A text information extraction method and system based on multi-agent large language model

Abstract

The application relates to a text information extraction method and system based on a multi-agent large language model. The method extracts a global named entity list in target text information based on a main-agent large language model, and the global named entity list at least includes a sample name list and an experimental method name list; constraint features are constructed according to the global named entity list, and prompt texts of at least two sub-agent large language models are constructed based on the constraint features; all sub-agent large language models are run in parallel, structured data fields are extracted from the target text information according to the constraint features in the corresponding prompt texts of each sub-agent large language model, and names in the structured data fields are unified based on the global named entity list. The application improves the accuracy and data integrity of long document multi-task extraction by constructing a two-stage collaborative architecture combining global coordination of a main agent and constraint of sub-agents in parallel and an explicit term constraint mechanism.

Description

Technical Field

[0001] This application relates to the field of text recognition technology, and in particular to a method and system for extracting text information based on a multi-agent large language model. Background Technology

[0002] In the field of materials science, for data-driven materials design and discovery research, the efficient, accurate, and automated extraction of structured experimental data from massive amounts of scientific literature has become a key technological bottleneck in constructing large-scale materials informatics databases and mining the materials genome. Current research relies on a large amount of high-quality published literature data. Taking silver-doped titanium-based antibacterial coatings as an example, a complete study typically includes a complete chain of experimental data, including material composition, preparation process, surface characterization, antibacterial properties, biocompatibility, and ion release. This data is scattered in unstructured form across multiple chapters, figures, and tables in the papers, such as Materials and Methods, Results and Discussion, forming a complex, multi-task information extraction challenge that relies on domain-specific knowledge.

[0003] Traditional methods for extracting information from literature primarily rely on rule-based pattern matching, such as regular expressions. These methods identify and extract specific entities and relationships in text through predefined string patterns. Their advantages lie in clear rules and computational efficiency. However, in complex texts like materials science literature, characterized by highly diverse expressions, specialized terminology, and numerous semantically equivalent variants, the limitations of traditional methods become glaringly apparent. First, they lack semantic understanding capabilities, failing to identify the same temperature expressed in different forms as the same entity, or to associate different names for the same sample, resulting in fragmented and inconsistent extraction results. Second, rule-based methods struggle to handle cross-sentence and cross-paragraph information connections, and cannot automatically integrate data on the same sample scattered across methods, results, and figures. Furthermore, maintaining and updating the rule base is costly and lacks scalability, failing to meet the accuracy requirements of large-scale automated extraction due to format differences and terminology variations across journals.

[0004] In recent years, large-scale language models have brought a new paradigm to scientific literature information extraction due to their powerful semantic understanding and context modeling capabilities. Methods based on single-agent large language models can theoretically overcome the semantic gap of traditional methods by handling multiple extraction tasks simultaneously with a single model call. However, when faced with the complex tasks contained in the literature, and each task having multiple fine-grained fields, a single agent needs to understand the entire text, identify all entities, extract all data, and maintain relationships in a single inference, which easily leads to cognitive overload. This manifests as the omission of key information, incorrect association of data from different samples, and the same sample being given different names in different tasks. The higher the document length, the number of tasks, and the complexity of the samples, the more significant this problem becomes, leading to a sharp decline in extraction accuracy and consistency, failing to meet the basic requirements for building a high-quality database.

[0005] To alleviate the cognitive overload of single-agent systems, multi-agent information extraction methods have emerged. These methods decompose the complex overall task and distribute it to multiple specialized sub-agents for parallel processing, theoretically allowing each agent greater focus. However, existing solutions typically employ completely independent parallel architectures, with each agent independently analyzing the full text and identifying entities within its own task, lacking explicit global coordination and constraint mechanisms. This leads to a new and more challenging problem: cross-task terminology inconsistency. For example, a materials composition agent identifies samples as one set of names, a preparation method agent identifies them as another, and an antibacterial testing agent reports data using a third set of names. Each agent uses completely different name identifiers for the same set of samples, rendering the extracted attribute data invalid due to the inability to automatically correlate them. Existing methods implicitly rely on the model's inherent stability for terminology consistency, but this cannot be guaranteed due to the lack of explicit constraints. In practical applications, materials science experts must perform time-consuming manual post-processing to unify terminology and correlate data, completely negating the efficiency advantages of automated extraction and severely hindering the large-scale application of this technical approach.

[0006] Therefore, how to design an effective global coordination mechanism in a multi-agent architecture, which can fully leverage the professional extraction capabilities of each agent while forcibly ensuring terminology consistency across tasks and chapters, thereby enabling structured data extraction that can be directly used for downstream analysis without human intervention, has become a core technical challenge that urgently needs to be overcome in the field of automated processing of materials science literature. Summary of the Invention

[0007] In view of the above-mentioned deficiencies in the existing technology, the purpose of this invention is to provide a method and system for extracting text information based on a multi-agent large language model. Specifically, the technical solution of this application is as follows: Firstly, this application provides a text information extraction method based on a multi-agent large language model, including the following steps: Based on the large language model of the main agent, extract the global named entity list from the target text information. The global named entity list includes at least the sample name list and the experimental method name list. Construct constraint features based on a global list of named entities, and construct prompt text for a large language model of at least two sub-agents based on the constraint features; All sub-agent large language models are run in parallel. Based on the constraint features in the prompt text corresponding to each sub-agent large language model, structured data fields are extracted from the target text information, and the names in the structured data fields are unified based on the global named entity list.

[0008] In some implementations, a global named entity list is extracted from the target text information based on the main agent's large language model, specifically including: The main intelligent agent's large language model can identify multiple different expressions of the same material sample and different expressions of the same experimental method in target text information; Based on preset priority rules, a standardized name is selected from multiple different expressions to form a sample name list and an experimental method name list; The priority rules include: first priority is the sample name marked in the predefined section; second priority is the description name containing chemometric information; and third priority is the descriptive name.

[0009] In some implementations, the structured data fields include supplementary additional fields; Add additional fields and corresponding named entities in the global named entity list to form a complete experimental description.

[0010] In some implementations, before extracting the list of globally named entities from the target text information based on the main agent's large language model, the following steps are also included: Based on a predefined list of keywords for non-experimental chapters, identify and filter content blocks of non-experimental chapters; The filtered and retained experimental chapter content blocks will be used as input for subsequent information extraction.

[0011] In some implementations, a global named entity list is extracted from the target text information based on the main agent's large language model, specifically including: Receive paragraph tag information based on document structure parsing and form chapter structure data for the target text information; Global variable extraction is performed based on chapter structure data to extract a list of global named entities from the target text information; Based on chapter structure data, each sub-agent's large language model is assigned text information fragments that are semantically relevant to its task.

[0012] In some implementations, the large language model of all sub-agents is run in parallel, specifically including: Receive constraint features containing a global list of named entities and the corresponding target text information; The extraction results of the current sub-agent are verified based on the predefined structured data. Based on constraint features, non-global named entities in the target text information are filtered out, and then mapped to their corresponding standardized names in the global named entity list through semantic understanding.

[0013] In some implementations, after unifying the names in the structured data fields based on the global named entity list, the following is also included: Based on the predefined output data patterns of each sub-agent's large language model, each agent's large language model is configured to generate structured output results that conform to its corresponding data pattern. The data pattern includes the name, data type, and constraints of each output field. The structured output results of each agent's large language model are verified using a verification procedure corresponding to the data pattern.

[0014] In some implementations, the main intelligent agent's large language model identifies chart information in the target text information, including chart number and title information; Associate each chart with its corresponding named entity; An image index is created that includes chart number, title information, named entity, and location identifier, and the image index is provided to the corresponding sub-agent large language model for data extraction.

[0015] In a second aspect, this application also provides a text information extraction system based on a multi-agent large language model, comprising: The main agent large language model is used to extract the global named entity list from the target text information, construct constraint features based on the global named entity list, and construct prompt text for at least two sub-agent large language models based on the constraint features. The global named entity list includes at least a sample name list and an experimental method name list. At least two sub-agent large language models are used to run in parallel and extract structured data fields from the target text information based on the constraint features in the corresponding prompt text, and unify the names in the structured data fields based on the global named entity list.

[0016] In a third aspect, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above methods.

[0017] This invention proposes an innovative method for automatically extracting information from materials science literature based on hierarchical multi-agent collaboration. By constructing a two-stage collaborative architecture combining global coordination by the main agent and parallel constraints by sub-agents, along with an explicit terminology constraint mechanism, it fundamentally solves the core problem of cross-task terminology inconsistency and reliance on manual post-processing in existing multi-agent methods. Compared with current multi-agent technologies that rely on single agents or lack coordination, this invention has the following outstanding advantages and beneficial effects: First, it solves the problem of cross-task terminology consistency, enabling directly usable extracted results and greatly improving the level of automation. The main agent first analyzes the full text and extracts a standardized global terminology list, which is then injected as a hard constraint into the prompts of each sub-agent, forcing all sub-tasks to use unified sample names and experimental method identifiers. This allows data scattered across different chapters, such as material composition, preparation processes, and performance testing, to be automatically linked and integrated based on the same identifiers. It completely eliminates the data silo problem caused by different agents using different names for the same sample. The extracted structured data can be directly used to build databases, train machine learning models, or construct knowledge graphs without manual correction, truly achieving a fully automated end-to-end processing flow.

[0018] Secondly, it significantly improves the accuracy and data integrity of multi-task extraction from complex and long documents, ensuring the data quality and value for downstream applications. On the one hand, through task decomposition and professional division of labor, it effectively avoids cognitive overload of single agents, reducing information omissions and erroneous associations. On the other hand, the unique paragraph awareness and hierarchical supplementation mechanism work synergistically. The main agent uses the document structure to accurately locate relevant information blocks to reduce noise interference, while each sub-agent supplements and extracts professional detail fields while adhering to global terminology, thus capturing experimental details completely while ensuring global consistency. Combined with a structured output verification mechanism based on a strict data model, it ensures the accuracy, structural integrity, and type correctness of the final data, laying a reliable data foundation for high-quality materials informatics analysis.

[0019] Third, it achieves efficient large-scale document batch processing, combining high system throughput with strong environmental adaptability, making it highly practical and easily scalable. Through a two-layer parallel processing mechanism combining sub-agent task-level parallelism and document batch-level parallelism, the overall system processing capacity is significantly improved, meeting the needs for large-scale information extraction from massive amounts of documents. Simultaneously, semantic understanding-based terminology standardization rules enable the system to adaptively handle differences in expression among different journals and authors, eliminating the need to write rules for specific formats. The preprocessing module automatically filters non-experimental chapters, further improving extraction accuracy while reducing computational costs. The system architecture is flexible, allowing for adjustments or additions of sub-task types according to actual needs, and exhibits strong scalability. Attached Figure Description

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

[0021] Figure 1 This is a flowchart of a text information extraction method based on a multi-agent large language model in one embodiment; Figure 2 This is another flowchart of a text information extraction method based on a multi-agent large language model in one embodiment; Figure 3 This is another flowchart of a text information extraction method based on a multi-agent large language model in one embodiment; Figure 4 This is a schematic diagram of the structure of a computer device in one embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] The method and system for automatic extraction of materials science literature information based on hierarchical multi-agent technology provided in this application can be widely applied to scenarios such as the construction of materials research databases, computational materials science-driven high-throughput screening, and cross-document knowledge discovery. The system interfaces with academic literature databases, institutional repositories, and open access platforms through standard data interfaces to automatically acquire research papers in the target field. A hierarchical multi-agent collaborative extraction engine is deployed on the data processing server to parse, extract structured data, and standardize the input literature, ultimately outputting experimental data in a unified format that can be directly used for downstream analysis.

[0024] Specifically, the materials science literature database serves as the primary data source, encompassing journal articles from major publishers such as Elsevier, Springer, and Wiley, as well as academic literature from preprint platforms like arXiv. The literature acquisition module periodically retrieves metadata and full-text documents from papers in the target field via API interfaces or web crawlers, transmitting them to a preprocessing server for format unification and parsing. The preprocessing server runs a document structure analysis program, performing deep analysis on PDF documents, identifying text blocks, table areas, and image content, and labeling each content block with its respective chapter tag, forming an intermediate representation with structural information. The core processing server runs a hierarchical collaborative engine composed of a main intelligent agent and multiple specialized sub-intelligent agents, completing the entire process from global terminology unification to fine-grained data extraction according to a two-level processing flow. A distributed database system is used to store the original literature, intermediate processing status, and the final output structured experimental data tables.

[0025] Furthermore, this system can be deeply integrated with materials computing platforms, laboratory information management systems, and knowledge graph construction tools. Extracted standardized data can be directly imported into first-principles calculation software and high-throughput computing frameworks via interfaces, serving as input parameters or validation benchmarks for materials design simulations. Simultaneously, structured data can be automatically imported into institutional knowledge bases or domain-specific databases, supporting knowledge graph construction and semantic retrieval based on SPARQL queries. When the system detects similar material systems with performance differences in multiple publications, it can automatically generate a knowledge discovery report, highlighting potential research opportunities or contradictory data points. In scientific research project management, the system can interface with enterprise R&D data platforms or national scientific research data centers to achieve automated archiving of experimental data and cross-project knowledge association.

[0026] The deployment architecture of this system is highly scalable, allowing for differentiated configurations for different application scenarios. For the construction of knowledge bases within research institutions, a local server cluster can be used for centralized processing; for large publishers or data service providers, a cloud-based distributed architecture can be adopted to achieve elastic resource scheduling and parallel pipeline processing; for scenarios requiring real-time processing, preprocessing and main agent analysis can be deployed on edge computing nodes, while sub-agent extraction tasks can be allocated to cloud computing resources, forming a hybrid computing architecture that meets the data processing needs of different levels, from individual researchers to scientific research infrastructure.

[0027] In the specific process of extracting text information, this application provides an embodiment, such as... Figure 1 As shown, a method for extracting text information based on a multi-agent large language model is provided, including the following steps: S100 extracts a list of globally named entities from target text information based on the main intelligent agent's large language model.

[0028] Specifically, the global named entity list includes at least a list of sample names and a list of experimental method names. Specifically, based on the target text information, the system first performs intelligent parsing and preprocessing on the original PDF document. Using a specialized document parsing tool, the system identifies all sections in the document and labels each text paragraph, table, and figure with its corresponding section title, such as Materials and Methods, Results, Discussion, etc.

[0029] After structural parsing, the system performs content filtering based on a pre-defined keyword list, removing non-experimental sections such as the abstract, introduction, references, and acknowledgments, retaining only the main text and figures containing core experimental data. This preprocessing step significantly reduces the amount of text processed subsequently and focuses on key information.

[0030] The preprocessed structured document, along with a carefully designed set of system instructions, was input into a large language model that served as the main agent. These instructions explicitly required the main agent to: first, browse the entire document, identify all the material samples studied, and assign a unique standardized name to each sample; second, identify all the experimental methods mentioned in the document and standardize their names.

[0031] During the analysis, the main agent performs semantic understanding and contextual association. For example, it identifies a sample named "Sample A" defined in the Materials section as a 2% silver-doped hydroxyapatite coating; in the Results section's charts, this sample is labeled as "two-silver-doped hydroxyapatite"; and in the Discussion section, it is referred to as a low-silver-content coating. The main agent recognizes that these different descriptions all refer to the same entity. Based on pre-defined standardization rules, it prioritizes the description containing precise stoichiometric information and explicitly defined for the first time in the text as the standard name, ultimately naming the sample uniformly "two-silver-doped hydroxyapatite." Regarding experimental methods, the agar diffusion method mentioned in the text is also standardized as the agar diffusion experiment.

[0032] Finally, the main agent outputs a structured global named entity list, which mainly contains two core parts: First, a standardized list of sample names, such as pure titanium control group, disilver-doped hydroxyapatite, and pentasilver-doped hydroxyapatite; Second, a standardized list of experimental method names, including preparation methods such as hydrothermal synthesis, characterization methods such as X-ray diffraction and scanning electron microscopy, and performance testing methods such as agar diffusion experiment and cytotoxicity test.

[0033] S200 constructs constraint features based on a global named entity list, and constructs prompt text for a large language model of at least two sub-agents based on the constraint features.

[0034] Specifically, the system prepares specialized instructions for subsequent parallel extraction tasks based on the standardized terminology generated by the main intelligent agent. The system pre-defines four specialized sub-tasks: material composition and preparation extraction, surface morphology and structure characterization extraction, antibacterial performance test data extraction, and biocompatibility assessment data extraction.

[0035] For each subtask, the system constructs a unique prompt text. Taking the antibacterial performance test data extraction subtask as an example, its prompt text contains several key parts. First, there are system-level instructions requiring the model to act as a materials science data extraction expert, ensuring that the output format is strictly standardized and that original text citations are included. Second, there are hard constraints, which are crucial to this invention. This part clearly lists the sample names that are allowed and required for this task in a prominent manner, namely, the pure titanium control group, disilver-doped hydroxyapatite, and pentasilver-doped hydroxyapatite defined by the main agent, as well as the names of the allowed antibacterial experimental methods, such as agar diffusion experiment and bacterial suspension counting method. The instructions explicitly require that even if the original text uses aliases such as sample A or low-silver coating, they must be mapped and replaced with the standard names. Finally, the prompt text will include a set of original text paragraphs selected by the main agent that are most relevant to the antibacterial test content, as specific material for the sub-agent analysis.

[0036] The S300 runs all sub-agent large language models in parallel. Based on the constraint features in the prompt text corresponding to each sub-agent large language model, it extracts structured data fields from the target text information and unifies the names in the structured data fields based on the global named entity list.

[0037] Specifically, the system simultaneously launches four sub-agent large language models, sending their respective customized prompt texts to them for parallel processing.

[0038] Upon receiving its prompt text, the antibacterial performance test data extraction sub-agent begins analyzing the relevant original text paragraphs. For example, upon recognizing that "Sample A has an inhibition zone diameter of 15.2 mm against E. coli," it immediately performs the following operations: First, it identifies that "Sample A" is a non-standard term used in the text. Then, guided by strict constraints, it utilizes its semantic understanding capabilities, combined with contextual information, to accurately map "Sample A" to the standard name "disilver-doped hydroxyapatite." Finally, it outputs the extraction results according to a predefined structured format, including fields such as standard sample name, antibacterial experimental method, test species, quantitative result value, and the corresponding original text sentence as a citation source. In the output, the value of the sample name field is strictly limited to one of pure titanium control group, disilver-doped hydroxyapatite, or pentasilver-doped hydroxyapatite. Simultaneously, other sub-agents are also working in parallel. The surface morphology and structure characterization extraction sub-agent, when extracting X-ray diffraction pattern analysis results, will also follow constraints, using the same standard sample name and outputting structured data such as phase composition and grain size. After all sub-agents have completed their extractions, the system collects their outputs. Since the identifiers for samples and methods in all outputs are strictly derived from the same globally named entity list generated by the main agent, data from different sub-tasks, such as the microstructure of the disilver-doped hydroxyapatite coating, its X-ray diffraction phase information, and its antibacterial performance data, can be automatically and accurately associated and integrated through the unique standard name "disilver-doped hydroxyapatite," forming a structured experimental data archive that can be directly used for the construction of materials databases or subsequent machine learning analysis.

[0039] In the above embodiments, the main agent first extracts global variables, including a list of sample names and a list of methods. Sub-agents then independently extract their respective task data under the constraints of these global variables, and finally, consistency is ensured through a verification layer. Unlike existing methods that allow each agent to extract data independently and in parallel before aligning it, this invention maintains hierarchical separation between global terminology standardization and fine-grained data extraction at the architectural level, avoiding the coordination overhead caused by direct communication between agents, and realizing a top-down control flow from global planning to local execution.

[0040] The global variable is passed through a prompt word concatenation method. The prompt word of the sub-agent is injected with "Sample name allowed by hard constraints: [list content], allowed methods: [list content]", and the global variable is passed directly as part of the context without the need for additional memory modules or knowledge bases.

[0041] Terminology unification relies on the semantic understanding capabilities of the main AI agent. The agent analyzes the full text context, mapping all sample descriptions appearing in the literature (such as "Ti-Ag", "Ti-5Ag", "5% Ag-doped Ti", "Sample A") to a unified standard name. The prompt explicitly requires: "Identify all material samples and determine a standard name for each sample, unifying all different expressions in the text to this standard name." This process relies on the cross-chapter reasoning capabilities of the large language model, eliminating the need for manually predefined mapping rules.

[0042] This invention employs a two-layer parallel mechanism to significantly improve processing efficiency: The first layer of parallelism is sub-agent task parallelism. Multiple sub-agents use Python's asyncio library to achieve true asynchronous concurrency. The `asyncio.gather` function simultaneously initiates LLM API calls for multiple sub-agents, ensuring that their extraction tasks are performed in parallel physical time, thus improving the extraction efficiency of a single document. The second layer of parallelism is document batch parallelism. When extracting multiple documents in batches, `ProcessPoolExecutor` is used to achieve multi-process parallel processing. The degree of parallelism is dynamically adjusted according to the API service provider's rate limit, typically set to four parallel processes, processing four documents simultaneously, significantly improving the overall throughput of large-scale document batch processing. By pre-unifying the terminology system through the main agent, terminology conflicts between sub-agents are fundamentally eliminated, and the extraction results can be directly used in downstream applications without manual post-processing.

[0043] In an exemplary embodiment, the main intelligent agent's large language model identifies multiple different expressions of the same material sample and different expressions of the same experimental method in the target text information. Based on preset priority rules, a standardized name is selected from multiple different expressions to form a sample name list and an experimental method name list; The priority rules include: first priority is the sample name marked in the predefined section; second priority is the description name containing chemometric information; and third priority is the descriptive name.

[0044] Specifically, the preprocessing module first parses the document, filtering out non-experimental sections and retaining the text containing core experimental content such as materials, methods, results, and discussion. The processed document, along with system instructions, is input into the main agent's large language model. The instructions require the main agent to complete a core task: to review the entire text, identify all material samples and experimental methods, and generate a unified standard name for each entity. During the analysis, the main agent identifies multiple expressions for the same material sample. For example, the sample code is explicitly defined in the materials section, the chemical name containing doping information is used in the results section, abbreviations are used in the figure and table titles, and descriptive terms appear in the discussion section. The main agent also finds that for the X-ray diffraction experimental method, there are multiple expressions in the text, including the full name, standard abbreviation, and shorthand. Based on a pre-defined three-level priority rule, the main agent processes all expressions uniformly: First priority: Use the formal sample code or full method name explicitly marked by the author in predefined sections such as materials and methods. Second priority: If not explicitly marked, use the expression containing complete stoichiometric information (such as elements, concentrations, and proportions) as the standard name. Third priority: When neither of the above applies, a descriptive name or a generic designation is used. After applying this rule, the main agent unifies various variations in the text into standard names. For example, the same group of samples is unified as BT0, BT1N, and BT2N; various descriptions of X-ray diffraction are unified as XRD. Finally, the main agent outputs a structured global named entity list. This process relies on the semantic understanding capabilities of a large language model for professional texts, rather than pre-defined hard-coded rules, thus achieving adaptive standardization of complex and diverse expressions in scientific literature.

[0045] In the above embodiments, when the main agent identifies multiple names for the same sample, a three-level priority rule is used to determine the standard name: the first priority is the expression explicitly marked as "samplename" or "sample code" in the Materials or Methods section, which is usually the formal name defined by the authors; the second priority is the expression containing explicit stoichiometric information (such as "Ti-5Ag", "Ti with 5 at% Ag"), which is complete and scientifically standardized; the third priority is descriptive names (such as "silver-modified titanium") or codes (such as "Sample A"), which are incomplete. When multiple names of the same priority exist, the name that appears first is selected as the standard name. When the main agent extracts global variables, it records various alias expressions simultaneously. These mapping relationships are passed to the sub-agents through prompt words to help the sub-agents map correctly.

[0046] The sub-agent maps non-standard terms in the literature to standard names using a synergistic effect of two mechanisms. The first is the cue word engineering mechanism: the sub-agent's cue words not only list allowed standard names but also provide mapping hints, such as "**GLOBAL CONTEXT:** Allowed sample names: ['cp-Ti','Ti-5Ag','Ti-10Ag']. **IMPORTANT NOTES:** Sample names MUST EXACTLY match the global `sample_groups` list. If the paper uses alternative names (e.g., 'Sample A', '5 wt% Agcoating'), you MUST map them to the standard names from the global context." This cue word design explicitly tells the large language model: which are standard names; non-standard expressions must be mapped to standard names; and mapping relies on semantic understanding.

[0047] Pydantic type validation mechanism: Data structures are defined using Pydantic BaseModel, which automatically validates required fields and data types. Although the actual implementation does not use Literal type constraints on the value range of the sample_name field, Pydantic still ensures that: the sample_name field must exist and be of string type; the output structure conforms to the predefined schema; and a ValidationError is thrown when a required field is missing.

[0048] The mapping relationships are not pre-coded rule tables, but are dynamically established through the semantic understanding of the main agent. The main agent understands the referential and stoichiometric relationships between various expressions while analyzing the full text, and determines standard names when extracting global variables. The sub-agent's large language model utilizes cue word constraints and its own semantic understanding capabilities to complete the mapping from non-standard terms to standard names. Pydantic type validation ensures the correctness of the output structure.

[0049] The method of this invention involves the main agent pre-extracting a standard terminology list. Sub-agents must adhere to these standard terms during extraction, achieving pre-constraint through cue word constraints and Pydantic verification. This follows a paradigm of constraint first, extraction later. This invention is preventative, relies on the semantic understanding of a large language model, and is applicable to the field of scientific literature extraction.

[0050] In one exemplary embodiment, in some implementations, the structured data fields include supplementary additional fields; the supplementary additional fields, together with the corresponding named entities in the global named entity list, constitute a complete experimental description.

[0051] Specifically, this invention standardizes both sample names and experimental methods, and employs a hierarchical supplementation mechanism to ensure both the uniformity of global terminology and the provision of rich detailed information.

[0052] Standardization is implemented at the main agent level: For sample names, specific naming conventions must be followed, such as "cp-Ti" (pure titanium control group), "1 nm Ag-coated MAO" (coated sample with thickness information), and "0.5 wt% Ag-TiO2" (doped sample with concentration information). Variations of different parameters (such as different thicknesses, concentrations, and processing times) must be listed as independent sample groups. For method names, abbreviations or standard terms should be used, such as "MAO" (instead of "Micro arcoxidation"), "XRD" (instead of "X-ray diffraction"), and "Agar diffusion test" (instead of "Agardiffusion method"). The standardization selection rule is: if there is already a clearly defined standardized name in the literature (such as the sample code defined in the Materials section), then the literature's name is adopted; otherwise, the large language model automatically generates a standard name based on the content and naming conventions.

[0053] At the sub-agent level, additional fields are added during fine-grained extraction to further differentiate and correlate information. For example, in the material sample task, the `dopant_elements` field (list of dopant elements) and the `description` field (description of key sample features) are added, allowing different samples to be distinguished even if the standard names of the main agent are similar. In the antibacterial testing task, the `experimental_model` field (list of test bacteria), the `evaluation_metric` field (evaluation metric), the `text_result` field (array of qualitative results descriptions), and the `tool_result` field (array of precise numerical values ​​extracted from charts) are added. These additional fields, along with the method names extracted by the main agent, form a complete experimental description. This hierarchical supplementation mechanism avoids information loss, ensuring both the consistency of global terminology and the capture of fine-grained experimental details.

[0054] In an exemplary embodiment, before extracting the list of globally named entities from the target text information based on the main agent's large language model, the method further includes: Based on a predefined list of keywords for non-experimental chapters, content blocks of non-experimental chapters are identified and filtered; the content blocks of experimental chapters that are retained after filtering are used as input for subsequent information extraction.

[0055] Specifically, after the preprocessing module starts, it first calls a specialized PDF document parsing tool to perform deep parsing of the original document. This parsing tool can not only extract text content, but also identify the document's layout structure and logical hierarchy. It automatically labels each block with its corresponding chapter information, recording it in the paragraph field. For example, a text block describing experimental methods might have its paragraph field labeled "2. Materials and Methods"; a figure block displaying results might have its paragraph field labeled "Fig. 3. Corrosion rate". The system then performs the core filtering operation. The system maintains a predefined list of non-experimental chapter keywords, which includes chapter names that typically do not contain core experimental data, such as "Abstract", "Introduction", "References", "Acknowledgements", "Funding", and "Supplementary Material". The filtering algorithm traverses all parsed content blocks, checking the paragraph field value of each block. Through case-insensitive substring matching, if a block's paragraph field contains any keyword from the list, the system determines that the block belongs to a non-experimental chapter and removes it from the set of document blocks to be processed.

[0056] Furthermore, the filtering strategy also includes location-based logic. The system identifies the starting positions of the "Introduction" and "References" sections. All content blocks preceding the "Introduction" section, such as the cover, title, and author list, are removed regardless of their paragraph field. Similarly, all content blocks following the "References" section, such as appendices, author biographies, and copyright notices, are also filtered out. This process effectively removes a large amount of text irrelevant to the core experiments. This not only significantly reduces the cost of calling the large language model API, but more importantly, it removes noisy information such as background information in the abstract, literature reviews in the introduction, and reference lists. This allows the main agent and sub-agents to focus their computations on the core experimental sections containing material formulations, preparation processes, characterization data, and performance results, such as "Materials and Methods," "Results," "Discussion," and "Characterization," thereby greatly improving the accuracy and reliability of information extraction. Finally, the filtered and purified set of experimental chapter content blocks is assembled into a structured intermediate representation, which serves as the direct input for the main agent to extract the global named entity list.

[0057] The preprocessing module uses a blacklist design to filter non-experimental sections, reducing input noise and computational cost for large language models. Filtering rules include: blacklisted sections (Abstract, Introduction, References, Acknowledgements, Funding, Supplementary Material, etc., which do not contain experimental data); case-insensitive substring matching (by checking the paragraph field of a block, if it contains blacklisted keywords (ignoring case), the block is removed); and positional filtering (removing preceding content before Abstract / Introduction, such as the cover and author information, and all content after References, such as appendices and copyright notices).

[0058] Actual testing shows that preprocessing filters an average of 40-50% of document blocks, reducing the average input length from approximately 14.5k tokens to approximately 8k tokens. This not only reduces the cost of LLM API calls, but more importantly, it reduces the interference of irrelevant information on extraction accuracy. The reserved sections (Methods, Results, Discussion, Characterization, Antibacterial Activity, etc.) concentrate key experimental data, enabling the large language model to focus more on extracting relevant information.

[0059] In one exemplary embodiment, such as Figure 2 As shown, step S100 extracts a list of globally named entities from the target text information based on the main agent's large language model, specifically including: S110 receives paragraph tag information based on document structure parsing and forms chapter structure data for the target text information.

[0060] S120 performs global variable extraction based on chapter structure data, extracting a list of global named entities from the target text information.

[0061] S130 assigns text information fragments related to the semantics of the task to each sub-agent's large language model based on chapter structure data.

[0062] Based on the above embodiments, the main intelligent agent adopts a three-layer structure design, namely a paragraph perception layer, a global variable extraction layer, and a task block selection layer, making full use of the document structure information pre-annotated by the PDF parsing tool.

[0063] The first layer (paragraph awareness layer) is responsible for using the pre-annotated paragraph information from the PDF parsing tool. When parsing a PDF, MineRU automatically annotates each document block (text block, table block, image block, etc.) with its corresponding section title (paragraph field), such as "Materials and Methods", "Results and Discussion", "Antibacterial Activity", etc. The main agent receives this pre-annotated paragraph information from the prompt words to form an overview of the section structure. Additional section and block summary information is added to the prompt words, such as: "DOCUMENT STRUCTURE OVERVIEW: The document is organized into the following paragraphs / sections: Introduction (3 blocks), Materials and Methods (12 blocks), Sample Preparation (8 blocks), Antibacterial Activity (15 blocks), Results and Discussion (20 blocks), Conclusions (2 blocks)".

[0064] The second layer (global variable extraction layer) guides the large language model to extract global variables based on paragraph-aware information. Paragraph-Guided Extraction Hints are injected into the prompt words, for example: "Paragraphs with titles containing 'Materials' or 'Samples' usually contain sample name information; paragraphs with titles containing 'Antibacterial' usually contain antibacterial method information." The main agent analyzes all document blocks in the full text and performs three sub-tasks: determining the research scope (is_titanium_base, is_surface_modified, is_doping_performed); extracting and standardizing global variables, including a list of sample names (sample_groups) and a list of five types of experimental methods (preparation_methods, surface_characterization, antibacterial_methods, biocompatibility_methods, ion_release_test_methods); and building an image index, mapping figure numbers (such as "Fig.1", "Table 2") to their titles and experimental method types.

[0065] The reason why large language models can achieve terminology standardization lies in their powerful semantic understanding and cross-textual association capabilities. When analyzing literature, the model can identify that "Ti-Ag" is defined in the Materials section, "5% Ag-doped Ti" describes the preparation process in the Methods section, "Sample A" reports data in the Results section, and "Ti-5Ag" is labeled in figures—all referring to the same material. By understanding the contextual referential relationships, stoichiometry (5 wt% Ag is approximately equal to 5 at% Ag), and the narrative logic of the literature, the model unifies these different expressions into a standard name. This is the core advantage of large language models over smaller models: smaller models lack sufficient semantic understanding and cross-textual reasoning capabilities, making them unable to complete such complex terminology mapping tasks. Furthermore, large language models can generate standard names that conform to the naming conventions in the prompts.

[0066] The third layer (task block selection layer) selects relevant document blocks for each of the six sub-agents based on a paragraph-aware strategy. The main agent identifies a list of block_ids containing relevant information for each sub-task based on the semantic relevance of the paragraph title to the task. For example: the material sample task selects blocks in the paragraph containing "Materials," "Samples," and "Preparation"; the antibacterial test task selects blocks in the paragraph containing "Antibacterial," "Antimicrobial," and "Results"; and the surface characterization task selects blocks in the paragraph containing "Characterization," "XPS," and "SEM."

[0067] Unlike hard-coded chapter-task mapping rules, paragraph-aware strategies allow large language models to adaptively identify information scattered across chapters and handle format differences between different journals. For example, some journals distribute antimicrobial data across multiple locations such as "Results," "Discussion," and figures, and the main agent can flexibly identify all relevant locations.

[0068] In one exemplary embodiment, such as Figure 3 As shown, step S300 involves running the large language model of all sub-agents in parallel, specifically including: S310 receives constraint features containing a global list of named entities and the corresponding target text information.

[0069] S320 verifies the extraction results of the current sub-agent based on the predefined structured data.

[0070] S330 filters non-global named entities in the target text information based on constraint features and maps them to the corresponding standardized names in the global named entity list through semantic understanding.

[0071] Based on the above embodiments, the six specialized sub-intelligent agents (material sample, preparation method, surface characterization, antibacterial test, biocompatibility test, and ion release test) adopt a three-layer structure design, such as... Figure 1 These consist of a constraint injection layer, a structured validation layer, and a semantic mapping layer, respectively, which ensure terminology consistency through explicit constraint mechanisms.

[0072] The first layer (constraint injection layer): The sub-agent receives three types of input: global variables (lists of sample_groups and methods), task-related document blocks (blocks corresponding to the block_id list provided by the main agent's task block selection layer), and image indexes (used to associate with chart data). The hard constraints are explicitly injected at the beginning of the prompt: "[CRITICAL REQUIREMENTS] 1. Use sample names EXACTLY from the global context (do not invent new names) 2. Use method names EXACTLY from the global context. **GLOBAL CONTEXT:** Allowed sample names: [list of sample_groups extracted by the main agent], Allowed [task type] methods: [list of corresponding methods extracted by the main agent]. If the paper uses alternative names (e.g., 'Sample A'), you MUST map them to the standard names from the global context."

[0073] Although the six sub-agents share the same structure, their specific content differs across multiple dimensions: System prompts are shared by all sub-agents, defining common extraction requirements (e.g., precise numerical extraction, mandatory inclusion of the `source` field); Task-specific prompts differ for each sub-agent, detailing the fields to be extracted, their meanings, extraction strategies, and examples. For instance, the material sample task emphasizes the extraction strategy for the `dopant_elements` field, while the antibacterial testing task specifies the extraction requirements for fields like `is_quantitative` and `evaluation_metric`; The global constraint context receives different lists of allowed methods for each sub-agent; for example, the antibacterial testing agent receives the `antibacterial_methods` list, while the preparation method agent receives the `preparation_methods` list; Task-related blocks differ for each sub-agent, allocated by the main agent's task block selection layer; and the Pydantic data model uses a different `BaseModel` to define its output structure for each sub-agent.

[0074] This type of explicitly constrained sub-agent differs from a general sub-agent: a general sub-agent independently identifies sample names and methods, with no constraints in the prompts, and its output may be inconsistent with other agents; an explicitly constrained sub-agent, on the other hand, explicitly lists allowed samples and methods in its prompts, requiring the output to use these standard names to ensure consistency with other agents. Existing methods employ implicit dependency models, hoping that a large language model will "automatically" maintain terminology consistency across tasks, but without providing any explicit constraint mechanisms (such as lists of allowed values, type checks, etc.). This implicit dependency is unreliable in practical applications because the model may adopt different naming styles in different tasks.

[0075] The second layer (structured validation layer) uses the Pydantic BaseModel to define the output data pattern. Each sub-agent automatically performs type checking and required field validation using a task-specific Pydantic model. Taking the antibacterial testing task as an example, the actual Pydantic model is defined as follows: class AntibacterialTest(BaseModel): antibacterial_method: str # Required, method name method_procedure: str # Required, description of the method steps is_quantitative: str # Required, "yes" or "no" evaluation_metric: Optional[str] = None # Optional, evaluation metric metric_antibacterial_relationship: Optional[str] = None # Optional, the relationship between the metric and antibacterial activity. evaluation_formula: Optional[str] = None # Optional, calculation formula experimental_model: List[str] = Field(default_factory=list) # List of test strains best_sample: List[str] = Field(default_factory=list) # List of best sample names text_result: List[str] = Field(default_factory=list) # Qualitative result description array tool_result: Optional[List[str]] = Field(default=None) # An array of precise values ​​extracted from the chart figure_id: Optional[str] = None # Related chart number class AntibacterialTestResult(BaseModel): antibacterial_test: List[AntibacterialTest] = Field(default_factory=list) The large language model outputs structured data in JSON format. The Pydantic framework automatically validates: whether required fields are complete (antibacterial_method, method_procedure, is_quantitative); whether the data type is correct (experimental_model must be a list); and whether field values ​​meet the constraints (is_quantitative can only be "yes" or "no"). If constraints are violated, Pydantic throws a ValidationError, and the extracted result is rejected.

[0076] It should be noted that in the actual implementation, the Pydantic model does not use Literal type dynamic constraints on sample names and method names. Instead, consistency is ensured through cue word constraints and manual review. Pydantic is primarily used to verify the integrity of data structures and the correctness of types.

[0077] The third layer (semantic mapping layer): The large language model understands various terminological expressions in the literature and maps them to standard names defined by the main agent. For example, "Sample B" in the text is mapped to "Ti-5Ag" through cross-chapter reasoning (finding the definition of Sample B in the Materials chapter); "5 wt% Ag coating" in the text is mapped to "Ti-5Ag" by understanding the stoichiometric relationship; and "Silver-modified titanium" in the text is mapped to "Ti-5Ag" through contextual semantics (concentration information mentioned in the preceding and following sentences).

[0078] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0079] In one exemplary embodiment, after the name in the unified structured data field based on the globally named entity list, the method further includes: Based on the predefined output data patterns of each sub-agent large language model, each agent large language model is configured to generate structured output results that conform to its corresponding data pattern. The data pattern includes the name, data type, and constraints of each output field. A verification program corresponding to the data pattern is used to verify the structured output results output by each agent large language model.

[0080] Based on the above embodiments, the verification module uses Pydantic BaseModel to define a strict output data pattern, implements two-layer verification, and ensures that all data entering downstream applications is in the correct format and of the same type.

[0081] The first layer of validation (model definition layer) defines data classes in Python code, declaring constraints such as the type, required status, and value range of each field. For example: class GlobalVariables(BaseModel): is_titanium_base: bool (must be a boolean value) sample_groups: List[str] (must be a list of strings) antibacterial_methods: List[str] biocompatibility_methods: List[str] preparation_methods: List[str] surface_characterization: List[str] ion_release_test_methods: List[str] is_doping_performed: bool The large language model is configured in JSON mode (response_format={"type":"json_object"}), directly generating structured output that conforms to the Pydantic schema, rather than free text.

[0082] The second layer of validation (runtime validation layer) automatically validates the data after the large language model returns JSON results using Pydantic's `model_validate_json` method: Field type checks (`sample_groups` must be a list, not a single string); required field checks (the `is_titanium_base` field cannot be missing); and data format checks (list fields must be list types). If any constraint is violated, Pydantic throws a `ValidationError`, the extracted results are rejected, and the system can choose to retry or mark the extraction as failed.

[0083] Unlike traditional string parsing methods (which use regular expressions to extract information from free text and then use post-processing scripts to clean the data), Pydantic declares constraints at the model definition level and automates the verification process, eliminating the need to write complex parsing and validation code.

[0084] Data traceability is achieved through the task block selection layer (task_block_references) and the source field. The main agent's output includes task_block_references, which records a list of block_ids that each subtask should process, for example: {"task_block_references": {"material_samples_related_block_ids": ["block_12","block_13","block_15"], "antibacterial_test_related_block_ids":["block_89","block_90","block_95"], ...} The sub-agent is required to provide a `source` field in the prompt, which records the original text citations that support the extracted results, for example: {"material_samples": [{"sample_group_name":"Ti-5Ag", "description":"Silver-doped titanium substrate", "dopant_elements": ["Ag"], "source": ["Ti-5Ag samples were prepared by magnetron sputtering with5 at% silver target.", "EDS analysis confirmed 5.2 wt% Ag content in the coating." ]}]}.

[0085] The `source` field contains text paragraphs copied verbatim from the document, allowing for verification of extraction accuracy during manual review. The combination of `task_block_references` and the `source` field enables a complete traceability chain from the block level to the sentence level.

[0086] In an exemplary embodiment, the main agent large language model identifies chart information in the target text information, the chart information including chart number and title information; associates each chart information with the corresponding named entity; establishes an image index containing chart number, title information, named entity and location identifier, and provides the image index to the corresponding sub-agent large language model for data extraction.

[0087] The main agent builds an image index to provide sub-agents with a mapping relationship between charts and experimental methods, supporting multimodal information extraction. The image index building process is as follows: the main agent analyzes all image and table type blocks; extracts chart numbers (such as "Figure 1", "Fig. 2", "Table 3") and titles (caption field) from the block's metadata; and determines the type of experimental method the chart belongs to based on the title content (titles containing keywords such as "SEM", "XRD", "morphology" are classified as surface_characterization; titles containing keywords such as "antibacterial", "inhibition", "bacterial" are classified as antibacterial; and titles containing keywords such as "cell viability" and "cytotoxicity" are classified as biocompatibility).

[0088] The actual output example is as follows: {"image_index": {"Fig.1": {"caption":"SEM images of Ti-Ag coatings at different magnifications","method_type":"surface_characterization","block_id":"image_block_23"}, "Fig.3": {"caption":"Antibacterial rates of different samples against E. coli and S. aureus","method_type":"antibacterial","block_id":"image_block_45"}, "Table 2": {"caption":"XPS elemental composition analysis results","method_type":"surface_characterization","block_id":"table_block_18"} }}.

[0089] The sub-agent uses image indexing in the following ways: after receiving the image index, it identifies which charts are relevant to its task; if a large language model that supports vision (such as GPT-4o, Gemini) is used, the sub-agent will add the base64 encoding of the relevant images to the prompt and extract data directly from the images; if a plain text model is used, the sub-agent relies on the chart caption and related text description for extraction.

[0090] The correspondence between the five categories of experimental methods and the six sub-tasks is as follows: the material sample task uses information on preparation and characterization methods; the preparation method task focuses on detailed parameters of the preparation method; the surface characterization task focuses on detailed data of the characterization method; the antibacterial testing task focuses on detailed data of the antibacterial method; the biocompatibility testing task focuses on detailed data of the biocompatibility method; and the ion release testing task focuses on detailed data of the ion release method. The five categories of methods represent a classification of all experimental methods in the literature, while the six sub-tasks represent a classification of the extraction targets. After the main agent extracts the list of five categories of methods, each sub-agent receives a list of methods related to its task as constraints and extracts only the data from these methods.

[0091] Based on the same inventive concept, this application also provides a text information extraction system based on a multi-agent large language model for implementing the text information extraction method based on a multi-agent large language model described above. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the text information extraction system based on a multi-agent large language model provided below can be found in the limitations of the text information extraction method based on a multi-agent large language model described above, and will not be repeated here.

[0092] In an exemplary embodiment, a text information extraction system based on a multi-agent large language model is provided, comprising: a main agent large language model for extracting a global named entity list from target text information, constructing constraint features based on the global named entity list, and constructing prompt texts for at least two sub-agent large language models based on the constraint features, wherein the global named entity list includes at least a sample name list and an experimental method name list; and at least two sub-agent large language models for running in parallel and extracting structured data fields from the target text information based on the constraint features in the corresponding prompt texts, and unifying the names in the structured data fields based on the global named entity list.

[0093] The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or they can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0094] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores information such as raw audio, extracted phoneme sequences, and recognition results for interrogative tone recognition. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a dynamic battery power protection method for a photovoltaic cleaning robot.

[0095] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: extracting a global named entity list from target text information based on a large language model of a main intelligent agent. The global named entity list includes at least a sample name list and an experimental method name list. Construct constraint features based on a global list of named entities, and construct prompt text for a large language model of at least two sub-agents based on the constraint features; All sub-agent large language models are run in parallel. Based on the constraint features in the prompt text corresponding to each sub-agent large language model, structured data fields are extracted from the target text information, and the names in the structured data fields are unified based on the global named entity list.

[0096] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the following steps: extracting a global named entity list from target text information based on a large language model of a main intelligent agent, wherein the global named entity list includes at least a sample name list and an experimental method name list; Construct constraint features based on a global list of named entities, and construct prompt text for a large language model of at least two sub-agents based on the constraint features; All sub-agent large language models are run in parallel. Based on the constraint features in the prompt text corresponding to each sub-agent large language model, structured data fields are extracted from the target text information, and the names in the structured data fields are unified based on the global named entity list.

[0097] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for extracting text information based on a multi-agent large language model, characterized in that, Including the following steps: Based on the large language model of the main agent, a global named entity list is extracted from the target text information. The global named entity list includes at least a sample name list and an experimental method name list. Based on the global named entity list, constraint features are constructed, and prompt text for at least two sub-agent large language models is constructed based on the constraint features; All the sub-agent large language models are run in parallel. Based on the constraint features in the prompt text corresponding to each sub-agent large language model, structured data fields are extracted from the target text information, and the names in the structured data fields are unified based on the global named entity list.

2. The text information extraction method according to claim 1, characterized in that, The extraction of the global named entity list from the target text information based on the main agent's large language model specifically includes: The main intelligent agent's large language model identifies multiple different expressions of the same material sample and different expressions of the same experimental method in the target text information; According to a preset priority rule, one of the various different expressions is selected as a standardized name to form the sample name list and the experimental method name list; The priority rules include a first priority for sample names marked in predefined sections, a second priority for descriptive names containing chemometric information, and a third priority for descriptive names.

3. The text information extraction method according to claim 1, characterized in that, The structured data fields include supplementary additional fields; The additional fields, together with the corresponding named entities in the global named entity list, constitute a complete experimental description.

4. The text information extraction method according to claim 1, characterized in that, Before extracting the global named entity list from the target text information based on the main agent's large language model, the method further includes: Based on a predefined list of keywords for non-experimental chapters, identify and filter content blocks of non-experimental chapters; The filtered and retained experimental chapter content blocks will be used as input for subsequent information extraction.

5. The text information extraction method according to claim 1, characterized in that, The extraction of the global named entity list from the target text information based on the main agent's large language model specifically includes: Receive paragraph tag information based on document structure parsing to form chapter structure data for the target text information; Based on the chapter structure data, global variable extraction is performed to extract the list of global named entities from the target text information; Based on the chapter structure data, each sub-agent's large language model is assigned a text information fragment related to its task semantics.

6. The text information extraction method according to claim 1, characterized in that, The parallel execution of all the sub-agent large language models specifically includes: Receive the constraint features containing the global named entity list and the corresponding target text information; Based on the predefined structured data of the sub-agent, the extraction results are verified. Based on the constraint features, non-global named entities in the target text information are filtered and mapped to their corresponding standardized names in the global named entity list through semantic understanding.

7. The text information extraction method according to any one of claims 1 to 6, characterized in that, After unifying the names in the structured data fields based on the global named entity list, the method further includes: Based on the predefined output data patterns of each sub-agent's large language model, each agent's large language model is configured to generate structured output results that conform to its corresponding data pattern. The data pattern includes the name, data type, and constraints of each output field. The structured output results of each agent's large language model are verified using a verification procedure corresponding to the data pattern.

8. The text information extraction method according to any one of claims 1 to 6, characterized in that, The main intelligent agent's large language model identifies the chart information in the target text information, and the chart information includes chart number and title information; Associate each of the chart information with its corresponding named entity; An image index is created that includes the chart number, the title information, the named entity, and the location identifier, and the image index is provided to the corresponding sub-agent large language model for data extraction.

9. A text information extraction system based on a multi-agent large language model, characterized in that, include: A main agent large language model is used to extract a global named entity list from the target text information, construct constraint features based on the global named entity list, and construct prompt text for at least two sub-agent large language models based on the constraint features. The global named entity list includes at least a sample name list and an experimental method name list. At least two sub-agent large language models are used to run in parallel and extract structured data fields from the target text information according to the constraint features in the corresponding prompt text, and unify the names in the structured data fields based on the global named entity list.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images