A large model-based thesis knowledge graph construction method, device, system and storage medium

Abstract

This invention discloses a method, apparatus, system, and storage medium for constructing a knowledge graph of academic papers based on a large-scale model. The method for constructing a knowledge graph of academic papers based on a large-scale model includes the following steps: S11, acquiring original paper data and preset extraction rules; S12, preprocessing the original paper data to obtain structured target text; S13, based on the extraction rules, using a large-scale speech model to extract entities from the target text to obtain an entity list; S14, using the large-scale speech model to extract the relationship between any two entities in the entity list to obtain a relationship list; S15, aligning, deduplicating, and detecting conflicts among knowledge from different data sources; S16, based on the entity list and the relationship list, fusing multi-source knowledge to construct a knowledge graph. This invention can efficiently extract structured knowledge from massive amounts of academic papers and reveal the evolution and cross-relationships of disciplines.

Description

Technical Field

[0001] This invention relates to the fields of computer software security and artificial intelligence technology, and more specifically, to a method and system for constructing a paper knowledge graph based on a large model. Background Technology

[0002] Knowledge graphs (KG), as a structured form of knowledge representation, use "entity-relationship-entity" triples as basic units. They can effectively organize and manage massive amounts of information, providing fundamental support for applications such as intelligent question answering, recommendation systems, and semantic search. In the field of academic literature, constructing knowledge graphs for academic papers aims to structurally organize core elements such as research methods, datasets, experimental conclusions, and academic institutions scattered throughout massive amounts of academic literature. This reveals the evolution and cross-relationships of disciplines, helping researchers quickly grasp the overall picture of the field and discover potential innovative points. However, despite the enormous potential of knowledge graphs in theoretical research and practical applications, existing technologies for constructing knowledge graphs for academic papers still face the following technical bottlenecks and challenges: First, there is insufficient utilization of domain knowledge systems. While general-purpose large models demonstrate strong semantic understanding and generation capabilities in natural language processing tasks, they still have significant shortcomings in integrating scientific domain knowledge. Existing methods often fail to effectively integrate existing knowledge systems such as subject classification systems, domain thesauri, and professional ontology libraries, resulting in a disconnect between the extracted entities and relations and the subject's cognitive structure. For example, when extracting the entity "attention mechanism," general-purpose models may not be able to accurately distinguish its different connotations and application scenarios in computer vision and natural language processing. This semantic bias not only affects the accuracy and completeness of the knowledge graph but also limits its practical value in advanced application scenarios such as precise retrieval and subject analysis. Furthermore, due to the lack of effective guidance from domain knowledge, existing methods struggle to identify subject-specific entity types and relational patterns, making it difficult for the constructed knowledge graph to fully reflect the subject's internal logic and development trajectory.

[0003] Second, the integration of multi-source heterogeneous knowledge is challenging. The construction of a paper knowledge graph involves multi-source heterogeneous data, including full-text papers, citation data, author information, and journal metrics. These data differ significantly in format, standards, and semantic levels. For example, full-text papers are primarily unstructured text, citation data has a complex network structure, and author information and journal metrics contain a large amount of attribute information. Existing technologies have significant shortcomings in the collaborative integration of multi-source data, often facing problems such as data conflicts, redundancy, and inconsistencies in concept levels. For instance, different data sources may use different names or identifiers for the same scholar or institution, leading to difficulties in entity alignment; the description of the same research conclusion in different papers may have semantic discrepancies, causing relational conflicts. Furthermore, existing methods have limited ability to mine deep semantic relationships (such as causal logic and temporal evolution), resulting in relatively shallow knowledge graph relationships and a lack of in-depth support for scientific research decision-making. This limitation severely restricts the application potential of knowledge graphs in revealing the laws of disciplinary development and assisting scientific research innovation.

[0004] Third, the construction process suffers from low automation and insufficient modularity. Existing knowledge graph construction systems typically employ a tightly coupled design, with strong dependencies between functional modules, making flexible expansion and optimization difficult. For example, entity extraction and relation extraction often rely on specific rules or models; once the domain or task changes, the entire system needs to be redesigned and adjusted, resulting in high development and maintenance costs. Simultaneously, existing systems exhibit significant deficiencies in time-dimensional information modeling, lacking effective support for dynamic attributes such as entity and relation expiration dates. For instance, new methods and data in the research field are often time-sensitive, and existing knowledge graphs struggle to accurately reflect the generation, evolution, and expiration of knowledge. This static modeling approach not only limits the timeliness of the knowledge graph but also affects its application in dynamic research environments. Furthermore, existing systems have weak capabilities for processing incremental data, making it difficult to support automatic identification and knowledge updates for newly added papers, resulting in high maintenance costs and delayed updates.

[0005] In summary, existing knowledge graph construction technologies still have many shortcomings in areas such as domain knowledge fusion, multi-source data integration, system automation, and modularization, making it difficult to meet the growing demands for scientific research knowledge management. Therefore, there is an urgent need for a construction method and system that can deeply integrate domain knowledge, efficiently integrate multi-source information, support dynamic updates, and possess high scalability, in order to enhance the practical value and application breadth of knowledge graphs in the scientific research field. Summary of the Invention

[0006] This invention overcomes the shortcomings of existing technologies and proposes a method, device, system and storage medium for constructing a paper knowledge graph based on a large model. It can efficiently extract structured knowledge from massive papers and reveal the evolution and cross-relationships of disciplines.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for constructing a paper knowledge graph based on a large model, comprising the following steps: Step S11: Data and extraction rules acquisition.

[0008] Obtain original paper data, which includes at least one of the following: full text of the paper, abstract, citation data, author information, and journal indicators; obtain preset extraction rules, which include entity type definition, relation type definition, and validity period rule definition.

[0009] Step S12: Data cleaning.

[0010] The original paper data is preprocessed to obtain structured target text; the preprocessing includes at least one of text cleaning, paragraph segmentation, sentence boundary detection, and terminology normalization.

[0011] Step S13: Entity extraction.

[0012] Based on the extraction rules, a large language model is used to extract entities from the target text to obtain an entity list. The entity list contains entity information for each entity, including entity name, entity type, validity period for each type, and a list of entity attributes.

[0013] Entity extraction further includes: A prompt template for embedding a scientific knowledge system into a large model, wherein the knowledge system includes a subject classification system, a domain thesaurus, and a professional ontology; A multi-granularity text analysis strategy is adopted, with different extraction granularities configured for the title, abstract, and full text. The extracted entities are subjected to conflict detection and normalization.

[0014] Step S14: Relation extraction.

[0015] A large language model is used to extract the relationship between any two entities in the entity list to obtain a relationship list; the relationship list contains relationship information for each relationship, including the names of the two corresponding entities, the validity period of the relationship, and the relationship description.

[0016] Relation extraction further includes: Design relationship extraction prompt templates to guide large models in recognizing semantic relationships between entities; The chain-of-thought technique is used to guide and verify the relational reasoning process of large models; Joint extraction of complex relationships across paragraphs and documents.

[0017] Step S15: Knowledge integration and conflict resolution.

[0018] Alignment, deduplication, and conflict detection are performed on knowledge from different data sources, and consistency verification is carried out using the reasoning capabilities of large models to ensure the integrity and accuracy of the knowledge graph.

[0019] Step S16: Map construction and dynamic update.

[0020] Based on the entity list and the relationship list, a knowledge graph is constructed by integrating multi-source knowledge; it supports incremental processing of newly added papers, automatically identifies newly added entities and relationships, and dynamically updates the knowledge graph.

[0021] Secondly, the present invention provides a paper knowledge graph construction device based on a large model, comprising: The data extraction module is used to acquire original paper data and extraction rules; the original paper data includes at least one of the following: full text of the paper, abstract, citation data, author information, and journal indicators; the extraction rules include entity type definition, relation type definition, and validity period rule definition. The data cleaning module is used to preprocess the original paper data to obtain structured target text; the preprocessing includes at least one of text cleaning, paragraph segmentation, sentence boundary detection, and terminology normalization. The entity extraction module is used to extract entities from the target text using a large language model based on the extraction rules, and obtain an entity list; the entity list contains entity information for each entity, including entity name, entity type, validity period of each type, and entity attribute list; Specifically, the entity extraction module includes: a knowledge system embedding unit for embedding scientific knowledge systems into prompt templates of a large model; a multi-granularity extraction unit for configuring different extraction granularities for titles, abstracts, and full texts; and a conflict detection unit for performing conflict detection and normalization processing on the extracted entities. The relation extraction module is used to extract the relationship between any two entities in the entity list using a large language model, resulting in a relation list. The relation list contains relation information for each relationship, including the names of the two corresponding entities, the validity period of the relationship, and a relation description. Specifically, the relation extraction module includes: a prompt template design unit for designing relation extraction prompt templates to guide the large model in recognizing semantic relationships between entities; a thought chain guidance unit for guiding and verifying the relation reasoning process of the large model using thought chain technology; and a joint extraction unit for jointly extracting complex relationships across paragraphs and documents. The knowledge fusion module is used to align, deduplicatize, and detect conflicts of knowledge from different data sources, and to perform consistency verification using the reasoning capabilities of large models. The knowledge graph construction module is used to construct a knowledge graph based on the entity list and the relationship list, by integrating knowledge from multiple sources.

[0022] Thirdly, this invention provides a paper knowledge graph construction system based on a large model, comprising: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to execute any of the above-described methods for constructing large-model-based paper knowledge graphs.

[0023] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the paper knowledge graph construction method based on a large model as described above.

[0024] Compared with the prior art, the present invention has the following beneficial effects: First, improve the domain adaptability of knowledge graphs by embedding scientific knowledge systems into prompt templates of large models, so that the extraction results are semantically consistent with the subject norms, the graph structure is aligned with the depth of subject cognition, and the application value in scenarios such as precise retrieval and subject analysis is enhanced.

[0025] Second, it enhances the reliability and depth of knowledge extraction by using multi-granular text analysis and thought chain reasoning techniques to effectively capture implicit knowledge across paragraphs and documents, reduce the risk of large model "illusions," and improve the accuracy of triples; by introducing validity period modeling, it supports the time dimension traceability of knowledge.

[0026] Third, it achieves the organic integration of multi-source knowledge, solves data conflict and redundancy problems through knowledge fusion and conflict resolution mechanisms, and builds a more complete and consistent paper knowledge network; it overcomes the current shortcomings of difficulty in mining causal logic, resulting in a lack of in-depth decision support capabilities for graph relationships.

[0027] Fourth, improve the level of automation and modularity in construction. Modular device design with clear functions and interfaces for each module supports flexible expansion and optimization, reduces the cost of manual intervention, and supports large-scale and efficient construction. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a specific embodiment of the paper knowledge graph construction method based on a large model according to the present invention. Figure 2This is a structural block diagram of a paper knowledge graph construction device based on a large model, according to a specific embodiment of the present invention. Figure 3 This is a visualization example of a paper knowledge graph based on a large model, which is a specific embodiment of the present invention. Detailed Implementation

[0029] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0031] Example 1: Method for constructing a knowledge graph for academic papers.

[0032] This embodiment provides a method for constructing a paper knowledge graph based on a large model, such as... Figure 1 As shown, it includes the following steps: Step S11: Data and extraction rules acquisition.

[0033] The original paper data is obtained, including the full text, abstract, citation data, author information, journal metrics, etc. In this embodiment, 10,000 papers in the field of artificial intelligence are collected from academic databases such as arXiv, PubMed Central, and Springer.

[0034] Obtain the preset extraction rules, including: Entity type definition: includes seven categories: "algorithm", "dataset", "evaluation metric", "research institution", "scholar", "journal conference", and "research task"; Relationship type definitions include nine categories: "Proposed", "Improved", "Applied to", "Compared", "Citation", "Published in", and "Belongs to". Validity period rule definition: Set the validity period of entities and relationships, such as the validity period of a conference paper being the year the conference was held.

[0035] Step S12: Text processing.

[0036] The original paper data is preprocessed to obtain structured target text. Specifically, this includes: Text cleaning: Remove HTML tags, special characters, and extra spaces; Paragraph segmentation: Divide the entire text into natural paragraphs; Sentence boundary detection: Sentence segmentation was performed using the Spacy tool; Terminology normalization: Construct a domain-specific terminology dictionary to normalize different expressions of the same concept, such as unifying "deep neural network" and "DNN" into "deep neural network".

[0037] Step S13: Entity extraction.

[0038] Based on the extraction rules, a large language model is used to extract entities from the target text.

[0039] This embodiment uses GPT-4 as the base large model and performs entity extraction in the following way: Embed scientific knowledge systems into prompt templates, such as building prompts that include a CS domain classification system and an ACL thesaurus; Configure different extraction granularities for the title, abstract, and full text: extract only the core entities at the title level, extract the main entities at the abstract level, and perform full extraction at the full text level. The model is guided by a small number of labeled examples provided in the prompts, employing in-context learning techniques.

[0040] Step S14: Relation extraction.

[0041] A large language model is used to extract the relationship between any two entities in the entity list.

[0042] This embodiment employs the following strategy: Design a relationship extraction prompt template with the format: "Given the following two entities: <Entity A> and <Entity B>, please determine whether there is a relationship between them. If there is, please provide the relationship type and relationship description." Using the thinking chain technique, the model is guided to reason step by step: "First, determine the type of entities A and B, then determine the interaction between them based on the context, and finally determine the relationship type." For cross-paragraph relationships, a sliding window strategy is adopted to input paragraph blocks containing two entities into the model for joint judgment.

[0043] Step S15: Knowledge integration and conflict resolution.

[0044] This involves fusing entities and relationships from different papers. Specifically, it includes: Entity alignment: Entity alignment is performed based on entity name similarity and attribute similarity, merging different representations that point to the same real-world entity; Conflict detection: Perform consistency checks on different descriptions of the same relation. If a conflict is found between "Transformer published in ICLR" and "Transformer published in NeurIPS", the correct relation is determined by voting or confidence comparison. Utilize large models for inference verification: For suspected erroneous triples, ask questions to the large model for verification.

[0045] Step S16: Map construction and dynamic update.

[0046] A knowledge graph is constructed based on the merged entity list and relationship list. This embodiment uses the Neo4j graph database for storage and develops a visual interface.

[0047] Example 2: Device for constructing a knowledge graph for academic papers.

[0048] This embodiment provides a device for constructing a paper knowledge graph based on a large model, such as... Figure 2 As shown, it includes the following modules: The data acquisition module 201 is used to acquire original paper data and extraction rules. In this embodiment, the data acquisition module supports automatic data scraping from multiple academic database APIs and supports user-defined extraction rules.

[0049] The data cleaning module 202 is used to preprocess the original paper data to obtain structured target text. This module integrates text cleaning tools, paragraph segmentation algorithms, and a terminology normalization dictionary.

[0050] The entity extraction module 203 is used to extract entities from the target text using a large language model based on extraction rules. This module further includes: a knowledge system embedding unit 2031, used to embed scientific domain knowledge systems into the prompt template of the large model; a multi-granularity extraction unit 2032, used to configure different extraction granularities for the title, abstract, and full text; and a conflict detection unit 2033, used to perform conflict detection and normalization processing on the extracted entities.

[0051] The relation extraction module 204 is used to extract the relationship between any two entities in the entity list using a large language model. This module further includes: a prompt template design unit 2041 for designing relation extraction prompt templates; a thought chain guidance unit 2042 for guiding model reasoning using thought chain technology; and a joint extraction unit 2043 for jointly extracting complex relationships across paragraphs and documents.

[0052] The knowledge fusion module 205 is used to align, deduplicatize, and detect conflicts of knowledge from different data sources.

[0053] Graph construction module 206 is used to construct a knowledge graph based on an entity list and a relationship list, including visualization units and dynamically updated units.

[0054] Example 3: Thesis Knowledge Graph Construction System.

[0055] This embodiment provides a paper knowledge graph construction system based on a large model, such as... Figure 3 As shown, it includes: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to execute the paper knowledge graph construction method based on a large model as described in Embodiment 1.

[0056] In this embodiment, the system is deployed on a cloud server, supporting concurrent access by multiple users. Users upload paper data or specified data sources through a web interface, and the system automatically constructs a knowledge graph, providing functional interfaces such as visual browsing, SPARQL querying, and intelligent question answering.

[0057] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A large model-based thesis knowledge graph construction method, characterized in that, Includes the following steps: S11. Obtain the original paper data and the preset extraction rules; S12. Preprocess the original paper data to obtain the structured target text; S13. Based on the extraction rules, a large speech model is used to extract entities from the target text to obtain an entity list; S14. Use a large speech model to extract the relationship between any two entities in the entity list to obtain a relationship list; S15. Align, deduplicate, and detect conflicts for knowledge from different data sources, and use the reasoning capabilities of large models to perform consistency verification to ensure the integrity and accuracy of the knowledge graph. S16. Based on the entity list and the relationship list, multi-source knowledge is integrated to construct a knowledge graph; it supports incremental processing of newly added papers, automatically identifies newly added entities and relationships, and dynamically updates the knowledge graph.

2. The method for constructing a paper knowledge graph based on a large model according to claim 1, characterized in that, In step S11, the original paper data includes at least one of the following: full text of the paper, abstract, citation data, author information, and journal indicators; the extraction rules include entity type definition, relation type definition, and validity period rule definition.

3. The method for constructing a paper knowledge graph based on a large model according to claim 1, characterized in that, In step S12, the preprocessing includes at least one of text cleaning, paragraph segmentation, sentence boundary detection, and terminology normalization.

4. The method for constructing a paper knowledge graph based on a large model according to claim 1, characterized in that, In step S13, the entity list contains entity information for each entity, including entity name, entity type, validity period for each type, and entity attribute list. The entity extraction includes the following steps: embedding a scientific domain knowledge system into a prompt template of a large model, the scientific domain knowledge system including a subject classification system, a domain thesaurus, and a professional ontology; employing a multi-granularity text analysis strategy, configuring different extraction granularities for titles, abstracts, and full texts; and performing conflict detection and normalization processing on the extracted entities.

5. The method for constructing a paper knowledge graph based on a large model according to claim 1, characterized in that, In step S14, the relationship list contains relationship information for each relationship, including the names of the two corresponding entities, the validity period of the relationship, and the relationship description. Relationship extraction includes the following steps: designing a relationship extraction prompt template to guide the large model to identify semantic relationships between entities; using Chain-of-Thought technology to guide and verify the relationship reasoning process of the large model; and jointly extracting complex relationships across paragraphs and documents.

6. A device for constructing a paper knowledge graph based on a large model, characterized in that, include: The data extraction module is used to obtain the original paper data and extraction rules; The data cleaning module is used to preprocess the original paper data to obtain structured target text; The entity extraction module is used to extract entities from the target text based on the extraction rules using a large language model, and obtain an entity list. The relation extraction module is used to extract the relationship between any two entities in the entity list using a large language model, and obtain a relation list. The knowledge fusion module is used to align, deduplicatize, and detect conflicts of knowledge from different data sources, and to perform consistency verification using the reasoning capabilities of large models. The knowledge graph construction module is used to construct a knowledge graph based on the entity list and the relationship list, by integrating knowledge from multiple sources.

7. The device for constructing a paper knowledge graph based on a large model according to claim 6, characterized in that, The entity extraction module includes: Knowledge system embedding unit, a prompt template used to embed scientific knowledge systems into a large model; Multi-granularity extraction unit, which allows for configuring different extraction granularities for title, abstract, and full text; The conflict detection unit is used to perform conflict detection and normalization on the extracted entities.

8. The device for constructing a paper knowledge graph based on a large model according to claim 6, characterized in that, The relation extraction module includes: The prompt template design unit is used to design relation extraction prompt templates to guide large models in recognizing semantic relationships between entities. The thought chain guidance unit is used to guide and verify the relational reasoning process of a large model using thought chain technology; The joint extraction unit is used to jointly extract complex relationships across paragraphs and documents.

9. A paper knowledge graph construction system based on a large model, characterized in that, include: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to execute any of the above-described methods for constructing large-model-based paper knowledge graphs.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements any of the above-described methods for constructing a paper knowledge graph based on a large model.

Images