A medical question and answer method and system based on cooperative dual-source retrieval enhancement generation

Abstract

The application provides a medical question and answer method and system based on cooperative dual-source retrieval enhancement generation, which comprises the following steps: according to an input original medical question and candidate items, performing task self-adaptive query rewriting through a large language model to generate differential diagnosis keywords, entity enhanced queries and hypothetical medical abstracts; for the differential diagnosis keywords, performing an iterative network retrieval with a reflection mechanism to obtain network candidate evidence; for the entity enhanced queries and the hypothetical medical abstracts, performing a hybrid retrieval in a local medical knowledge base to obtain local candidate evidence; aggregating and deduplicating the network candidate evidence and the local candidate evidence, obtaining the aggregated heterogeneous evidence, performing deep semantic correlation scoring through a cross-encoder model, screening out a target evidence set, and splicing the original medical question as context input into a generative large language model to generate a final medical answer. The application improves the accuracy, timeliness and explainability of the medical question and answer system.

Description

Technical Field

[0001] This invention relates to the field of medical question-and-answer service technology, and in particular to a medical question-and-answer method and system based on collaborative dual-source retrieval enhancement generation. Background Technology

[0002] Medical question answering and clinical decision support are critical fields requiring extremely high accuracy and logical rigor. With the rapid development of artificial intelligence technology, large-scale language models, with their superior natural language understanding and generation capabilities, have demonstrated enormous application potential in building medical knowledge question answering systems, assisting medical teaching, and simulating clinical reasoning. Through human-computer interaction, this technology aims to provide medical students and professionals with a convenient way to acquire knowledge, which has significant application value in improving the efficiency of medical talent training, reducing learning costs, and assisting in preliminary diagnosis.

[0003] However, current mainstream general-purpose large language model-based direct question-answering solutions have significant technical limitations in high-reliability vertical scenarios such as medicine. These solutions primarily rely on parameterized memories learned from general corpora during model pre-training to directly generate answers. Limited by the insufficient depth of coverage of general corpora in the medical field, the models often exhibit a lack of factual accuracy when answering specialized medical questions, failing to meet the stringent standards of precision and professionalism in the medical field. More importantly, large models inherently follow a probability-driven generation mechanism, which, in the absence of reliable external knowledge constraints, is highly susceptible to "illusion" phenomena—generated content that, while linguistically fluent, deviates significantly from the truth in terms of logic and facts. This inherent unreliability makes it difficult for directly applied large models to explain their reasoning, thus posing a potential risk of misleading in medical question answering and diagnostic assistance.

[0004] To alleviate the limitations of general-purpose large-scale models in specialized fields, existing technologies have proposed a retrieval-enhanced generative architecture. This architecture aims to construct a proprietary, closed knowledge base by integrating external data such as medical textbooks, clinical guidelines, and professional literature. Its basic implementation process typically involves: first, preprocessing and indexing domain-specific document data; when the system receives a medical query from a user, using a retrieval algorithm to match and retrieve semantically relevant content fragments from the knowledge base; then, inputting the retrieved fragments as contextual information along with the original query into a large-scale language model, guiding the model to generate an answer based on these references. While this method improves the standardization and credibility of the answer to some extent by introducing external reference information, its core drawback lies in its heavy reliance on a static, local knowledge base. Once the knowledge base is built, the information it contains is locked at the point of construction, and the system cannot automatically acquire new knowledge generated subsequently. In the medical field, new drug research, clinical trial data, and treatment standards are updated extremely frequently. This static and closed nature leads to significant knowledge lag in the system, making it unable to effectively handle complex question-and-answer scenarios involving cutting-edge medical advancements.

[0005] To overcome the timeliness bottleneck of static knowledge bases, another existing technology attempts to introduce open-domain web retrieval and adopt iterative retrieval strategies to obtain dynamic information. This type of method typically connects to internet search engines, uses large models to generate query terms, and conducts multiple rounds of "search-feedback-research" interactions to gradually acquire deeper information. However, directly applying this general method to medical question-answering scenarios faces the dual challenges of "semantic mismatch" and "noise accumulation." On the one hand, medical questions or clinical problems often contain lengthy case descriptions, complex professional terminology, and distracting options. General query generation mechanisms lack the ability to adapt to tasks in vertical domains, making it difficult to translate complex professional intents into accurate keywords suitable for general search engines, resulting in low relevance of search results. On the other hand, open web data sources are filled with unstructured noise, advertisements, and low-quality information. Existing iterative retrieval methods often lack effective information verification and reflection mechanisms, easily introducing and amplifying erroneous information during multiple rounds of interaction, leading to "retrieval drift," which severely damages the logical rigor and factual reliability of the final generated answer.

[0006] In summary, existing medical question-and-answer systems have the following shortcomings: (1) Lagging knowledge timeliness and high maintenance costs: Most existing medical question answering systems (such as the aforementioned solutions based on static knowledge bases or knowledge graphs) rely on pre-built closed data sources. Medical knowledge is updated rapidly, and static libraries face the risk of becoming outdated once built. Furthermore, frequent full updates of the index are extremely costly, making it difficult for the system to answer questions involving the latest medical advancements.

[0007] (2) Semantic mismatch of general retrieval mechanisms in medical scenarios: Although some technologies attempt to introduce web search, existing general retrieval or simple iterative retrieval methods lack task-adaptive optimization for the medical field. Medical problems often contain complex case descriptions and professional terminology, and directly using raw queries often makes it difficult to match accurate discriminative evidence in the open network, resulting in low relevance of search results.

[0008] (3) Lack of unified standards for heterogeneous evidence fusion: Existing solutions use simple concatenation or linear weighting when processing two types of heterogeneous data, "local documents" and "online text". Due to the lack of a reordering mechanism based on deep semantic interaction, the system cannot accurately identify high-quality evidence with real medical interpretability, resulting in the answer being interfered with by low-quality information.

[0009] Therefore, it is evident that overcoming the timeliness barrier of static knowledge bases, resolving semantic mismatch in medical scenarios for general retrieval, and establishing unified evaluation standards for heterogeneous evidence have become urgent technical problems to be solved. Summary of the Invention

[0010] In view of this, the present invention provides a medical question-answering method and system based on collaborative dual-source retrieval enhancement generation, so as to at least solve the above-mentioned problems.

[0011] This invention provides a medical question-answering method based on collaborative dual-source retrieval enhancement, comprising: based on the input original medical question and candidate options, performing task adaptive query rewriting through a large language model to generate differential diagnosis keywords adapted to web retrieval, entity enhancement queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval; performing iterative web retrieval with a reflective mechanism for the differential diagnosis keywords to obtain web candidate evidence; performing hybrid retrieval in a local medical knowledge base for the entity enhancement queries and the hypothetical medical summaries to obtain local candidate evidence; aggregating and deduplicating the web candidate evidence and the local candidate evidence to obtain aggregated heterogeneous evidence; performing deep semantic relevance scoring on the aggregated heterogeneous evidence through a cross-encoder model to select a target evidence set; concatenating the original medical question and the target evidence set as context input into a generative large language model to generate a final medical answer containing evidence citation tags.

[0012] Optionally, the step of performing adaptive query rewriting through a large language model to generate differential diagnosis keywords adapted to web retrieval, entity augmentation queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval includes: constructing a first prompt instruction to guide the large language model to analyze the original medical question and candidate options, extract core medical features or differential points used to distinguish the correctness of each option, and generate the differential diagnosis keywords; constructing a second prompt instruction to guide the large language model to identify core medical entities from the original medical question and candidate options, and perform full name completion and synonym expansion to generate the entity augmentation query; and constructing a third prompt instruction to guide the large language model to generate a declarative summary simulating the style of a medical textbook or clinical guideline based on the original medical question and candidate options, thereby obtaining the hypothetical medical summary.

[0013] Optionally, the step of performing an iterative web search with a reflective mechanism to obtain candidate evidence for the differential diagnosis keywords includes: search initialization: submitting the differential diagnosis keywords to an internet search engine and setting a search pointer; pagination execution and retrieval: retrieving the original web page data returned by the internet search engine in batches according to a set window size; cleaning and refining: cleaning the original web page data to obtain cleaned web page text; extracting medical facts, clinical data, or definitions related to the current differential diagnosis keywords from the cleaned web page text using a large language model and storing them in a temporary evidence pool for information accumulation; reflection based on information saturation: inputting the accumulated information in the temporary evidence pool and the current differential diagnosis keywords into the large language model to determine whether the medical information of the current differential diagnosis keywords has been sufficiently covered; if the preset maximum search depth threshold has not been reached, moving the search pointer and returning to the pagination execution and retrieval steps; if the preset maximum search depth threshold has been reached, stopping the search; and output: aggregating all the accumulated information in the temporary evidence pool obtained after stopping the search into the candidate evidence for the network.

[0014] Optionally, the step of performing a hybrid retrieval in a local medical knowledge base to obtain local candidate evidence for the entity augmentation query and the hypothetical medical summary includes: for the entity augmentation query, performing a keyword retrieval based on the BM25 algorithm using a pre-built inverted index to obtain a first candidate document list; for the hypothetical medical summary, mapping it to a high-dimensional dense query vector using a pre-trained vector encoding model, and performing an approximate nearest neighbor search by calculating cosine similarity to obtain a second candidate document list; and fusing and reordering the document fragments in the first and second candidate document lists using an inverse sorting fusion algorithm to generate the local candidate evidence.

[0015] Optionally, the step of fusing and reordering document fragments in the first candidate document list and the second candidate document list using a reverse sorting fusion algorithm to generate the local candidate evidence includes: calculating the fusion score of each document fragment in the first candidate document list and the second candidate document list; sorting all document fragments in descending order according to the fusion score to generate a fused candidate document list; and selecting the Top-M document fragments with the highest fusion scores from the fused candidate document list as the local candidate evidence.

[0016] Optionally, the fusion score for each document fragment can be calculated using the following formula:

[0017] in, It is a smoothing constant. For document fragments In the list The ranking position in the system.

[0018] Optionally, the step of performing deep semantic relevance scoring on the aggregated heterogeneous evidence using a cross-encoder model to select a target evidence set includes: concatenating each piece of evidence in the aggregated heterogeneous evidence with the original medical question to construct a text pair; inputting the text pair into a pre-trained cross-encoder model, and calculating the relevance confidence score of the evidence relative to the original medical question through the self-attention mechanism within the cross-encoder model; sorting all evidence in descending order based on the relevance confidence score; and selecting the top-K high-confidence evidence from the descendingly sorted evidence to constitute the target evidence set.

[0019] In another aspect, the present invention provides a medical question-answering system based on collaborative dual-source retrieval enhancement generation, comprising: a task-adaptive query rewriting module, used to perform task-adaptive query rewriting through a large language model based on the input original medical question and candidate options, generating differential diagnostic keywords adapted to web retrieval, entity-enhanced queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval; a collaborative dual-source retrieval module, used to perform an iterative web retrieval with a reflective mechanism for the differential diagnostic keywords to obtain web candidate evidence; and to perform a hybrid retrieval in a local medical knowledge base for the entity-enhanced queries and the hypothetical medical summaries to obtain local candidate evidence; a heterogeneous evidence reordering module, used to aggregate and deduplicate the web candidate evidence and the local candidate evidence to obtain aggregated heterogeneous evidence; and to perform deep semantic relevance scoring on the aggregated heterogeneous evidence through a cross-encoder model to select a target evidence set; and a traceable answer generation module, used to concatenate the original medical question and the target evidence set as context input into a generative large language model to generate a final medical answer containing evidence citation tags.

[0020] In another aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of a medical question-answering method based on collaborative dual-source retrieval enhancement as described in any of the preceding claims. In another aspect, the present invention provides a computer storage medium storing a computer program that, when executed by a processor, implements the steps of a medical question-answering method based on collaborative dual-source retrieval enhancement as described in any of the preceding claims.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) The collaborative dual-source retrieval architecture proposed in this invention realizes the complementary integration of local authoritative medical knowledge and dynamic Internet information, effectively solves the problem of information lag in a single local knowledge base and the problem of insufficient professionalism in a single network search, and significantly improves the knowledge coverage and timeliness of the medical question-and-answer system.

[0022] (2) The task-adaptive query rewriting mechanism proposed in this invention can generate optimized query representations for the characteristics of different retrieval paths, overcoming the shortcomings of the traditional single query method with low recall rate of related documents on heterogeneous data sources, and greatly improving the retrieval accuracy of medical term variations and cutting-edge medical knowledge.

[0023] (3) The iterative network retrieval strategy based on “retrieval-cleaning-refinement-reflection” proposed in this invention controls the retrieval depth by dynamically verifying the completeness of evidence, avoiding invalid web page browsing and wasting computing resources, and improving the system’s operating efficiency.

[0024] (4) The heterogeneous evidence reordering scheme based on cross encoder proposed in this invention eliminates the differences in scoring criteria between local documents and online texts, effectively removes noise from a large number of search results, and ensures high relevance and low interference of the input context of the large language model.

[0025] (5) The traceable answer generation method proposed in this invention forces the model to mark the source of the medical facts cited, making the generated medical advice verifiable and effectively enhancing the interpretability and user trust of the system in clinical auxiliary scenarios. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention. In the accompanying drawings: Figure 1 This is a flowchart of the steps of the method of the present invention.

[0027] Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0029] This invention provides a medical question-answering method and system based on collaborative dual-source retrieval enhancement generation, aiming to solve the problems of knowledge lag, semantic mismatch and large model illusion in existing medical question-answering systems.

[0030] See Figure 1 The present invention provides a medical question-answering method based on collaborative dual-source retrieval enhancement generation, which mainly includes: S1. Based on the input original medical question and candidate options, the large language model performs adaptive query rewriting to generate differential diagnosis keywords adapted to web retrieval, entity augmentation queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval. S2. For the aforementioned diagnostic keywords, perform an iterative network search with a reflective mechanism to obtain network candidate evidence; S3. For the entity enhancement query and the hypothetical medical summary, perform a hybrid search in the local medical knowledge base to obtain local candidate evidence; S4. Aggregate and deduplicate the network candidate evidence and the local candidate evidence to obtain aggregated heterogeneous evidence; S5. The aggregated heterogeneous evidence is scored with deep semantic relevance using a cross-encoder model to select the target evidence set. S6. The original medical question and the target evidence set are concatenated and used as context input into the generative large language model to generate a final medical solution containing evidence citation tags.

[0031] It should be understood that the final medical solution is generated based on explicit traceability constraint instructions.

[0032] Through the above method, the present invention achieves the synergistic integration of local normative knowledge and dynamic network information, effectively alleviating the problem of knowledge lag while eliminating retrieval noise, and improving the accuracy, timeliness and interpretability of medical question-and-answer generated based on retrieval enhancement.

[0033] Optionally, the step of performing adaptive query rewriting through a large language model to generate differential diagnosis keywords adapted to web retrieval, entity augmentation queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval includes: constructing a first prompt instruction to guide the large language model to analyze the original medical question and candidate options, extract core medical features or differential points used to distinguish the correctness of each option, and generate the differential diagnosis keywords; constructing a second prompt instruction to guide the large language model to identify core medical entities from the original medical question and candidate options, and perform full name completion and synonym expansion to generate the entity augmentation query; and constructing a third prompt instruction to guide the large language model to generate a declarative summary simulating the style of a medical textbook or clinical guideline based on the original medical question and candidate options, thereby obtaining the hypothetical medical summary.

[0034] Optionally, the step of performing an iterative web search with a reflective mechanism to obtain candidate evidence for the differential diagnosis keywords includes: search initialization: submitting the differential diagnosis keywords to an internet search engine and setting a search pointer; pagination execution and retrieval: retrieving the original web page data returned by the internet search engine in batches according to a set window size; cleaning and refining: cleaning the original web page data to obtain cleaned web page text; extracting medical facts, clinical data, or definitions related to the current differential diagnosis keywords from the cleaned web page text using a large language model and storing them in a temporary evidence pool for information accumulation; reflection based on information saturation: inputting the accumulated information in the temporary evidence pool and the current differential diagnosis keywords into the large language model to determine whether the medical information of the current differential diagnosis keywords has been sufficiently covered; if the preset maximum search depth threshold has not been reached, moving the search pointer and returning to the pagination execution and retrieval steps; if the preset maximum search depth threshold has been reached, stopping the search; and output: aggregating all the accumulated information in the temporary evidence pool obtained after stopping the search into the candidate evidence for the network.

[0035] Optionally, the step of performing a hybrid retrieval in a local medical knowledge base to obtain local candidate evidence for the entity augmentation query and the hypothetical medical summary includes: for the entity augmentation query, performing a keyword retrieval based on the BM25 algorithm using a pre-built inverted index to obtain a first candidate document list; for the hypothetical medical summary, mapping it to a high-dimensional dense query vector using a pre-trained vector encoding model, and performing an approximate nearest neighbor search by calculating cosine similarity to obtain a second candidate document list; and fusing and reordering the document fragments in the first and second candidate document lists using an inverse sorting fusion algorithm to generate the local candidate evidence.

[0036] Optionally, the step of fusing and reordering document fragments in the first candidate document list and the second candidate document list using a reverse sorting fusion algorithm to generate the local candidate evidence includes: calculating the fusion score of each document fragment in the first candidate document list and the second candidate document list; sorting all document fragments in descending order according to the fusion score to generate a fused candidate document list; and selecting the Top-M document fragments with the highest fusion scores from the fused candidate document list as the local candidate evidence.

[0037] Optionally, the fusion score for each document fragment can be calculated using the following formula:

[0038] in, It is a smoothing constant. For document fragments In the list The ranking position in the system.

[0039] Optionally, the step of performing deep semantic relevance scoring on the aggregated heterogeneous evidence using a cross-encoder model to select a target evidence set includes: concatenating each piece of evidence in the aggregated heterogeneous evidence with the original medical question to construct a text pair; inputting the text pair into a pre-trained cross-encoder model, and calculating the relevance confidence score of the evidence relative to the original medical question through the self-attention mechanism within the cross-encoder model; sorting all evidence in descending order based on the relevance confidence score; and selecting the top-K high-confidence evidence from the descendingly sorted evidence to constitute the target evidence set.

[0040] Another aspect of the present invention provides a medical question-answering system based on collaborative dual-source retrieval enhancement generation, comprising: The task-adaptive query rewriting module is used to perform task-adaptive query rewriting based on the input original medical question and candidate options through a large language model, generating differential diagnosis keywords adapted to web retrieval, entity augmentation queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval. The collaborative dual-source retrieval module is used to perform an iterative network retrieval with a reflective mechanism for the differential diagnosis keywords to obtain network candidate evidence; and to perform a hybrid retrieval in the local medical knowledge base for the entity augmentation query and the hypothetical medical summary to obtain local candidate evidence. The heterogeneous evidence reordering module is used to aggregate and deduplicate the network candidate evidence and the local candidate evidence to obtain aggregated heterogeneous evidence; and to perform deep semantic relevance scoring on the aggregated heterogeneous evidence through a cross-encoder model to select the target evidence set. The traceable answer generation module is used to concatenate the original medical question with the target evidence set as context input into the generative large language model to generate a final medical answer containing evidence citation tags.

[0041] It should be understood that the system of this embodiment is used to implement the corresponding methods in the foregoing multiple method embodiments and has the beneficial effects of the corresponding method embodiments.

[0042] Specifically, the solution of the present invention is further described with reference to the following examples: See Figure 2 The system of this invention has a collaborative dual-source retrieval enhanced generation architecture, which mainly consists of four core layers: the task adaptive query rewriting layer, the collaborative dual-source retrieval layer, the heterogeneous evidence reordering layer, and the traceable answer generation layer.

[0043] (1) Task-Adaptive Query Rewriting Layer: As the input processing layer of the system, its configuration is used to receive the user's original medical questions (usually including a question description and candidate options). This layer has a built-in task-adaptive rewriting mechanism based on a large language model, which is used to convert unstructured original medical questions into three query representations adapted to different downstream retrieval paths: differential diagnosis keywords adapted to the network retrieval path, entity augmentation query adapted to the local sparse retrieval path, and hypothetical medical summary adapted to the local dense retrieval path.

[0044] (2) Collaborative Dual-Source Retrieval Layer: This layer is connected to the task adaptive query rewriting layer and is configured to perform local and network retrieval tasks in parallel to obtain medical evidence.

[0045] Local retrieval pathway: Based on the entity-enhanced query and hypothetical medical summary, sparse and dense searches are performed respectively in the pre-built local medical knowledge base, and the retrieval results are fused using the inverse ranking fusion (RRF) algorithm to output local candidate evidence. ).

[0046] Network retrieval pathway: Based on the aforementioned diagnostic keywords, an iterative retrieval with a reflective mechanism is performed by connecting to an internet search engine. This pathway includes a closed loop of retrieval, cleaning, refinement, and information saturation reflection, used to obtain and output candidate evidence from the network. ).

[0047] (3) Heterogeneous Evidence Reordering Layer: This layer connects the local retrieval pathway and the network retrieval pathway, and is configured to receive and aggregate the local candidate evidence and the network candidate evidence. This layer has a built-in cross-encoder model, which is used to perform deep semantic interaction calculations on the aggregated heterogeneous evidence and the original medical question, output the relevance score of each piece of evidence, and select the Top-K high-confidence evidence to form the target evidence set. ).

[0048] (4) Traceable Answer Generation Layer: This layer connects to the heterogeneous evidence reordering layer and is configured to concatenate the original medical question with the target evidence set in context. This layer utilizes a generative large language model and, based on explicit traceability constraints, generates a final medical answer containing evidence citation annotations. .

[0049] The specific implementation methods for each layer are described below: 1. Detailed implementation of the task-adaptive query rewriting layer The task-adaptive query rewriting layer first receives the original medical question input by the user. Then, adaptive query rewriting is performed for network retrieval pathways, local sparse retrieval pathways, and local dense retrieval pathways respectively. The original medical questions are usually unstructured text, containing specific questions and multiple candidate options.

[0050] For ease of explanation, this embodiment uses the following input example as an example, and all subsequent retrieval paths are rewritten based on the original query in this format: Question: Which of the following statements about the characteristics of bilirubin metabolism in the liver of newborns is incorrect? Candidates: A: Insufficient amount and activity of UDPGT B: Poor function in excreting conjugated bilirubin C: Y and Z proteins reach adult levels 5-15 days after birth. D: UDPGT reached adult level 1 week after birth. 1.1 Generate diagnostic keywords based on network retrieval pathways: To align with internet search engines' preference for short text matching and reduce noise from long texts, the system implements a first prompt instruction. This instruction aims to guide the model to analyze the differences between candidate options and extract core feature words for differential diagnosis, rather than simply asking a restatement question.

[0051] Example of the first prompt instruction template: "Role setting: Medical differential diagnosis expert. Task objective: Analyze the given clinical problem and candidates, and extract the core medical features or differential points used to distinguish the correctness of each option. Constraints: Ignore irrelevant background descriptions; output only 3-5 of the most specific keywords; keywords are separated by spaces. Input content:" ".

[0052] The system parses the output of the large language model into a set of keywords. Used for subsequent internet searches. Output results are as follows: "Neonatal bilirubin metabolism UDPGT activity Y protein Z protein excretion function".

[0053] 1.2 Keyword entity enhancement and normalization for local sparse retrieval pathways: To address the reliance of sparse retrieval algorithms such as BM25 (Best Mathcing 25) on precise keyword matching, and the issue of abbreviations and aliases in medical terminology, the system constructs a second prompt instruction. This instruction aims to guide the model to identify core entities and perform full name completion and synonym expansion.

[0054] Example of the second prompt instruction template: "Role setting: Medical terminology search optimization expert. Task objective: Extract core medical entities (such as diseases, indicators, drugs) from questions and options. Processing logic: Extract and identify highly specific entities (such as "UDPGT"); standardize and provide the full name (such as "urinary diphosphate glucuronide transferase"); expand by providing 2-3 clinically commonly used synonyms or variations. Output format: Output only the deduplicated keyword list, separated by commas, with a total number not exceeding 20. Input content:" ".

[0055] The system obtains entity-enhanced queries. Used for BM25 retrieval. Output results are as follows: "Neonatal liver, bilirubin metabolism, UDPGT, urinary bisphosphoglucuronyl transferase, glucuronyl transferase, Y protein, Z protein, carrier protein, bilirubin excretion, conjugated bilirubin."

[0056] 1.3 Generating Hypothetical Medical Abstracts for Local Dense Search Pathways: To address the semantic space mismatch between the distributional differences between medical queries (specific and descriptive) and knowledge base text (abstract and declarative) in vector retrieval, the system employs a hypothetical document embedding (HyDE) strategy to construct a third prompt instruction. This instruction aims to generate a declarative text containing standard medical knowledge, making it semantically closer to paragraphs in textbooks or guidelines.

[0057] Example of the third prompt instruction template: "Based on the following medical question, please generate a standard summary that may appear in medical textbooks or clinical guidelines. The summary should include the relevant pathological mechanisms and treatment principles that can answer the question, and even if the details are not completely accurate, the semantic direction should be correct." System for obtaining hypothetical medical summaries The input vector encoding model maps to high-dimensional semantic vectors, thereby achieving precise alignment with local knowledge base documents in the semantic space. The output is as follows: "The metabolism of bilirubin in the liver of newborns has its own unique characteristics. At birth, the levels of Y and Z proteins in liver cells are extremely low, resulting in insufficient bilirubin uptake. It usually takes 5 to 15 days after birth to reach adult levels. At the same time, the content and activity of urinary bisphosphoglucuronide transferase (UDPGT) in the liver of newborns are significantly low, usually starting to increase around 1 week after birth, and reaching adult levels by about 2 weeks. In addition, the function of the liver in excreting conjugated bilirubin is relatively poor in newborns, which can easily lead to temporary bilirubin accumulation." After completing the above three query rewriting methods, the system will generate Send to the network retrieval channel, and Send the data to the sparse search and dense search paths in the local search channel, respectively.

[0058] 2. Implementation plan for the collaborative dual-source retrieval layer 2.1 Construction and Preprocessing of Local Medical Knowledge Base To construct a local medical search source with high coverage and low noise, this embodiment implements the following rigorous data processing flow: (1) Data Source Acquisition. The system first collects authoritative medical literature in PDF format as the original data source. The authoritative medical literature includes, but is not limited to: standard medical textbooks designated by the Chinese National Medical Licensing Examination (MCMLE) (covering disciplines such as Internal Medicine, Surgery, Psychiatry, and Pharmacology), current clinical practice guidelines, and authoritative expert consensus documents.

[0059] (2) Format conversion and text extraction. PyMuPDFLoader, integrated with the LangChain framework, was used to parse the collected PDF medical documents page by page. The system reads the underlying streaming data of the document and accurately extracts elements such as plain text content, layout coordinates, heading levels, and table structure, and converts them into an editable TXT plain text sequence.

[0060] (3) Rule-based deep cleaning. Since the TXT text converted from PDF parsing often contains formatting noise, the system performs the following regular expression-based cleaning steps to eliminate semantic interference: a. Remove header, footer, and page number noise: Identify header text and consecutive page numbers at fixed positions on each page, match and remove them using regular expressions to prevent them from truncating the semantics of the main text.

[0061] b. Filter non-textual interference: Scan the text sequence and remove garbled characters caused by parsing failures, invisible control characters, and meaningless fragment lines with a length less than a preset threshold.

[0062] c. Citation and Reference Separation: Identify and remove the reference list at the end of the article and the superscript citations in the text, focusing on the description of medical facts themselves.

[0063] d. Cross-line text merging: To address the issue of forced line breaks within paragraphs caused by PDF formatting, the system detects the end character of each line. If the end of the line is not a period, question mark, or other punctuation mark, it is determined to be a line break within the paragraph. The system then appends the next line of text to the current line to restore the complete natural language paragraph structure.

[0064] (4) Adaptive Recursive Text Slicing Strategy Based on Hierarchical Structure: For medical textbooks and clinical guidelines with clearly defined chapter hierarchies, this embodiment employs a strategy combining delimiter-based recursive slicing with a sliding window to ensure semantic integrity after slicing. The specific steps are as follows: a. Recursive coarse segmentation based on delimiters: The system pre-defines a list of text delimiters with priorities, from highest to lowest: double newline character, single newline character, sentence-ending punctuation, and comma. The system prioritizes using high-priority delimiters to hierarchically decompose the text. If the length of the segmented fragment does not exceed the preset maximum threshold, the fragment's integrity is preserved; if the fragment length exceeds the threshold, the system recursively downgrades the delimiter to the next level, breaking the fragment down into finer-grained semantic blocks, until the length of all basic semantic blocks meets the threshold requirement.

[0065] b. Sliding window-based block merging and overlap settings: After completing the basic semantic block division, the system uses a sliding window mechanism to merge adjacent semantic blocks to fill them to a length close to the maximum threshold. Simultaneously, to prevent key contextual semantics from being severed at the segmentation boundaries, the system sets a fixed-length overlap area between adjacent final slices, ensuring that semantic information at the boundary can exist simultaneously in both the preceding and following slices.

[0066] 2.2 Detailed Implementation of Local Hybrid Retrieval Path The system performs sparse and dense searches in parallel on the local knowledge base and then merges the results. The specific process is as follows: (1) Perform sparse retrieval based on entity augmentation query Input reception: The system receives entity enhancement queries from the query rewriting layer. This query includes the full name of the medical entity and its synonym expansion.

[0067] BM25 Search Execution: The system utilizes a pre-built inverted index to perform keyword searches based on the BM25 algorithm. The system calculates the query terms. The relevance score of each document to the query is calculated by combining the word frequency in the document and the inverse document frequency in the entire corpus with the document length normalization factor.

[0068] Candidate set generation: The system sorts the documents from high to low according to their relevance scores and outputs a list of first candidate documents. This list contains Top-N (N=15) document fragments, focusing on accurately hitting specific disease names, drug names or clinical indicators.

[0069] (2) Perform intensive retrieval based on hypothetical summaries Input reception: The system receives hypothetical medical summaries from the query rewrite layer. ).

[0070] Vector Encoding: The system calls a pre-trained vector encoding model to process the hypothetical medical summary. The mapping is to a high-dimensional dense query vector.

[0071] Vector retrieval execution: The system performs an approximate nearest neighbor search on a pre-built vector index. The system calculates the cosine similarity between the query vector and all document vectors in the knowledge base.

[0072] Candidate set generation: The system sorts the documents from largest to smallest based on their similarity scores and outputs a second candidate document list. This list contains Top-N (N=15) document fragments, focusing on capturing medical principle descriptions that are semantically consistent but expressed differently.

[0073] (3) Perform Reverse Rank Merge (RRF) Differentiated fusion: Given the different numerical distributions of BM25 scores and cosine similarity, the system uses a reciprocal sorting fusion algorithm to uniformly sort the two candidate document lists.

[0074] Score Calculation: The system iterates through all documents in both lists and calculates the score for each document segment. Calculate its fusion score using the following formula. :

[0075] in, This is the smoothing constant (set to 50). For document fragments In the list The ranking position in the system.

[0076] Final output: The system reorders all documents based on their RRF scores and extracts the top-M (M=10) segments with the highest scores to form the final local candidate evidence set. ), output to the reordering layer.

[0077] 2.3 Detailed Implementation of the Network Iterative Retrieval Path The online retrieval pathway targets specific diagnostic keywords and achieves a balance between recall and computational resource consumption by adjusting the search depth. The specific implementation process is as follows: (1) Search initialization and pagination execution Input reception: The system receives a set of differential diagnostic keywords from the query rewriting layer. ).

[0078] Batch Retrieval Mechanism: The system employs a batch retrieval mechanism, submitting keywords to the Internet search engine API. To control retrieval depth, the system sets the current retrieval pointer P=0 (representing the starting ranking position of the search results) and adopts a batch retrieval strategy, where the window size for each retrieval is N (e.g., N=3 webpage links). In the first iteration, the system only retrieves the original webpage content of the Top-1 to Top-N results returned by the search engine.

[0079] (2) Cleaning and refining of unstructured web page data Noise cleaning: For the acquired raw HTML webpage data, the system uses a pre-defined set of regular expression rules to perform cleaning, removing HTML tags, navigation bars, pop-up ad text and script code, while retaining the main text of the webpage.

[0080] Evidence Refinement: In order to extract high-density medical information from lengthy web page text, the system inputs the cleaned text into a large language model and executes evidence refinement instructions.

[0081] Evidence Refinement Instruction Logic: "Read the following webpage fragment and extract keywords..." Directly relevant medical facts, clinical data, or definitions. Ignore irrelevant background information. Output 'None' if no valid information is provided. Output: The system collects a refined summary of the output of the large model and stores it in a temporary evidence pool.

[0082] (3) Reflection based on information saturation Search keyword information saturation determination: The system inputs the accumulated information in the temporary evidence pool along with the current search keyword into the large language model and executes the information saturation verification instruction to determine whether the amount of information obtained is sufficient to fully analyze the medical meaning of the keyword.

[0083] Information saturation verification instruction logic: "Based on the currently collected information about..." The evidence provided should be assessed to determine whether it adequately covers the core clinical features of the concept, such as etiology, symptoms, and indicators. If the information is saturated and there are no significant gaps, output [STOP]; if the information is insufficient or contains critical gaps, output [CONTINUE]. Network Iterative Retrieval Judgment: The system analyzes the output of the large model. If the output is [STOP] or the preset maximum retrieval depth threshold has been reached, it determines that the retrieval for that keyword has been saturated, stops acquiring subsequent web pages, and outputs the current evidence pool content; if the output is [CONTINUE], it determines that the current evidence is insufficient. The system keeps the search keywords unchanged and moves the search pointer down (…). Continue to retrieve subsequent ranked web pages (e.g., Top-N+1 to Top-2N results) and return to perform the "cleaning and refining" step.

[0084] (4) Output of network evidence set Aggregated Output: After the retrieval process for all keywords has terminated, the system aggregates the refined summaries from all temporary evidence pools to construct the final network candidate evidence set. And then it is sent to the heterogeneous evidence reordering layer.

[0085] 3. Detailed Implementation of the Sorting Layer In this embodiment, the heterogeneous evidence reordering layer is configured to perform unified deep semantic evaluation and filtering of multi-source evidence. The system first receives a local candidate evidence set from the local retrieval pathway ( ) and network candidate evidence sets from network retrieval pathways ( The system then merges these two sets of evidence from different sources into a unified initial candidate pool. Subsequently, the system analyzes each piece of evidence in the candidate pool... To link it with original medical problems The system concatenates the text pairs to construct text pairs conforming to the input specifications of the cross-encoder. It then calls a pre-trained cross-encoder model to perform inference on these text pairs, utilizing the model's internal self-attention mechanism to capture the deep semantic interaction between the query and the evidence. The system calculates and outputs the relevance confidence score of each piece of evidence relative to the original question. Finally, the system sorts all evidence in descending order based on the scores, selecting the top-K (K=5) high-confidence pieces of evidence to form the final target evidence set. These pieces of evidence are then sequentially pieced together to form contextual text blocks, which are then sent to the traceable answer generation layer.

[0086] 4. Detailed Implementation of the Source-Tracing Answer Generation Layer In this embodiment, the traceable answer generation layer aims to generate evidence-supported medical answers using a large language model. The system first processes the final target evidence set output from the previous layer (…). As background context, this is combined with the original medical question to construct the final generated prompts. To ensure the reliability and interpretability of the answers, the system employs explicit source citation constraints, requiring the large language model to adhere to the "evidence first" principle during generation. This means that when stating each medical fact, the system must explicitly cite the corresponding evidence source in the context at the end of the sentence using an index marker. Finally, the system executes the large language model's reasoning and generation process, outputting medical question-and-answer results containing a complete citation chain, thereby achieving accurate source citation of the factual basis of the answer.

[0087] (1) The proposed solution differs from traditional single knowledge base retrieval or simple network retrieval. It proposes a collaborative framework that integrates "local standard medical knowledge" and "open domain dynamic network information" in parallel, namely a collaborative dual-source retrieval enhancement generation architecture. Heterogeneous fusion is carried out in the reordering layer to achieve complementary fusion of local authoritative medical knowledge and Internet dynamic information. It effectively solves the problem of information lag in a single local knowledge base and the problem of insufficient professionalism in a single network search, and significantly improves the knowledge coverage and timeliness of the medical question answering system.

[0088] (2) The task-adaptive query rewriting mechanism proposed in this invention can generate optimized query representations for the characteristics of different retrieval paths, overcoming the shortcomings of the traditional single query method with low recall rate of related documents on heterogeneous data sources, and greatly improving the retrieval accuracy of medical term variations and cutting-edge medical knowledge.

[0089] (3) The iterative network retrieval strategy based on “retrieval-cleaning-refinement-reflection” proposed in this invention controls the retrieval depth by dynamically verifying the completeness of evidence, avoiding invalid web page browsing and wasting computing resources, and improving the system’s operating efficiency.

[0090] (4) The heterogeneous evidence reordering scheme based on cross encoder proposed in this invention eliminates the differences in scoring criteria between local documents and online texts, effectively removes noise from a large number of search results, and ensures high relevance and low interference of the input context of the large language model.

[0091] (5) The traceable answer generation method proposed in this invention forces the model to mark the source of the medical facts cited, making the generated medical advice verifiable and effectively enhancing the interpretability and user trust of the system in clinical auxiliary scenarios.

[0092] Another aspect of the present invention provides an electronic device, which includes a processor, a memory, a communication bus, and a communication interface.

[0093] in: The processor, memory, and communication interface communicate with each other via a communication bus.

[0094] A communication interface is used to communicate with other electronic devices or servers.

[0095] The processor is used to execute programs, specifically, to execute any of the steps of the medical question-answering method based on collaborative dual-source retrieval enhancement in the above embodiments.

[0096] Specifically, the program may include program code, which includes computer operation instructions.

[0097] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0098] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0099] Specifically, the program can be used to cause the processor to execute the steps of any of the medical question-answering methods based on collaborative dual-source retrieval enhancement described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed by any of the medical question-answering methods based on collaborative dual-source retrieval enhancement described above, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0100] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods of various embodiments of this application.

[0101] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0102] Specific embodiments of the present invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result.

[0103] It should be noted that in the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0104] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0105] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0106] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical question-answering method based on collaborative dual-source retrieval enhancement generation, characterized in that, include: Based on the input original medical question and candidate options, the large language model performs adaptive query rewriting to generate differential diagnosis keywords adapted to web retrieval, entity augmentation queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval. For the aforementioned diagnostic keywords, perform an iterative network search with a reflective mechanism to obtain candidate network evidence; For the entity augmentation query and the hypothetical medical summary, a hybrid search is performed in the local medical knowledge base to obtain local candidate evidence; The network candidate evidence and the local candidate evidence are aggregated and deduplicated to obtain aggregated heterogeneous evidence; The target evidence set is selected by performing a deep semantic relevance score on the aggregated heterogeneous evidence using a cross-encoder model. The original medical question and the target evidence set are concatenated and used as context input into a generative large language model to generate a final medical solution containing evidence citation tags.

2. The method according to claim 1, characterized in that, The adaptive query rewriting process, which utilizes a large language model to generate differential diagnostic keywords adapted for web retrieval, entity augmentation queries adapted for local sparse retrieval, and hypothetical medical summaries adapted for local dense retrieval, includes: Construct a first prompt instruction to guide the large language model to analyze the original medical question and candidate options, extract core medical features or identification points to distinguish the correctness of each option, and generate the differential diagnosis keywords; A second prompt instruction is constructed to guide the large language model to identify core medical entities from the original medical questions and candidate options, and to perform full name completion and synonym expansion to generate the entity-enhanced query; A third prompt instruction is constructed to guide the large language model to generate a declarative summary in the style of a medical textbook or clinical guideline based on the original medical question and candidate options, thus obtaining the hypothetical medical summary.

3. The method according to claim 1, characterized in that, The step of performing an iterative network retrieval with a reflective mechanism for the aforementioned diagnostic keywords to obtain network candidate evidence includes: Search initialization: Submit the diagnostic keywords to the Internet search engine and set the search pointer; Pagination execution and retrieval: Retrieve raw webpage data returned by internet search engines in batches according to the set window size; Cleaning and Refining: The original web page data is cleaned of noise to obtain the cleaned web page text; medical facts, clinical data or definitions related to the current differential diagnosis keywords are extracted from the cleaned web page text using a large language model and stored in a temporary evidence pool for information accumulation; Reflection based on information saturation: Input the accumulated information in the temporary evidence pool and the current differential diagnosis keywords into the large language model to determine whether the medical information of the current differential diagnosis keywords has been sufficiently covered; if the preset maximum search depth threshold has not been reached, move the search pointer and return to the pagination execution and retrieval steps; if the preset maximum search depth threshold has been reached, stop the search. Output: Aggregate all accumulated information in the temporary evidence pool obtained after stopping the search into the network candidate evidence.

4. The method according to claim 1, characterized in that, The process of performing a hybrid search in the local medical knowledge base to obtain local candidate evidence for the entity-enhanced query and the hypothetical medical summary includes: For the entity enhancement query, keyword retrieval based on the BM25 algorithm is performed through a pre-built inverted index to obtain a first candidate document list; For the hypothetical medical summary, it is mapped into a high-dimensional dense query vector using a pre-trained vector encoding model, and an approximate nearest neighbor search is performed by calculating cosine similarity to obtain a second candidate document list; The document fragments in the first candidate document list and the second candidate document list are merged and reordered using a reciprocal sorting fusion algorithm to generate the local candidate evidence.

5. The method according to claim 4, characterized in that, The step of fusing and reordering document fragments in the first and second candidate document lists using a reciprocal sorting fusion algorithm to generate the local candidate evidence includes: Calculate the fusion score for each document fragment in the first candidate document list and the second candidate document list: All document fragments are sorted in descending order based on the fusion score to generate a fused candidate document list. From the fused candidate document list, the top-M document fragments with the highest fusion scores are selected as the local candidate evidence.

6. The method according to claim 5, characterized in that, The fusion score for each document fragment is calculated using the following formula: in, It is a smoothing constant. For document fragments In the list The ranking position in the system.

7. The method according to claim 1, characterized in that, The process involves using a cross-encoder model to perform deep semantic relevance scoring on the aggregated heterogeneous evidence, and then filtering out the target evidence set, including: Each piece of evidence in the aggregated heterogeneous evidence is concatenated with the original medical question to construct a text pair; The text pairs are input into a pre-trained cross-encoder model, and the self-attention mechanism within the cross-encoder model is used to calculate the relevance confidence score of the evidence relative to the original medical question. All evidence was sorted in descending order based on the relevance confidence scores. The top-K pieces of high-confidence evidence are selected from the evidence arranged in descending order to form the target evidence set.

8. A medical question-answering system based on collaborative dual-source retrieval and enhanced generation, characterized in that, include: The task-adaptive query rewriting module is used to perform task-adaptive query rewriting based on the input original medical question and candidate options through a large language model, generating differential diagnosis keywords adapted to web retrieval, entity augmentation queries adapted to local sparse retrieval, and hypothetical medical summaries adapted to local dense retrieval. The collaborative dual-source retrieval module is used to perform an iterative network retrieval with a reflective mechanism for the differential diagnosis keywords to obtain network candidate evidence; and to perform a hybrid retrieval in the local medical knowledge base for the entity augmentation query and the hypothetical medical summary to obtain local candidate evidence. The heterogeneous evidence reordering module is used to aggregate and deduplicatize the network candidate evidence and the local candidate evidence to obtain aggregated heterogeneous evidence; The target evidence set is selected by performing a deep semantic relevance score on the aggregated heterogeneous evidence using a cross-encoder model. The traceable answer generation module is used to concatenate the original medical question with the target evidence set as context input into the generative large language model to generate a final medical answer containing evidence citation tags.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

Images