientific and technological literature based on generative artificial intelligence as described in any one of claims 1 to 7 are implemented.

Images