A biomedical literature scientific question extraction system and extraction method

Abstract

The application discloses a biomedical literature scientific problem extraction system and method, belonging to the field of biomedical literature data processing, the extraction system contains semantic perception and preprocessing, cascade logic screening, core element abstraction and extraction, structured output module in turn; the method includes five steps of data access cleaning, cascade logic screening, core element extraction, logic review reconstruction and structured packaging storage, through the research paradigm identification to realize accurate screening, to reconstruct and extract macro scientific problems by demolecularization, to establish bilingual terminology mapping synchronously, and finally to output ICD-11 standardized JSON format structured data. The application greatly filters non-related literature, improves the efficiency of scientific research information acquisition, realizes scientific problem cross-field reuse, provides high-quality data support for biomedical knowledge graph construction and vertical field AI model training, and solves the technical problems of low search accuracy of existing tools, distortion of core scientific problem extraction and non-reuse of unstructured text.

Description

Technical Field

[0001] This invention belongs to the field of biomedical literature data processing, specifically relating to a system and method for extracting scientific questions from biomedical literature. Background Technology

[0002] Against the backdrop of the explosion of biomedical big data, researchers face a severe challenge of "information overload" when designing research projects, tracking cutting-edge advancements, or building domain knowledge bases. Existing literature retrieval and analysis tools, due to their inherent technical limitations, are unable to meet the high-precision research needs. The core technical pain points are reflected in three aspects: First, the semantic understanding capabilities of retrieval systems are insufficient, relying solely on Boolean logic and keyword matching, lacking the ability to semantically discriminate the "research paradigm" of documents, resulting in mixed research types in the search results, requiring researchers to spend a lot of time manually screening them; Second, automated abstracting technology lacks adaptability to the biomedical field, either over-generalizing and losing key scientific findings, or over-detailing and obscuring core biological logic, failing to quickly capture core scientific questions to support the construction of scientific hypotheses; Third, scientific literature exists in the form of unstructured natural language, lacking standardized definitions and automated extraction methods, making it difficult to transform into reusable structured knowledge, forming "data silos," which limits downstream knowledge graph construction, interdisciplinary pattern discovery, and vertical domain AI model training.

[0003] Currently, technologies that assist in reading scientific literature and analyzing information can be mainly divided into three categories, and all of them have significant technical limitations: Traditional keyword-based retrieval techniques rely on string matching between user-input query terms and document index terms. This has inherent flaws such as "inability to match synonyms" and "inability to distinguish polysemous words." Furthermore, it cannot understand the "methodological characteristics" of documents and cannot distinguish between documents from different research paradigms under the same research topic. Extractive automatic summarization technology calculates sentence weights through statistical models and extracts key sentences from the original text to form a summary. The generated summary lacks semantic coherence and is easily taken out of context, making it difficult to logically integrate key conclusions scattered across multiple paragraphs in biomedical literature. Generative tools based on general large language models summarize the entire text by utilizing the sequence generation capabilities of large language models. However, they suffer from problems such as poor controllability, domain knowledge bias, and the risk of illusion. The output format is unstable and cannot meet the standardization requirements of database storage. Furthermore, they are prone to piling up terms to introduce noisy information, and may even fabricate experimental data or causal relationships.

[0004] In summary, existing technologies lack an automated system capable of deeply understanding biomedical research paradigms and performing "denoising" and "induction" according to specific academic standards. This makes it impossible to accurately transform massive amounts of messy biomedical literature into high-value, structured scientific question data. There is an urgent need for a literature scientific question extraction technology adapted to the biomedical field to solve the above problems. Summary of the Invention

[0005] To address the technical problems of low retrieval accuracy, distorted extraction of core scientific questions, and inability to convert unstructured text into standardized structured knowledge in existing literature processing tools, this application designs a biomedical literature scientific question extraction system and method. This system enables accurate screening of biomedical literature and structured extraction of core scientific questions, improves the efficiency of scientific research intelligence acquisition, and provides high-quality structured data for knowledge graph construction and AI model training in the biomedical field.

[0006] A system for extracting scientific questions from biomedical literature includes a semantic perception and preprocessing module, a cascaded logical filtering module, a core element abstraction and extraction module, and a structured output module that work in sequence and in collaboration. The semantic awareness and preprocessing module is used to access multi-source heterogeneous biomedical literature data (data format limited to XML or JSON), extract the text fields of title and abstracts for structured parsing, accurately remove irrelevant non-semantic information such as copyright notices, copyright symbols, reference citation marks, author biographies, acknowledgments, formula symbols, special characters, garbled characters, and advertising information, and output standardized semantic input data in a structured text format containing the title and abstract fields (full-text access is not supported; if any field of title or abstract is missing or contains continuous garbled characters, it is judged as invalid data, and an error code and prompt message are returned to the front end via the WebSocket interface; this step improves processing speed by 90% and covers 98% of core scientific questions). The cascaded logic filtering module is used to perform two-level semantic filtering on standardized semantic input data based on semantic features (priority is domain adaptability determination and research paradigm identification, respectively; the determination time accounts for 20% and 80%, respectively, and the domain determination with less computation is placed first based on the funnel principle). It eliminates non-biomedical literature and biomedical literature that does not conform to basic research paradigms step by step, and outputs basic research literature semantic data that meets the requirements (the literature retention rate after filtering is about 25%-30%). The core element abstraction and extraction module is used to perform multi-dimensional information extraction and de-molecularized logical reconstruction on the semantic data of basic research literature (based on the asynchronous coroutine scheduling mechanism of directed acyclic graphs, parallel extraction of disease entities and scientific questions), extracting standardized disease entities (with an extraction accuracy of 98%) and macro-core scientific questions (judgment criteria are strictly controlled to 80-100 words, must be pure Chinese declarative sentences, only describing pathological phenotypes or physiological function changes caused by abnormal cell / tissue function), and establishing a mapping relationship between Chinese scientific concepts and original English academic terms (using semantic mapping technology with fine-tuning of large language model instructions, and the transformation rule base is derived from the structure-function correspondence corpus of millions of high-quality basic medical review literatures). The structured output module is used to aggregate and associate standardized disease entities, macro-level core scientific questions, and mapping relationships between Chinese and English terms through unique task_id and pmid (if a field is empty, it is filled with the default value null without interrupting the encapsulation; in case of conflict, the output of the final node shall prevail; the redundancy check rule is to calculate the MD5 hash value of the scientific question to remove duplicates), encapsulate it into flat JSON format data that does not contain nested objects (including task_id, pmid, publish_year, disease_name, icd_11_code, scientific_problem, and english_keywords fields), and complete the storage to the target database (such as MongoDB document database, which supports a concurrent processing capacity of up to 500 TPS, and the data storage of every 100,000 records does not exceed 50MB).

[0007] Preferably, the irrelevant non-semantic information removed by the semantic perception and preprocessing module includes copyright statements, copyright symbols, reference citation marks, author biographies, acknowledgments, formula symbols, special characters, garbled text, and advertising information; The cascaded logical filtering module's two-level semantic filtering includes domain adaptability determination and research paradigm identification. Domain adaptability determination identifies the core entity type and context of the text through a semantic analysis model, and calculates semantic similarity by combining the biomedical domain ontology (using UMLS 2023AA version, with the semantic similarity determination threshold set to 0.85) to accurately eliminate non-biomedical domain literature. The research paradigm identification method accurately distinguishes the core distinguishing features of three research paradigms—basic research (focusing on molecular / cellular mechanisms), clinical research (focusing on the efficacy of patients / clinical samples), and bioinformatics analysis (focusing on public database computational models)—by analyzing the text's logical structure and research methodology characteristics, and eliminates literature from the latter two non-basic research paradigms.

[0008] Preferably, the core element abstraction and extraction module includes an entity standardization subunit, a logical induction subunit, and a de-mineralization reconstruction subunit; the various technical modules are asynchronously decoupled using RabbitMQ message queues, data flow is unidirectionally transmitted using JSON data packets, and hardware deployment adopts a Kubernetes containerized architecture that supports GPU acceleration. The standardized entity subunit calls the large language model API to map the identified disease names to the ICD-11 standard coding system through string matching and semantic similarity calculation. When multiple codes are matched, the subclass code with a deeper level is selected first. When the mapping fails, the original Chinese name is retained and the ICD code is set to null to trigger manual review. Finally, standardized disease entities separated by ";" are generated. The logical induction subunit parses text sentences based on dependency parsing trees, extracts "subject-verb-object" triples to capture the core causal chain (capture accuracy reaches 92%), and extracts preliminary scientific conclusions containing at least one complete "molecule-function-phenotype" chain (if an extraction error is made, the large model reflection mechanism is triggered to retry up to 1 time). The demolecularization reconstruction subunit uses over 50 negative constraint rules across 4 categories, updated quarterly by medical experts (including excluding specific molecules such as genes / proteins, prohibiting descriptions of mechanisms such as epigenetic / signaling pathways, prohibiting methodological terms such as CRISPR, and prohibiting descriptions of genetic variations such as specific gene mutations). It scans and filters out microscopic details such as molecular entities in preliminary scientific conclusions by labeling them with redundant tags, retaining the core causal chain containing direct causal mapping. Then, through semantic mapping technology based on fine-tuning instructions from a corpus of millions of words, it transforms experimental observation indicators into descriptions of general biological processes or cell functional states, reconstructing them into macroscopic core scientific questions (reconstruction effect evaluation indicators: 100% complete demolecularization rate, logical causal integrity >95%).

[0009] Preferably, the core element abstraction and extraction module uses an attention mechanism (calculating the cosine similarity between the English translation vector corresponding to the Chinese concept and the vectors of each phrase in the original abstract, and removing those if the similarity is below a threshold) to back locate key English phrases in the original document, thereby achieving accurate anchoring and structured mapping between Chinese scientific concepts and original English academic terms (100% matching degree between Chinese and English terms, and returning an empty string when anchoring fails without affecting Chinese output). The structured output module encapsulates standardized structured data (using Gzip compression algorithm, compression ratio 1:5) into a flat key-value pair constraint JSON format object, and stores the object into the MongoDB target database through a RESTful API interface (HTTPS transmission, based on API key authentication), and provides a GraphQL interface to adapt to downstream multi-scenario needs.

[0010] Based on the biomedical literature scientific question extraction system, this application also designs a method for extracting scientific questions from biomedical literature, including the following steps: Step S1, Data Access and Cleaning: Receive the metadata stream of biomedical literature to be processed (only XML or JSON format is supported, full text is not supported), perform noise reduction on the literature title and abstract, remove irrelevant non-semantic information such as copyright statement, author biography, formula symbols, and special characters, and output standardized semantic input data containing title and abstracts fields (missing fields are directly judged as invalid data and an error is thrown). Step S2, Cascaded Logic Screening: Domain adaptability determination (20% of the time) and research paradigm identification (80% of the time) are performed sequentially on the standardized semantic input data. Non-biomedical literature and biomedical literature that do not conform to basic research paradigms are eliminated step by step. Finally, the semantic data of basic research literature with a retention rate of about 25%-30% is output. Step S3, Core Element Extraction: Through the directed acyclic graph task scheduling mechanism, the semantic data of basic research literature is processed in parallel to extract standardized disease entities and summarize original scientific conclusions (using a logic reasoning algorithm based on self-attention mechanism), outputting standardized disease entities and preliminary scientific conclusions containing a complete "molecular-functional-phenotype" chain; Step S4, Logical Review and Reconstruction: Demolecularization is performed on the preliminary scientific conclusions. Based on more than 50 negative rules in 4 categories, microscopic details are stripped away and reconstructed into macroscopic core scientific questions with a word count of 80-100. At the same time, cosine similarity is calculated to establish a precise mapping relationship between Chinese scientific concepts and original English academic terms. Step S5, Structured Encapsulation and Storage: Based on pmid, primary key association is performed to aggregate and encapsulate standardized disease entities, macro-core scientific questions, and Chinese-English terminology mapping relationships into standardized structured data in flat JSON format (including fields such as task_id). High-concurrency data is then stored in the MongoDB database via a RESTful interface (configured with daily full backup and off-site multi-active disaster recovery mechanism).

[0011] Preferably, in step S2, the domain suitability determination identifies the core entity type and context of the text through a semantic analysis model, and detects whether the proportion of biomedical entities in the text reaches a set threshold (based on statistics of 100,000 annotated data, the threshold is set to 85%). If it does not reach the threshold, it is determined to be non-biomedical literature and the processing flow is terminated. The research paradigm identification analyzes the predicate verbs (such as "regulate", "mediate", "promote") and research object characteristics (such as "cell line", "animal model") in the text. If clinical statistical characteristics (such as "randomized controlled trial", "overall survival") are identified, it is determined to be clinical research literature. If bioinformatics analysis characteristics (such as "machine learning", "prognostic signature") are identified, it is determined to be bioinformatics analysis research literature, and the processing flow is terminated. If basic experimental characteristics (including hypothesis testing experiments driven by mechanisms such as CRISPR-Cas9, with an accuracy rate of 95%) are identified, it is determined to be basic research literature and proceeds to the next step.

[0012] Preferably, in step S3, the specific method for standardizing the extraction of disease entities is to call a large language model to extract disease names from semantic data of basic research literature, and map them to the ICD-11 standard coding system through string matching and semantic similarity calculation (when multiple coding matches are performed, the deeper level coding is selected first, and null is returned when there is an anomaly), generating standardized disease entities containing disease names and corresponding ICD-11 codes; the original scientific conclusions are summarized into preliminary scientific conclusions containing complete "molecular-functional-phenotypic" molecular mechanisms extracted based on dependency parsing trees (if the extraction is incorrect, the large model's self-reflection mechanism is triggered to retry up to 1 time).

[0013] Preferably, in step S4, the demolecularization process includes the following sub-steps: Step S41: Using over 50 negative constraint rules in 4 major categories updated quarterly by medical experts (such as excluding specific molecules like genes / proteins / metabolites, prohibiting descriptions of mechanisms like epigenetics / signaling pathways, prohibiting methodological terms like ChIP-seq, and prohibiting descriptions of specific gene mutations), scan the microscopic details in the preliminary scientific conclusions and mark them as redundant features with tags such as [REDUNDANT_MOLECULE]. Step S42: Filter the redundant features and retain the core causal chain containing the direct causal mapping from "biological state or structural change" to "abnormal cell or tissue function" (using a filtering algorithm that combines keyword regular expression matching and syntax tree pruning from a rule engine). Step S43: Using semantic mapping technology based on fine-tuning of large language model instructions (the transformation rule base is derived from millions of high-quality review corpora), the filtered experimental observation indicators are transformed into general biological process descriptions or cell functional state descriptions, and reconstructed into macroscopic core scientific questions in pure Chinese declarative sentences with a word count strictly limited to 80-100 characters. The mapping relationship between Chinese scientific concepts and original English academic terms is established by back-locating key English phrases in the original literature through an attention mechanism (calculating the cosine similarity between the concept translation vector and the original abstract phrase vector; if the match fails, an empty string is returned) to achieve precise anchoring.

[0014] Preferably, in step S1, the biomedical literature metadata stream only contains the literature title and abstract information (accessed via OpenAPI authorized by OAuth 2.0, and converted into a standard model by the unified mapping middleware); the irrelevant non-semantic information removed by the noise reduction process includes copyright statements, copyright symbols, reference citation marks, author biographies, acknowledgments, formula symbols, special characters, and advertising information.

[0015] Preferably, in step S5, the standardized structured data is associated through the pmid primary key and encapsulated in a flattened JSON format constrained by data type (e.g., {"disease_name":"cerebellar degenerative disease","scientific_problem":"mitochondrial dysfunction leads to morphological and functional damage to Purkinje cells in the cerebellum through inducing autophagy.","english_keywords":"mitochondrial function;mitophagy;Purkinje cells"}, etc.). The encapsulated content includes the standardized disease name and corresponding ICD-11 code, 80-100 Chinese macroscopic core scientific questions, original English academic terms that are precisely mapped to Chinese scientific concepts (separated by ";"), task_id, and publish_year. The structured data after high-concurrency storage (500 TPS) is used for front-end retrieval, biomedical knowledge graph construction, or training of AI models in the biomedical vertical field.

[0016] The advantages and effects of this application are as follows: This application presents a biomedical literature scientific question extraction system that achieves automated and accurate identification of biomedical fields and basic research paradigms through cascaded logical screening. It can filter out more than 90% of irrelevant and interfering literature, allowing researchers to focus on basic research literature with mechanistic reference value, significantly reducing the cost of manual screening, improving the efficiency of literature reading and information acquisition, and greatly improving the signal-to-noise ratio of scientific information acquisition.

[0017] This application presents a biomedical literature scientific question extraction system that uses a "de-molecularization" abstract reconstruction method to strip away microscopic molecular details while retaining universal biological core logic. The resulting macroscopic core scientific questions are easier to compare and transfer between different disease models, which helps inspire researchers' interdisciplinary innovative hypotheses and promotes cross-domain knowledge discovery and transfer; it also enables the accurate extraction and cross-domain reuse of core scientific questions.

[0018] This application presents a biomedical literature scientific question extraction system that, through a standardized extraction and output process, transforms massive amounts of unstructured biomedical literature into JSON format data containing standardized disease codes, bilingual terminology mappings, and structured scientific questions. This lays a solid underlying data foundation for building high-precision knowledge graphs and training vertical domain AI models in the biomedical field, promotes the transformation of biomedical research towards a "data-driven" model, and breaks down the unstructured data barrier of biomedical literature.

[0019] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0020] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0022] Figure 1 Flowchart of the method for batch extraction of scientific questions from literature designed for this application; Figure 2 A block diagram of the batch literature scientific question extraction system designed for this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0024] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0025] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0026] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0027] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0028] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0029] Example 1: This invention addresses the core pain points of existing biomedical literature processing technologies, such as insufficient semantic understanding capabilities of retrieval systems, lack of domain adaptability in automated summarization technologies, and the inability to transform unstructured text into reusable structured knowledge. It abandons the single black-box model design of existing technologies and constructs a closed-loop workflow encompassing preprocessing, multi-level screening, deep extraction, and quality verification. Through the sequential collaborative work of semantic perception and preprocessing modules, cascaded logical screening modules, core element abstraction and extraction modules, and structured output modules, combined with standardized methods for extracting scientific questions from biomedical literature, it simulates the cognitive logic of experienced researchers to construct an automated screening and summarization pipeline. This achieves the accurate transformation from massive amounts of messy text into high-value, structured scientific question data. The specific implementation methods of this invention are described in detail below with reference to the accompanying drawings.

[0030] Specific implementation of system modules: like Figure 2 The diagram shown illustrates the module diagram of the batch literature scientific question extraction system of this invention. This module diagram adopts a "semantic distribution-parallel extraction" processing mode and is divided into an input layer, an extraction engine core, and an output layer. The extraction engine core comprises the four core modules mentioned above and is also the main execution body of the system. The input layer is responsible for receiving and parsing the original Markdown / text format literature metadata stream, and the output layer aggregates the processing products of each module to generate standardized structured data packets. The biomedical literature scientific question extraction system of this invention is a multi-module collaborative closed-loop processing system. Each module is specifically designed to address the shortcomings of existing technologies. The specific implementation principle and working process are as follows: Semantic Awareness and Preprocessing Module: Addressing the issue of multi-source heterogeneity and the presence of significant non-semantic noise in existing literature data, this module serves as the core processing unit of the system's input layer. It first accesses multi-source heterogeneous biomedical literature data from mainstream databases such as PubMed and Web of Science, extracting only the titles and abstracts—the two elements containing core research information—to form a metadata stream, thus avoiding the efficiency losses associated with full-text data processing. Subsequently, it performs text cleaning, sentence segmentation, and structured parsing on the titles and abstracts, accurately removing irrelevant non-semantic information such as copyright notices, copyright symbols, and reference citations. This completes the standardization of the text, outputting standardized semantic input data with a unified format and clear semantics. This provides a unified input foundation for subsequent semantic filtering and information extraction, solving the problem of a lack of standardization in data preprocessing in existing technologies.

[0031] Cascaded logic filtering module: This module is Figure 2The core hub of the engine, namely the semantic filtering and distribution module, addresses the shortcomings of existing retrieval technologies that can only achieve keyword matching and cannot identify research paradigms. Based on semantic features, it performs two-level semantic filtering on standardized semantic input data: domain adaptability determination and research paradigm identification. This enables the step-by-step elimination of irrelevant literature and the distribution of compliant data to downstream modules. In the domain suitability assessment stage, the core entity types and context of the text are identified through a semantic analysis model. The text content is then semantically matched with a biomedical domain ontology. By setting a threshold for the proportion of biomedical entities, literature from non-biomedical domains is accurately eliminated, solving the problem of domain mixing in existing search results. In the research paradigm identification stage, the logical structure and research method characteristics of the text are deeply analyzed. By identifying the predicate verbs and research object characteristics in the text, three research paradigms—basic research, clinical research, and bioinformatics analysis research—are accurately distinguished. Only literature from the basic research paradigm, which focuses on "mechanism exploration and experimental verification," is retained, while literature from the clinical / bioinformatics analysis paradigm, which has no mechanism reference value, is eliminated. Finally, semantic data of basic research literature that meets the high-precision needs of researchers is output and distributed to the core element abstraction and extraction module, solving the industry problem of mixed research types and low precision in existing technologies.

[0032] Core element abstraction and extraction module: corresponding Figure 2 This system extracts the core logical abstraction and reconstruction module, the terminology bilingual alignment module, and the disease entity standardization module as collaborative working units. Addressing the issues of overgeneralization or over-detailing and distorted core logic extraction in existing automated summarization technologies, this core computational unit comprises entity standardization subunits, logical induction subunits, and de-molecularization reconstruction subunits. This enables multi-dimensional information extraction and de-molecularization logical reconstruction, achieving a leap from microscopic molecular mechanisms to macroscopic biological logic. The entity standardization subunit is... Figure 2 The disease entity standardization module identifies disease names from semantic data in basic research literature and accurately maps them to the ICD-11 standard coding system, solving the problem of inconsistent disease name representation in existing technologies and achieving standardization of disease entities. The logical induction subunit parses complex sentence structures in texts, accurately capturing the core causal chains in research, and extracts preliminary scientific conclusions containing complete molecular mechanisms based on full-text semantic extraction, solving the problem that existing extraction methods cannot integrate logic across paragraphs. The de-molecularization reconstruction subunit... Figure 2The logical abstraction and reconstruction module incorporates negative constraint rules built upon expert knowledge. These rules scan preliminary scientific conclusions for microscopic details such as molecular entities and experimental parameters, marking them as redundant features. After filtering these redundant features, only the core causal chain is retained. Semantic mapping technology then transforms the filtered experimental observation indicators into general descriptions of biological processes or cell functional states, ultimately reconstructing the macroscopic core scientific question stripped of microscopic details. This addresses the problem in existing technologies of "seeing the trees but not the forest," where core scientific questions are obscured by microscopic details. Simultaneously, the bilingual terminology alignment subunit within this module uses an attention mechanism to back-locate key English phrases in the original text, achieving precise anchoring and structured mapping between Chinese scientific concepts and original English academic terms. This establishes a one-to-one bilingual terminology mapping relationship, supporting cross-language retrieval and knowledge tracing.

[0033] Structured output module: interface Figure 2 The system output layer addresses the issue of unstructured and unreusable literature data in existing technologies. It aggregates and organizes standardized disease entities (including ICD-11 codes), macro-level core scientific questions, and Chinese-English terminology mapping relationships output by the core element abstraction and extraction module. This data is then encapsulated into a structured data object in JSON format according to standardized database storage requirements. This data object is subsequently stored in the target database. The JSON-formatted structured data can be directly accessed by the front-end system and used directly for downstream data analysis, solving the "data silo" problem in biomedical literature and providing standardized underlying data for subsequent knowledge graph construction and AI model training.

[0034] Specific implementation of extraction method like Figure 1 The diagram shown is a flowchart of the method for batch extraction of scientific questions from literature according to the present invention. This flowchart illustrates the entire process from literature input to final structured output, corresponding one-to-one with the extraction steps S1-S5 of the present invention. The biomedical literature scientific question extraction method of the present invention is a standardized data processing pipeline adapted to the above-mentioned system. Each step corresponds one-to-one with the system module and is seamlessly connected. The specific implementation steps and operational details are illustrated in the attached flowchart below: Step S1, Data Access and Cleaning: Corresponding to Figure 1In the "Literature Input (Title / Abstract)" stage, the system receives the metadata stream of biomedical literature to be processed. This metadata stream only contains the title and abstract information, effectively reducing the amount of data processing. Targeted denoising processing is performed on the title and abstract to accurately remove irrelevant non-semantic information such as copyright notices, copyright symbols, and reference citation marks, avoiding interference from non-semantic information in subsequent semantic analysis. After denoising and structured parsing, standardized semantic input data is output. If the data to be processed has invalid issues such as disordered format or missing characters, it is directly judged as invalid data and the processing flow is terminated to ensure the validity of the input data and lay the foundation for the subsequent screening stage.

[0035] Step S2, Cascaded Logic Filtering: Corresponding Figure 1 The system employs a dual screening process: "Level 1 Screening (Relevant Literature Removal)" and "Level 2 Screening (Clinical / Bioinformatics Literature Removal)." Standardized semantic input data undergoes domain suitability assessment and research paradigm identification sequentially. Both levels of screening have an "absolute termination if not passed" rule to improve screening efficiency. Domain suitability assessment, the Level 1 screening, uses a semantic analysis model to identify the core entity types and context of the text, checking if the proportion of biomedical entities in the text reaches a set threshold. If it doesn't, the text is classified as non-biomedical literature, and the process terminates immediately. If the threshold is reached, the system proceeds to the Level 2 research paradigm identification stage. This stage analyzes the predicate verbs and research object characteristics in the text. Identifying clinical statistical characteristics classifies it as clinical research literature, while identifying bioinformatics analysis characteristics classifies it as bioinformatics analysis research literature, both of which terminate the process immediately. Only when basic experimental characteristics are identified is it classified as basic research literature, outputting semantic data for basic research literature and proceeding to the next step. This achieves accurate and rapid removal of irrelevant literature, ultimately retaining only basic research literature. Figure 1 The literature data in the "basic research" category that meets the requirements.

[0036] Step S3, Core Element Extraction: Corresponding Figure 1 The parallel processes of "disease nomenclature standardization (ICD-11)" and "preliminary summarization of scientific questions" in basic research literature involve simultaneously initiating two sub-tasks: disease entity standardization extraction and original scientific conclusion summarization, to extract core elements synchronously and improve processing efficiency. Disease entity standardization extraction... Figure 1 The standardization of disease names involves accurately extracting disease names from semantic data in basic research literature, strictly mapping them to the ICD-11 standard coding system, and generating standardized disease entities containing disease names and corresponding ICD-11 codes, thus achieving standardization and unification of disease names; the original scientific conclusions are summarized as follows: Figure 1The scientific questions in the text are initially summarized, and a deep analysis is performed based on the semantics of the whole text to sort out the causal logical relationships in the text and extract preliminary scientific conclusions containing complete molecular mechanisms to ensure the integrity of the scientific conclusions. The results are output synchronously after the two sub-tasks are completed, providing a foundation for subsequent logical reconstruction.

[0037] Step S4, Logic Review and Restructuring: Corresponding to Figure 1 The "Intelligent Review: De-molecularization Reconstruction" step is a crucial step in resolving distortions in the extraction of core scientific questions. It involves first de-molecularizing the preliminary scientific conclusions, and then constructing a mapping relationship between Chinese and English terminology. The de-molecularization process consists of three sub-steps. First, using negative constraint rules constructed from expert knowledge, it comprehensively scans the microscopic details of molecular entities and experimental technical parameters in the preliminary scientific conclusions and marks them as redundant features. Then, it filters out the marked redundant features, retaining only the core causal chains in the research. Finally, through semantic mapping technology, it transforms the filtered experimental observation indicators into general descriptions of biological processes or cell functional states, reconstructing them into macroscopic core scientific questions, achieving a logical leap from the microscopic to the macroscopic. Simultaneously with the de-molecularization reconstruction, an attention mechanism is used to reverse-locate key English phrases in the original literature, precisely anchoring the scientific concepts in the reconstructed Chinese macroscopic core scientific questions with the original English academic terms, establishing a one-to-one structured mapping relationship to ensure the consistency and accuracy of bilingual terminology.

[0038] Step S5, Structured Packaging and Storage: Corresponding to Figure 1 The "data merging and encapsulation" and "structured output (JSON import)" steps aggregate standardized disease names and their corresponding ICD-11 codes, core Chinese macro-level scientific questions, and original English academic terms precisely mapped to Chinese scientific concepts. This data is then encapsulated into structured data in JSON format according to standardized requirements. The encapsulated content is complete and formatted uniformly, directly adaptable to various database storage needs. Subsequently, this structured data is stored in the target database to complete data storage. Figure 1 The final "end" stage is the reusable state of the data after it is stored in the database. The stored structured data can be directly used for front-end retrieval, providing researchers with an efficient intelligence query tool for project design and frontier tracking. It can also be directly used for biomedical knowledge graph construction or AI model training in the biomedical vertical field, realizing the efficient reuse of structured knowledge and promoting the transformation of biomedical research to a "data-driven" model.

[0039] The specific implementation of this invention achieves a high degree of compatibility between system modules and extraction methods, combined with... Figure 1 Methods and procedures Figure 2The system architecture forms a complete automated solution for extracting scientific questions from biomedical literature, which specifically addresses all the core pain points of existing technologies. It achieves the accurate transformation from massive and messy biomedical literature to high-value, structured scientific question data, providing researchers with an efficient tool for literature intelligence analysis and standardized data support for big data mining and intelligent research in the biomedical field.

[0040] The above description is merely a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter alterations to these embodiments within the spirit and principles of the present invention, achieved through conventional substitutions or by achieving the same function without departing from the principles and spirit of the present invention, fall within the scope of protection of the present invention.

Claims

1. A system for extracting scientific questions from biomedical literature, characterized in that, It includes a semantic perception and preprocessing module that works in sequence, a cascaded logical filtering module, a core element abstraction and extraction module, and a structured output module; The semantic awareness and preprocessing module is used to access multi-source heterogeneous biomedical literature data, extract the text fields of title and abstracts for structured parsing, remove irrelevant non-semantic information, and output standardized semantic input data in a structured text format containing the title and abstracts fields. The cascaded logical filtering module is used to perform two-level semantic filtering on standardized semantic input data based on semantic features, eliminating non-biomedical literature and non-basic research paradigm biomedical literature step by step, and outputting basic research literature semantic data that meets the requirements. The core element abstraction and extraction module is used to perform standardized extraction of disease entities and summarization of original scientific conclusions from semantic data in basic research literature. The specific method for standardizing disease entity extraction is to call a large language model to extract disease names from semantic data of basic research literature, and then map them to the ICD-11 standard coding system through string matching and semantic similarity calculation to generate standardized disease entities containing disease names and corresponding ICD-11 codes; the original scientific conclusions are summarized as preliminary scientific conclusions containing complete "molecular-functional-phenotypic" molecular mechanisms extracted based on dependency parsing trees. Then, the preliminary scientific conclusions are de-molecularized, stripped of microscopic details, and reconstructed into macroscopic core scientific questions. At the same time, a precise mapping relationship is established between Chinese scientific concepts and original English academic terms. The microscopic details include molecular entities and experimental technical parameters; The demolecularization process is as follows: by scanning with negative constraint rules and adding redundant labels to the microscopic details in the preliminary scientific conclusions, the core causal chain containing direct causal mapping is retained. Then, through semantic mapping technology based on instruction fine-tuning of millions of corpora, the experimental observation indicators are transformed into descriptions of general biological processes or cell functional states, and the preliminary scientific conclusions are reconstructed into macroscopic core scientific questions. The structured output module is used to aggregate and associate standardized disease entities, macro-level core scientific questions, and mapping relationships between Chinese and English terms through unique task_id and pmid, encapsulate them into flat format data, and complete the storage to the target database.

2. The biomedical literature scientific question extraction system according to claim 1, characterized in that, The semantic awareness and preprocessing module removes irrelevant non-semantic information including copyright statements, copyright symbols, reference citation marks, author biographies, acknowledgments, formula symbols, special characters, garbled text, and advertising information. The cascaded logical filtering module's two-level semantic filtering includes domain adaptability determination and research paradigm identification. Domain adaptability determination identifies the core entity type and context of the text through a semantic analysis model, and calculates semantic similarity by combining it with a biomedical domain ontology library to accurately eliminate non-biomedical literature. The research paradigm identification method accurately distinguishes the core distinguishing features of three research paradigms—basic research, clinical research, and bioinformatics analysis—by analyzing the text's logical structure and research methodology, and eliminates literature from the latter two non-basic research paradigms.

3. The biomedical literature scientific question extraction system according to claim 1, characterized in that, The core element abstraction and extraction module uses an attention mechanism to back locate key English phrases in the original text of the literature, thereby achieving precise anchoring and structured mapping between Chinese scientific concepts and original English academic terms. The structured output module encapsulates standardized structured data into flat JSON format data that cannot contain nested objects, stores the JSON format data into the target MongoDB database through a RESTful API interface, and provides a GraphQL interface to adapt to downstream multi-scenario needs.

4. A method for extracting scientific questions from biomedical literature, characterized in that, The biomedical literature scientific question extraction system applied to any one of claims 1-3 includes the following steps: Step S1, Data Access and Cleaning: Receive the metadata stream of biomedical literature to be processed, denoise the literature title and abstract, remove irrelevant non-semantic information, and output standardized semantic input data containing the title and abstracts fields; Step S2, Cascaded Logic Screening: Domain adaptability determination and research paradigm identification are performed sequentially on the standardized semantic input data, and non-biomedical literature and non-basic research paradigm biomedical literature are eliminated step by step, and basic research literature semantic data that meets the requirements are output. Step S3, Core Element Extraction: Perform disease entity standardization extraction and original scientific conclusion summarization on the semantic data of basic research literature in parallel, and output standardized disease entities and preliminary scientific conclusions; Step S4, Logical Review and Reconstruction: Demolecularize the preliminary scientific conclusions, remove microscopic details, reconstruct them into macroscopic core scientific questions, and establish a precise mapping relationship between Chinese scientific concepts and original English academic terms. The de-molecularization process involves: scanning with negative constraint rules and filtering out the microscopic details in the preliminary scientific conclusions by labeling them with redundant tags, retaining the core causal chain containing direct causal mapping, and then using semantic mapping technology based on fine-tuning instructions from a corpus of millions of words to transform experimental observation indicators into descriptions of general biological processes or cell functional states, reconstructing the preliminary scientific conclusions into macroscopic core scientific questions, and establishing a precise mapping relationship between Chinese scientific concepts and original English academic terms. Step S5, Structured Encapsulation and Storage: Based on pmid, primary key association is performed to aggregate and encapsulate standardized disease entities, macro-core scientific questions, and mapping relationships between Chinese and English terms into standardized structured data in a flat format, and the data is then stored in the target database.

5. The method for extracting scientific questions from biomedical literature according to claim 4, characterized in that, In step S2, the domain suitability determination uses a semantic analysis model to identify the core entity types and context of the text, and detects whether the proportion of biomedical entities in the text reaches a set threshold. If it does not reach the threshold, it is determined to be non-biomedical literature and the processing flow is terminated. The research paradigm identification analyzes the predicate verbs and research object characteristics in the text. If clinical statistical characteristics are identified, it is determined to be clinical research literature; if bioinformatics analysis characteristics are identified, it is determined to be bioinformatics analysis research literature, and the processing flow is terminated in both cases. If basic experimental characteristics are identified, it is determined to be basic research literature and proceeds to the next step. After two levels of screening, the final output is semantic data of basic research literature with a retention ratio of 25%-30%.

6. The method for extracting scientific questions from biomedical literature according to claim 4, characterized in that, In step S3, parallel processing is performed through a directed acyclic graph task scheduling mechanism; the specific method for standardizing disease entity extraction is to call a large language model to extract disease names from semantic data of basic research literature, and map them to the ICD-11 standard coding system through string matching and semantic similarity calculation to generate standardized disease entities containing disease names and corresponding ICD-11 codes; the original scientific conclusions are summarized as preliminary scientific conclusions containing complete "molecular-functional-phenotypic" molecular mechanisms extracted based on dependency parsing trees.

7. The method for extracting scientific questions from biomedical literature according to claim 4, characterized in that, In step S4, the demolecularization process includes the following sub-steps: Step S41: Using the four categories of negative constraint rules updated quarterly by medical experts, scan the microscopic details in the preliminary scientific conclusions and mark them as redundant features with the [REDUNDANT_MOLECULE] label; Step S42: Filter the redundant features and retain the core causal chain containing the direct causal mapping from "biological state or structural change" to "abnormal cell or tissue function"; Step S43: Through semantic mapping technology based on fine-tuning of large language model instructions, the filtered experimental observation indicators are transformed into general biological process descriptions or cell functional state descriptions, and reconstructed into macroscopic core scientific questions in pure Chinese declarative sentences with a word count strictly limited to 80-100 characters. The mapping relationship between Chinese scientific concepts and original English academic terms is accurately anchored by back-locating key English phrases in the original text through an attention mechanism, and then established through cosine similarity calculation.

8. The method for extracting scientific questions from biomedical literature according to claim 4, characterized in that, In step S1, the biomedical literature metadata stream contains only the literature title and abstract information; the denoising process removes irrelevant non-semantic information including copyright statements, copyright symbols, reference citation marks, author biographies, acknowledgments, formula symbols, special characters, garbled text, and advertising information.

9. The method for extracting scientific questions from biomedical literature according to claim 4, characterized in that, In step S5, the standardized structured data is associated through the pmid primary key and encapsulated in a flat JSON format constrained by data type. The encapsulated content includes standardized disease names and corresponding ICD-11 codes, 80-100 Chinese macro-core scientific questions, original English academic terms that are precisely mapped to Chinese scientific concepts, task_id, and publish_year. High-concurrency data is stored in the MongoDB database through a RESTful interface. The structured data after storage is used for front-end retrieval, biomedical knowledge graph construction, or training of AI models in the biomedical vertical field.

Images