A large model-based literature knowledge base self-evolution construction method and system

Abstract

The application relates to the field of natural language processing, and provides a literature knowledge base self-evolution construction method and system based on a large model. The method comprises the following steps: performing semantic extension on a retrieval request input by a user to obtain a standardized query word set; performing retrieval and quality evaluation on a literature database through a double-channel retrieval mechanism according to the standardized query word set to obtain a candidate literature list; performing paragraph structure analysis and template-driven extraction on a plurality of literatures in the candidate literature list to obtain an original knowledge fragment set; performing term standardization and semantic fusion processing on the original knowledge fragment set to obtain unified knowledge representation; and performing incremental comparison between the unified knowledge representation and an existing knowledge graph to trigger conflict detection and version evolution management, and obtaining an updated knowledge graph. The application improves the cross-literature integration capability and improves the reliability of literature retrieval output.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for the self-evolutionary construction of a document knowledge base based on a large model. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, large language models have been widely used in fields such as scientific literature understanding, knowledge extraction, and knowledge graph construction. However, due to the continuous growth in the number of scientific literatures, the pressure of reading, organizing, and refining massive amounts of literature has become an important research direction in the field of AI-assisted scientific research.

[0003] Existing scientific literature knowledge extraction systems typically employ a single large language model or rule-based methods, integrating literature retrieval, information extraction, and semantic fusion into a single model. This leads to issues such as terminology differences, hierarchical inconsistencies, and variations in experimental conditions among knowledge descriptions from different sources. Consequently, these systems lack effective semantic alignment and conflict resolution mechanisms, making it difficult to verify the consistency and credibility of knowledge from multiple sources. The resulting knowledge graph suffers from inconsistent knowledge fusion and the inability to trace conflicting conclusions. Summary of the Invention

[0004] This invention provides a method and system for the self-evolution of a literature knowledge base based on a large model, in order to overcome the shortcomings of existing technologies.

[0005] This invention provides a method for the self-evolutionary construction of a literature knowledge base based on a large model, comprising: S1. Semantically expand the user's input search request to obtain a standardized set of query terms; S2. Based on the standardized query term set, the literature database is searched and its quality is evaluated through a dual-channel retrieval mechanism to obtain a list of candidate documents; S3. Perform paragraph structure parsing and template-driven extraction on multiple documents in the candidate document list to obtain the original knowledge fragment set; S4. Perform terminology standardization and semantic fusion processing on the original knowledge fragment set to obtain a unified knowledge representation; S5. Incrementally compare the unified knowledge representation with the existing knowledge graph to trigger conflict detection and version evolution management, and obtain an updated knowledge graph to complete the self-evolutionary construction of the document knowledge base.

[0006] According to the self-evolutionary construction method of a large-model-based literature knowledge base provided by the present invention, step S1 further includes: S11. Input the user's search request into the large language model, and rewrite the search request through multiple tasks including domain keyword generation, method term expansion, abbreviation restoration and semantic completion to obtain extended query terms; S12. Map each query term in the extended query term group to the semantic embedding space to obtain a standardized query term set.

[0007] According to the self-evolutionary construction method of a document knowledge base based on a large model provided by the present invention, step S2 further includes: S21. Using the standardized query term set as input, the document is matched using the BM25 scoring function, and the document is matched using the cosine similarity function, and the term matching score and semantic vector matching score are obtained respectively. S22. The term matching scores and semantic vector matching scores are linearly merged and sorted by weight to obtain a comprehensive relevance ranking result. An initial candidate set is selected from the comprehensive relevance ranking result. S23. The standardized citation values, journal impact factor, and time decay factor of multiple documents in the initial candidate set are weighted and summed according to preset weights to obtain the comprehensive quality score of multiple documents. S24. Sort and filter the initial candidate set according to the comprehensive quality score, and attach structured tags including topic classification, method type and data type to each document to obtain a candidate document list.

[0008] According to the self-evolutionary construction method of a large-model-based literature knowledge base provided by the present invention, the time decay factor in step S23 is calculated using an exponential decay formula and is denoted as: ; in, The time decay factor, The attenuation coefficient is... For the current year, This refers to the year the corresponding document was published.

[0009] According to the self-evolutionary construction method of a document knowledge base based on a large model provided by the present invention, step S3 further includes: S31. Perform discourse structure recognition on multiple document texts in the candidate document list and split them into single information document blocks that each carry discourse type annotations; S32. Input multiple single information document blocks into a predefined scientific knowledge extraction template, perform target-oriented extraction of experimental fields through large language model reasoning, and reason and complete the implicit content in the original text to obtain fine-grained scientific knowledge items. S33. Convert the fine-grained scientific knowledge items into a structured knowledge representation format to obtain the original knowledge fragment set.

[0010] According to the self-evolutionary construction method of a document knowledge base based on a large model provided by the present invention, step S4 further includes: S41. For multiple scientific research term entities in the original knowledge fragment set, semantic vectors are calculated through a term embedding model, and the multiple scientific research term entities are preprocessed according to the semantic vectors to obtain a standardized term set. S42. Map multiple concept nodes in the standardized terminology set to a unified concept hierarchy structure, assign a complete concept path to each node, and obtain an aligned concept graph; S43. For the multiple source attributes of the same entity in the alignment concept map, a weighted average fusion is performed using the comprehensive quality score of multiple source documents as the weight. When the semantic similarity of the text attributes is lower than the first preset threshold, it is marked as an attribute conflict identifier to obtain a unified knowledge representation.

[0011] According to the self-evolutionary construction method of a literature knowledge base based on a large model provided by the present invention, step S41, the preprocessing of multiple scientific research term entities specifically includes: When the semantic vector between any two scientific term entities is less than the second preset threshold, synonym term mapping is performed, and the average similarity of the context vectors before and after merging is controlled to be higher than the third preset threshold. When the semantic vector between any two scientific research term entities is greater than the second preset threshold, a new concept node is generated.

[0012] According to the self-evolutionary construction method of a literature knowledge base based on a large model provided by the present invention, step S43 further includes: Conflict resolution is performed on conflict entries marked with attribute conflict indicators, further including: S431. Calculate the normalized difference value between the same attribute values ​​from different documents. If the normalized difference value exceeds the fourth preset threshold, it is determined to be an attribute conflict entry. S432. Organize the source documents of conflict attributes into a logical chain structure. By comparing the differences between the condition layer and the result layer of different logical chains, distinguish between conditional conflicts and real conflicts, and obtain the conflict determination result. S433. When the conflict determination result is a real conflict entry, the conflict attribute value is weighted by the product of the comprehensive quality score of multiple source documents and the model extraction confidence level. The version with the highest confidence is selected as the main conclusion, and the other versions are written into the candidate version field to obtain a credible unified conclusion containing multiple version records.

[0013] According to the self-evolutionary construction method of a large-model-based literature knowledge base provided by the present invention, step S5 further includes: S51. By continuously optimizing and scanning scientific research databases, subscription update sources and user-uploaded literature, new unprocessed literature is detected by abstract hash comparison. If new literature is found, an update task identifier is generated. S52. Convert the newly added documents into a new set of knowledge fragments, compare them with the existing knowledge graph in multiple dimensions, determine the update type by calculating the entity structure fingerprint, add the new attributes, and obtain the updated knowledge graph.

[0014] This invention also provides a self-evolving document knowledge base construction system based on a large model, used to execute a self-evolving document knowledge base construction method based on a large model as described in any of the above claims, including: Extension module: Used to semantically expand the user's input search request to obtain a standardized set of query terms; Search module: Used to search and evaluate the quality of the literature database based on the standardized query term set through a dual-channel search mechanism to obtain a list of candidate documents; Extraction module: used to perform paragraph structure parsing and template-driven extraction on multiple documents in the candidate document list to obtain a set of original knowledge fragments; Fusion module: used to perform terminology standardization and semantic fusion processing on the original knowledge fragment set to obtain a unified knowledge representation; Update module: Used to incrementally compare the unified knowledge representation with the existing knowledge graph, trigger conflict detection and version evolution management, and obtain an updated knowledge graph to complete the self-evolutionary construction of the literature knowledge base.

[0015] This invention introduces a query rewriting mechanism and a dual-channel retrieval scoring fusion strategy during the literature retrieval stage. This allows the retrieval results to simultaneously consider precise term matching and implicit semantic associations. Combined with a comprehensive quality score based on citation volume, journal influence, and time decay factors, the candidate literature entering subsequent processing is effectively guaranteed in both relevance and authority, thus laying a high-quality data foundation for subsequent knowledge extraction. Secondly, in the knowledge extraction stage, this invention, through the joint drive of discourse structure recognition and predefined scientific knowledge extraction templates, decomposes complex literature content into single information document blocks carrying discourse type annotations. It then uses a large language model to infer and complete implicit knowledge, resulting in extraction results that outperform traditional rule-based methods in both structural integrity and semantic coverage. In the semantic fusion stage, a term merging mechanism using vector distance threshold judgment and context similarity dual verification, along with a unified concept hierarchy mapping rooted in scientific concepts, accurately merges heteronyms from different documents. Cross-document knowledge is integrated within a unified semantic framework, effectively eliminating the knowledge fragmentation problem caused by terminological ambiguity. Furthermore, in the conflict resolution phase, a logical chain structure of experimental conditions → research methods → experimental results is constructed to distinguish between conditional conflicts and real conflicts. The credibility of conflict conclusions is weighted using the product of the literature's overall quality score and the model's confidence level, ensuring that the final output knowledge graph possesses the characteristics of traceable conflict origins and interpretable conclusions, significantly enhancing the credibility of the knowledge base content. In the knowledge evolution phase, a joint judgment mechanism of abstract hash comparison and entity structure fingerprint calculation enables continuous awareness and incremental updates of newly added literature. Simultaneously, relying on the evolution history record table, the version number, source literature, and reasoning path of each change are fully preserved, allowing the knowledge graph to continuously evolve with research progress and support historical state rollback, ensuring the long-term effectiveness and dynamic maintainability of the knowledge base. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A schematic diagram of the process for constructing a self-evolving literature knowledge base based on a large model, provided by this invention; Figure 2 This invention provides a schematic diagram of a self-evolving document knowledge base construction system based on a large model. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] The embodiments of the present invention are described below with reference to the figures.

[0020] like Figure 1 As shown, this invention provides a self-evolving construction method for a literature knowledge base based on a large model, comprising: S1. Semantically expand the user's search request to obtain a standardized set of query terms.

[0021] Step S1 further includes: S11. Input the user's search request into the large language model, and rewrite the search request through multiple tasks including domain keyword generation, method term expansion, abbreviation restoration and semantic completion to obtain an expanded query term set; S12. Map each query term in the expanded query term set to the semantic embedding space to obtain a standardized query term set.

[0022] In step S1, the user-input search request suffers from issues such as non-standard expression, mixed abbreviations, and incomplete terminology coverage. Therefore, step S11 of this invention first feeds the user-input research topic or keywords into a large language model. Through four tasks—domain keyword generation, methodological terminology expansion, abbreviation restoration, and semantic completion—the original input is rewritten, correcting and supplementing errors and omissions in the user's description to ensure consistency with the terminology used in the literature, resulting in an expanded query term set. Subsequently, in step S12, this invention maps each query term in the expanded query term set to a 1024-dimensional semantic embedding space. Each query term corresponds to a fixed-dimensional semantic vector representation, ultimately yielding a standardized query term set. The semantic embedding space refers to a representation space that maps text to high-dimensional numerical vectors. The distance between vectors reflects the semantic similarity of the text; the higher the dimension, the more refined the semantic expression.

[0023] S2. Based on the standardized query term set, the literature database is searched and its quality is evaluated through a dual-channel retrieval mechanism to obtain a list of candidate documents.

[0024] Step S2 further includes: S21. Using the standardized query term set as input, the document is matched using the BM25 scoring function, and the document is matched using the cosine similarity function, so as to obtain the term matching score and the semantic vector matching score respectively.

[0025] In step S21, the present invention first takes a standardized query term set as input and simultaneously activates two independent retrieval channels. The first channel uses the BM25 scoring function. BM25 is a classic text matching algorithm based on term frequency and inverse document frequency. It comprehensively calculates the occurrence frequency of each query term in the document, the total number of words in the document, and the average number of words in the document set, outputting a term matching score for each document relative to the query term set. A higher term matching score indicates a greater degree of literal overlap between the document and the query term. The second channel calculates the semantic vector of each query term and the semantic vector of the document using the cosine similarity formula. Cosine similarity measures the directional closeness of two vectors, with a value ranging from 0 to 1. The closer the value is to 1, the more semantically similar the two texts are, thus obtaining a semantic vector matching score for each document.

[0026] Furthermore, the BM25 scoring function is used to calculate the degree of term matching, expressed as: ; in, Literature representing calculation For query The term matching score, among which Indicates the document object, This represents the set of query terms after query rewriting. For the vector index in the query term set, Indicates the first One query term, Documents Chinese search terms Frequency of occurrence (word frequency). Documents Total word count This represents the average word count of a document collection. Inverse citation frequency is used to measure The importance of This indicates that BM25 parameters have been adjusted to control word frequency saturation, set to 1.2. This represents the document length normalization factor, set to 0.75.

[0027] The semantic retrieval subchannel uses a vector database to calculate cosine similarity, expressed as: ; in, For literature For query The semantic vector matching score, Indicates query terms The vector representation of , Documents The vectorized representation of .

[0028] S22. The term matching scores and semantic vector matching scores are linearly merged and sorted by weight to obtain a comprehensive relevance ranking result. An initial candidate set is selected from the comprehensive relevance ranking result.

[0029] In step S22, the present invention linearly weights and merges the term matching score and semantic vector matching score with weights of 0.6 and 0.4 respectively to obtain a comprehensive relevance score for each document. The documents are then sorted from highest to lowest score, and the top 200 documents are selected as the initial candidate set. The different weights assigned to the two scores are because term matching is more sensitive to precise term matching, while semantic vector matching excels at capturing implicit connections with different expressions but similar meanings. The combined score, resulting from the complementarity of the two, better reflects the true relevance between the document and the query request than a single channel.

[0030] S23. The standardized citation values, journal impact factor, and time decay factor of multiple documents in the initial candidate set are weighted and summed according to preset weights to obtain the comprehensive quality score of multiple documents.

[0031] The time decay factor in step S23 is calculated using the exponential decay formula and is denoted as: ; in, For the literature object, The time decay factor, The attenuation coefficient is... For the current year, This refers to the year the corresponding document was published.

[0032] After obtaining the initial candidate set in step S23, this invention further calculates a comprehensive quality score for each document. The comprehensive quality score is obtained by weighted summation of three components. The citation frequency standardization value reflects the frequency of citation of the document in the academic community. After normalization, the difference in magnitude is eliminated. The journal impact factor reflects the academic status of the journal in the field. The time decay factor is calculated using the difference between the current year and the publication year of the document as the independent variable, and the decay coefficient is set to 0.15. The earlier the publication time of the document, the smaller the decay factor, reflecting the law that the timeliness of the document decreases over time. Finally, the three components are weighted and summed with weights of 0.5, 0.3, and 0.2 to obtain the comprehensive quality score of each document.

[0033] Furthermore, the calculated overall quality score is expressed as: ; in, For the overall quality score, This is a standardized value for citation count. It is the journal's impact factor.

[0034] S24. Sort and filter the initial candidate set according to the comprehensive quality score, and attach structured tags including topic classification, method type and data type to each document to obtain a candidate document list.

[0035] In step S24, the present invention reorders the initial candidate set based on the comprehensive quality score and adds three types of structured tags to each document: topic classification, method type, and data type, to obtain a candidate document list. The structured tags are automatically generated by the large language model based on the document abstract and full text content, and are used to mark the research topic category, research method type, and data type of the document, providing a classification index basis for parsing the document content as needed in the subsequent knowledge extraction stage.

[0036] S3. Perform paragraph structure parsing and template-driven extraction on multiple documents in the candidate document list to obtain the original knowledge fragment set.

[0037] Step S3 further includes: S31. Perform discourse structure recognition on multiple document texts in the candidate document list and split them into single information document blocks, each carrying a discourse type label.

[0038] Furthermore, the original text structure of the literature in the candidate literature list is complex, containing various discourse structures such as title, abstract, introduction, methods, experiments, results and discussion. Directly extracting knowledge from the full text can easily cause context confusion and information misalignment. Therefore, in step S31, the present invention first performs discourse structure recognition on each literature text in three ways: title style, paragraph indentation and syntactic delimiters. The literature is then divided into several literature blocks, each containing only a single type of information content. A discourse type labeling field is added to each literature block to indicate its discourse category, including abstract, methods, experiments, discussion, etc., resulting in a set of single information literature blocks with discourse type labels.

[0039] S32. Input multiple single information document blocks into a predefined scientific knowledge extraction template, perform target-oriented extraction of experimental fields through large language model reasoning, and reason and complete the implicit content in the original text to obtain fine-grained scientific knowledge items.

[0040] In step S32, the present invention inputs each of the aforementioned single information document blocks into a predefined scientific knowledge extraction template. This template predefines target extraction fields such as experimental variables, model names, model components, input data, evaluation indicators, experimental procedures, and experimental conclusions. Subsequently, the present invention uses a large language model, with the document block text and template field definitions as input, to perform target-oriented extraction on each field. Content explicitly described in the original text is directly extracted and filled in; for content implicit in the original text but inferred from the context, the large language model performs reasoning to complete the extraction. The goal is for all fields in the template to output complete content, resulting in fine-grained scientific knowledge entries corresponding to each document block.

[0041] S33. Convert the fine-grained scientific knowledge items into a structured knowledge representation format to obtain the original knowledge fragment set.

[0042] Step S33: After obtaining the fine-grained scientific knowledge entries, this invention converts them into a structured knowledge representation format. Each knowledge entry uses an entity number prefix plus a 64-bit hash value as a unique entity identifier, and is organized into three categories: entity, attribute, and relation. The entity category is divided into two types: scientific terminology and experimental content. The attributes of scientific terminology are the relevant descriptions of the term, while the attributes of experimental content include fields such as experimental variables, model components, evaluation indicators, data sources, experimental procedures, and experimental conclusions. The performance score field records the quantitative results in the experimental conclusions in the form of evaluation indicators, numerical values, and dataset triples. After all document blocks have been processed, the original knowledge fragment set is obtained.

[0043] S4. Perform terminology standardization and semantic fusion processing on the original knowledge fragment set to obtain a unified knowledge representation.

[0044] Step S4 further includes: S41. For multiple scientific research term entities in the original knowledge fragment set, semantic vectors are calculated using a term embedding model, and the multiple scientific research term entities are preprocessed based on the semantic vectors to obtain a standardized term set. In step S41, the preprocessing of multiple scientific research term entities specifically includes: when the semantic vector between any two scientific research term entities is less than a second preset threshold, synonym term mapping is performed, and the average similarity of the context vectors before and after merging is controlled to be higher than a third preset threshold; when the semantic vector between any two scientific research term entities is greater than the second preset threshold, a new concept node is generated.

[0045] Furthermore, since the original knowledge fragment set comes from multiple documents, there are problems such as different names of terms, different abbreviations, and cross-domain transfer in the expression of the same scientific research concept in different documents. In step S41, the present invention first inputs all scientific research term entities in the original knowledge fragment set into the term embedding model. The term embedding model is a pre-trained model that maps text terms into fixed-dimensional numerical vectors. The distance between the vectors reflects the semantic similarity of the terms.

[0046] Specifically, this invention calculates the vector distance between the semantic vectors of any two scientific research term entities pairwise. When the vector distance between the two is less than a second preset threshold (set to 0.18 in this embodiment), they are determined to be synonymous terms, and the less standardized terms are merged and mapped to more standardized terms. After merging, this invention further extracts the contextual semantic vectors of the document blocks containing the two terms before and after merging, calculates the average contextual vector similarity, and requires that this similarity be higher than a third preset threshold (set to 0.72 in this embodiment). If it is lower than this threshold, the merging is canceled and independent concept nodes are regenerated. In addition, when the vector distance between any two term entities is greater than the second preset threshold, they are regarded as independent concepts, and new concept nodes are directly generated, ultimately resulting in a standardized term set.

[0047] S42. Map multiple concept nodes in the standardized terminology set to a unified concept hierarchy structure, assign a complete concept path to each node, and obtain an aligned concept graph.

[0048] In step S42, the present invention maps all concept nodes in the standardized terminology set to a unified concept hierarchy structure. This hierarchy structure takes scientific concepts as the root node and includes sub-nodes such as models, datasets, evaluation indicators, and experimental factors. Each concept node is assigned to the corresponding sub-level according to its semantic category, and a complete concept path string is generated along the hierarchy path. The complete hierarchy affiliation of the node from the root node to the current node is recorded to obtain the aligned concept graph.

[0049] S43. For the multiple source attributes of the same entity in the alignment concept map, a weighted average fusion is performed using the comprehensive quality score of multiple source documents as the weight. When the semantic similarity of the text attributes is lower than the first preset threshold, it is marked as an attribute conflict identifier to obtain a unified knowledge representation.

[0050] Furthermore, after obtaining the aligned concept map, the present invention performs fusion processing on the multi-source attributes of the same entity in step S43. Numerical attributes are calculated using a weighted average based on the comprehensive quality score of each source document, which is calculated in step S2. Documents with higher scores have a larger proportion of their attribute values ​​in the fusion result. For textual attributes, the semantic similarity between the multi-source texts is first calculated. If the similarity is higher than a first preset threshold (set to 0.65 in this example), they are directly merged into a unified description. If the semantic similarity is lower than the first preset threshold, an attribute conflict identifier is marked, and the attribute is passed to the conflict resolution process. The remaining non-conflicting attributes are then fused to obtain a unified knowledge representation.

[0051] Step S43 is followed by: performing conflict resolution on conflict entries marked with attribute conflict identifiers, which further includes: S431. Calculate the normalized difference value between the same attribute values ​​from different documents. If the normalized difference value exceeds the fourth preset threshold, it is determined to be an attribute conflict entry. S432. Organize the source documents of the conflicting attribute into a logical chain structure. By comparing the differences between the condition layer and the result layer of different logical chains, distinguish between conditional conflicts and real conflicts to obtain the conflict determination result. S433. When the conflict determination result is a real conflict entry, use the product of the comprehensive quality score of multiple source documents and the model extraction confidence level as the weight to perform a weighted calculation on the conflicting attribute value. Select the version with the highest confidence as the main conclusion, and write the other versions into the candidate version field to obtain a credible unified conclusion containing multiple version records.

[0052] Furthermore, for conflict entries marked with attribute conflict identifiers, this invention first calculates the normalized difference value between the same attribute values ​​from different documents. The normalized difference value is calculated by dividing the absolute value of the difference between the two attribute values ​​by the larger of the two values. If the normalized difference value exceeds the fourth preset threshold (set to 0.25 in this example), it is formally determined to be an attribute conflict entry and enters the subsequent logical chain analysis.

[0053] The above numerical conflict calculation is expressed as: ; in , This indicates the same attribute value from different documents. This represents the attribute difference value after normalization. If It was ruled a conflict.

[0054] In logical chain analysis, this invention organizes the conflicting information from various sources into a three-layer logical chain structure of "experimental conditions → research methods → experimental results". Conflict attribution is performed by comparing the differences in the condition layer and the differences in the result layer of logical chains from different sources: if there are differences in the condition layer and the differences in the result layer are caused by differences in the condition, it is determined to be a conditional conflict; if the condition layer is consistent but there are still differences in the result layer, it is determined to be a real conflict, and the conflict determination result is obtained.

[0055] For entries determined to be genuine conflicts, this invention uses the product of the overall quality score of each source document and the model extraction confidence score as the weight to perform a weighted calculation on each conflict attribute value. The model extraction confidence score is the confidence score output by the large language model when extracting that attribute, reflecting the reliability of the extraction result. After the weighted calculation is completed, the attribute value version with the highest overall weight is selected as the main conclusion and written into the unified knowledge representation, while the remaining versions are written into the candidate version field for storage, ultimately obtaining a credible unified conclusion containing multiple version records.

[0056] S5. Incrementally compare the unified knowledge representation with the existing knowledge graph to trigger conflict detection and version evolution management, and obtain an updated knowledge graph to complete the self-evolutionary construction of the document knowledge base.

[0057] Step S5 further includes: S51. By continuously optimizing and scanning scientific research databases, subscribed update sources, and user-uploaded documents, new unprocessed documents are detected by abstract hash comparison. If new documents are found, an update task identifier is generated.

[0058] In step S51, this invention aims to continuously optimize and scan scientific research databases, subscribed update sources, and user-uploaded documents. New documents are detected using abstract hash comparison. The abstract hash is a fixed-length string obtained by calculating the hash value of the document abstract text; different abstracts correspond to different hash values. This invention compares the newly scanned document abstract hash with the abstract hash records of already processed documents one by one. If a new hash value does not match, it is determined to be a new document, and an update task identifier is generated, triggering the incremental inference process. Furthermore, when a user initiates a search request, this invention also determines whether the current knowledge base content needs to be updated; if so, the above process is triggered simultaneously.

[0059] S52. Convert the newly added documents into a new set of knowledge fragments, compare them with the existing knowledge graph in multiple dimensions, determine the update type by calculating the entity structure fingerprint, add the new attributes, and obtain the updated knowledge graph.

[0060] In step S52, the present invention first processes the newly added documents through steps S3 and S4 to obtain a new set of knowledge fragments. Then, the present invention compares the new set of knowledge fragments with the existing knowledge graph in three dimensions: concept path, attribute hashing, and relation graph isomorphism. The entity structure fingerprint is generated by hashing the concatenation of four fields: entity type, entity label, attribute set, and relation set, and is used to uniquely identify the complete structural state of a knowledge entity.

[0061] Subsequently, this invention compares the structural fingerprint of each entity in the new knowledge fragment with the structural fingerprint of the corresponding entity in the existing knowledge graph. Entities with completely identical fingerprints are not written repeatedly. Entities with new attributes are directly appended to the existing entity record. Old attributes that are overturned by new literature are replaced after being judged by the conflict resolution process, thus obtaining an updated knowledge graph.

[0062] like Figure 2 As shown, the present invention also provides a self-evolving document knowledge base construction system based on a large model, comprising: Extension module 100: used to semantically expand the user's input search request to obtain a standardized set of query terms; Search module 200: Used to search and evaluate the quality of the literature database based on the standardized query term set through a dual-channel search mechanism to obtain a list of candidate documents; Extraction module 300: used to perform paragraph structure parsing and template-driven extraction on multiple documents in the candidate document list to obtain a set of original knowledge fragments; Fusion module 400: used to perform terminology standardization and semantic fusion processing on the original knowledge fragment set to obtain a unified knowledge representation; Update module 500: Used to incrementally compare the unified knowledge representation with the existing knowledge graph, trigger conflict detection and version evolution management, and obtain an updated knowledge graph to complete the self-evolutionary construction of the literature knowledge base.

[0063] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0064] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the self-evolutionary construction method of a large model-based document knowledge base described in the various embodiments or some parts of the embodiments.

[0065] The purpose of this invention is to overcome the problems of insufficient single-model capabilities, inconsistent extraction results, lack of knowledge fusion mechanism, and lack of knowledge evolution capability in existing scientific literature knowledge extraction systems, and to provide a scientific literature knowledge extraction and integration system based on multi-model collaborative reasoning, which realizes automatic retrieval, structured extraction, semantic fusion and conflict verification of scientific literature information.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for self-evolutionary construction of a document knowledge base based on a large model, characterized in that, include: S1. Semantically expand the user's input search request to obtain a standardized set of query terms; S2. Based on the standardized query term set, the literature database is searched and its quality is evaluated through a dual-channel retrieval mechanism to obtain a list of candidate documents; S3. Perform paragraph structure parsing and template-driven extraction on multiple documents in the candidate document list to obtain the original knowledge fragment set; S4. Perform terminology standardization and semantic fusion processing on the original knowledge fragment set to obtain a unified knowledge representation; S5. Incrementally compare the unified knowledge representation with the existing knowledge graph to trigger conflict detection and version evolution management, and obtain an updated knowledge graph to complete the self-evolutionary construction of the document knowledge base.

2. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 1, characterized in that, Step S1 further includes: S11. Input the user's search request into the large language model, and rewrite the search request through multiple tasks including domain keyword generation, method term expansion, abbreviation restoration and semantic completion to obtain extended query terms; S12. Map each query term in the extended query term group to the semantic embedding space to obtain a standardized query term set.

3. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 1, characterized in that, Step S2 further includes: S21. Using the standardized query term set as input, the document is matched using the BM25 scoring function, and the document is matched using the cosine similarity function, and the term matching score and semantic vector matching score are obtained respectively. S22. The term matching scores and semantic vector matching scores are linearly merged and sorted by weight to obtain a comprehensive relevance ranking result. An initial candidate set is selected from the comprehensive relevance ranking result. S23. The standardized citation values, journal impact factor, and time decay factor of multiple documents in the initial candidate set are weighted and summed according to preset weights to obtain the comprehensive quality score of multiple documents. S24. Sort and filter the initial candidate set according to the comprehensive quality score, and attach structured tags including topic classification, method type and data type to each document to obtain a candidate document list.

4. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 3, characterized in that, The time decay factor in step S23 is calculated using the exponential decay formula and is denoted as: ； in, The time decay factor, The attenuation coefficient is... For the current year, This refers to the year the corresponding document was published.

5. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 1, characterized in that, Step S3 further includes: S31. Perform discourse structure recognition on multiple document texts in the candidate document list and split them into single information document blocks that each carry discourse type annotations; S32. Input multiple single information document blocks into a predefined scientific knowledge extraction template, perform target-oriented extraction of experimental fields through large language model reasoning, and reason and complete the implicit content in the original text to obtain fine-grained scientific knowledge items. S33. Convert the fine-grained scientific knowledge items into a structured knowledge representation format to obtain the original knowledge fragment set.

6. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 1, characterized in that, Step S4 further includes: S41. For multiple scientific research term entities in the original knowledge fragment set, semantic vectors are calculated through a term embedding model, and the multiple scientific research term entities are preprocessed according to the semantic vectors to obtain a standardized term set. S42. Map multiple concept nodes in the standardized terminology set to a unified concept hierarchy structure, assign a complete concept path to each node, and obtain an aligned concept graph; S43. For the multiple source attributes of the same entity in the alignment concept map, a weighted average fusion is performed using the comprehensive quality score of multiple source documents as the weight. When the semantic similarity of the text attributes is lower than the first preset threshold, it is marked as an attribute conflict identifier to obtain a unified knowledge representation.

7. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 6, characterized in that, In step S41, the preprocessing of multiple scientific research term entities specifically includes: When the semantic vector between any two scientific term entities is less than the second preset threshold, synonym term mapping is performed, and the average similarity of the context vectors before and after merging is controlled to be higher than the third preset threshold. When the semantic vector between any two scientific research term entities is greater than the second preset threshold, a new concept node is generated.

8. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 6, characterized in that, Step S43 is followed by: Conflict resolution is performed on conflict entries marked with attribute conflict indicators, further including: S431. Calculate the normalized difference value between the same attribute values ​​from different documents. If the normalized difference value exceeds the fourth preset threshold, it is determined to be an attribute conflict entry. S432. Organize the source documents of conflict attributes into a logical chain structure. By comparing the differences between the condition layer and the result layer of different logical chains, distinguish between conditional conflicts and real conflicts, and obtain the conflict determination result. S433. When the conflict determination result is a real conflict entry, the conflict attribute value is weighted by the product of the comprehensive quality score of multiple source documents and the model extraction confidence level. The version with the highest confidence is selected as the main conclusion, and the other versions are written into the candidate version field to obtain a credible unified conclusion containing multiple version records.

9. The self-evolutionary construction method for a document knowledge base based on a large model according to claim 1, characterized in that, Step S5 further includes: S51. By continuously optimizing and scanning scientific research databases, subscription update sources and user-uploaded literature, new unprocessed literature is detected by abstract hash comparison. If new literature is found, an update task identifier is generated. S52. Convert the newly added documents into a new set of knowledge fragments, compare them with the existing knowledge graph in multiple dimensions, determine the update type by calculating the entity structure fingerprint, add the new attributes, and obtain the updated knowledge graph.

10. A self-evolving document knowledge base construction system based on a large model, used to execute the self-evolving document knowledge base construction method based on a large model as described in any one of claims 1 to 9, characterized in that, include: Extension module: Used to semantically expand the user's input search request to obtain a standardized set of query terms; Search module: Used to search and evaluate the quality of the literature database based on the standardized query term set through a dual-channel search mechanism to obtain a list of candidate documents; Extraction module: used to perform paragraph structure parsing and template-driven extraction on multiple documents in the candidate document list to obtain a set of original knowledge fragments; Fusion module: used to perform terminology standardization and semantic fusion processing on the original knowledge fragment set to obtain a unified knowledge representation; Update module: Used to incrementally compare the unified knowledge representation with the existing knowledge graph, trigger conflict detection and version evolution management, and obtain an updated knowledge graph to complete the self-evolutionary construction of the literature knowledge base.

Images