A heterogeneous scientific-technological knowledge graph construction method and system

Abstract

This invention provides a method and system for constructing a heterogeneous science and technology knowledge graph, comprising: analyzing multi-source scientific and technological literature and expert interview data; defining and verifying the consistency of knowledge structures in scientific papers and technological patents through ontology modeling; constructing a hierarchical heterogeneous ontology framework containing two types of knowledge sub-ontologies and cross-domain mapping relationships; building an entity-relation joint extraction framework based on this framework and a literature database using a domain-fine-tuned large language model; extracting sets of scientific and technological knowledge triplets through dual-channel retrieval enhancement; learning unified semantic vectors through graph knowledge representation; achieving cross-domain entity disambiguation and alignment through pre-filtering and large model inference; generating enhanced fusion triplets by adding metadata; constructing a multi-granularity temporal hypergraph based on this; integrating temporal information and storing it in a graph database to form a graph database that supports multi-granularity temporal queries. This invention solves the problems of difficulty in cross-source heterogeneous knowledge fusion, low accuracy of entity disambiguation, and insufficient representation of graph temporal evolution.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and knowledge engineering technology, specifically involving a method and system for constructing heterogeneous science and technology knowledge graphs, which can be applied to scenarios such as science and technology intelligence analysis, R&D decision support and science and technology transformation path mining. Background Technology

[0002] Knowledge graph technology is a crucial infrastructure for semantic knowledge organization and intelligent reasoning, and is widely used in fields such as scientific and technological intelligence analysis, R&D decision support, and innovation path discovery. Scientific and technological knowledge graphs typically store domain knowledge in the form of triples (entity-relationship-entity), supporting cross-document and cross-institutional knowledge association queries and reasoning.

[0003] Existing methods for constructing scientific and technological knowledge graphs mainly fall into two categories: one is the traditional information extraction method based on rules and dictionaries, which extracts knowledge from structured documents through manually defined extraction rules, relies on manually maintained rule bases, and has limited generalization ability; the other is the neural network method based on pre-trained language models, which achieves named entity recognition and relation extraction by fine-tuning models such as BERT and GPT, and shows better extraction accuracy in specific domains. The existing technologies most relevant to this invention focus on knowledge graph construction for single document types: either constructing scientific knowledge graphs for academic papers or constructing patent knowledge graphs for technical patents. Both methods have relatively mature technical approaches for their respective document types.

[0004] However, the existing technology has the following two main technical problems:

[0005] First, the heterogeneous knowledge from scientific papers and technological patents is difficult to align and integrate within a unified framework. Scientific papers focus on theoretical logic and causal reasoning, while technological patents emphasize functional implementation and application scenarios. Existing methods struggle to process both types of literature simultaneously, leading to a long-term separation between scientific and technological knowledge and failing to effectively reflect the evolutionary path of scientific discoveries transforming into technological applications.

[0006] Secondly, the cross-domain disambiguation accuracy for large-scale scientific and technological entities is insufficient, and knowledge graphs lack the ability to express multi-granular temporal evolution. Existing entity alignment methods mainly rely on literal similarity matching, which results in a high disambiguation error rate when faced with a large number of synonymous but different terms in the scientific and technological field (such as "convolutional neural network" and "CNN"). At the same time, existing graphs only label knowledge nodes with a single timestamp, which cannot distinguish the differences in knowledge evolution patterns at different time granularities (annual macro trends, quarterly evolution, and instant dissemination of citations). Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for constructing heterogeneous scientific and technological knowledge graphs, so as to solve the technical problems in the prior art, such as the difficulty in unifying and integrating cross-source heterogeneous knowledge from scientific papers and technical patents, the insufficient accuracy of cross-domain entity disambiguation, and the lack of multi-granular temporal evolution expression capabilities of knowledge graphs.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a method for constructing a heterogeneous science and technology knowledge graph, including:

[0010] We analyze multi-source scientific literature and domain expert interview data, and formally define and verify the differences in knowledge structure between scientific papers and technical patents through ontology modeling methods. We construct a hierarchical heterogeneous ontology framework that includes scientific knowledge sub-ontologies, technical knowledge sub-ontologies and cross-domain mapping relationship sets.

[0011] Using the aforementioned hierarchical heterogeneous ontology framework and multi-source scientific and technological literature database as input, an entity-relation joint extraction framework is constructed through a domain-fine-tuned large language model. Dual-channel retrieval is performed iteratively in batches to enhance triple extraction. All batch results are merged to obtain a set of scientific knowledge triples and a set of technical knowledge triples.

[0012] Based on the aforementioned scientific knowledge triplet set and technical knowledge triplet set, entities and relations are mapped to a unified semantic vector space through graph knowledge representation learning. Cross-domain entity disambiguation and alignment are completed through dual-threshold pre-filtering and two-stage progressive large language model reasoning. After adding document metadata, an enhanced fusion triplet set is obtained.

[0013] Based on the enhanced fusion triplet set and semantic vectors of entities in each domain, a multi-granularity temporal hypergraph is constructed and the temporal information of each granularity layer is weighted and integrated through a cross-scale conflict perception mechanism. The resulting heterogeneous science-technology semantic knowledge graph is persistently stored in a graph database, resulting in a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

[0014] Optionally, the analysis of multi-source scientific literature and domain expert interview data, and the formal definition and consistency test of the knowledge structure differences between scientific papers and technological patents using ontology modeling methods, constructing a hierarchical heterogeneous ontology framework including scientific knowledge sub-ontologies, technological knowledge sub-ontologies, and cross-domain mapping relationship sets, includes:

[0015] Based on multi-source scientific and technological literature and interview data with domain experts, we sorted out the differences in knowledge structure between scientific papers and technology patents, and obtained a draft of the concept system of scientific and technological knowledge domain that includes a set of entity types and a set of relation types.

[0016] Based on the aforementioned draft concept system of scientific and technological knowledge domain, the entity type set and the relation type set are formally defined and attribute constraints are added using ontology modeling tools. Consistency checks are then performed to obtain an empirical scientific and technological knowledge ontology that includes the entity type set, relation type set, and attribute constraint set.

[0017] Based on the entity type set and relation type set in the verified scientific and technological knowledge ontology, a scientific knowledge sub-ontology containing research questions, research methods, and experimental conclusions, and a technical knowledge sub-ontology containing technical efficacy, application scenarios, and invention methods are defined. A cross-domain mapping relation set between the scientific knowledge sub-ontology and the technical knowledge sub-ontology is established based on the attribute constraint set, resulting in a hierarchical heterogeneous ontology framework that supports cross-domain alignment.

[0018] Optionally, the entity-relation joint extraction framework constructed by fine-tuning a large language model through domain includes:

[0019] Based on the hierarchical heterogeneous ontology framework and multi-source scientific and technological literature database, the knowledge entities and relationships between the texts in the multi-source scientific and technological literature database are manually annotated, and the multi-source scientific and technological literature database is divided into multiple batches according to semantic topics to obtain an annotated sample set and a batched literature index table.

[0020] Based on the labeled sample set, the entity type set and relation type set defined in the hierarchical heterogeneous ontology framework are used as type constraints to fine-tune the pre-trained large language model in the domain. The batch iterative extraction order of the fine-tuned large language model is set according to the batch literature index table to obtain an entity-relation joint extraction framework that can simultaneously identify the relationship between scientific entities, technical entities and knowledge entities.

[0021] Optionally, the batch-iterative dual-channel retrieval enhanced triple extraction includes:

[0022] Based on the entity-relation joint extraction framework, the current batch of documents is initially extracted to obtain a candidate pre-entity set and a candidate triple set, and the intermediate incremental graph up to the current batch is loaded.

[0023] Based on the candidate pre-entity set and the intermediate incremental graph, each pre-entity in the candidate pre-entity set is used as a search query term to retrieve cross-batch text fragments in the full document of the multi-source scientific and technological literature database, so as to obtain the semantic context corresponding to each pre-entity.

[0024] Based on the candidate pre-entity set and the intermediate incremental graph, the neighborhood relationship information of each pre-entity is retrieved from the intermediate incremental graph to obtain the graph structure information corresponding to each pre-entity;

[0025] The semantic context and the graph structure information are combined according to the pre-entity alignment to obtain the dual-channel enhanced context corresponding to each pre-entity.

[0026] The dual-channel enhanced context and the candidate triple set are jointly input into the large language model in the entity-relation joint extraction framework to obtain the refined triple set of the current batch. The refined triple set is then merged into the intermediate incremental graph to update the intermediate incremental graph. This process is iterated until all batches are processed. Finally, the refined triple sets of all batches are merged to obtain the scientific knowledge triple set and the technical knowledge triple set.

[0027] Optionally, the hallucination filtering in the dual-channel retrieval enhanced triple extraction includes:

[0028] Based on each candidate triplet in the refined triplet set and the original text of the current batch of documents, the candidate triplet and the original text are used to construct a verification prompt, which is then input into a large language model. The large language model scores the semantic consistency between the candidate triplet and the original text to obtain the semantic consistency confidence score corresponding to each candidate triplet.

[0029] Based on the semantic consistency confidence score, candidate triples below the preset confidence threshold are filtered out from the refined triple set to obtain the filtered triple set for the current batch. The filtered triple set is then merged into the intermediate incremental graph to update the intermediate incremental graph. This process is iterated until all batches are processed. Finally, the filtered triple sets from all batches are merged to obtain the scientific knowledge triple set and the technical knowledge triple set.

[0030] Optionally, the step of mapping entities and relations to a unified semantic vector space through graph knowledge representation learning includes:

[0031] Based on the scientific knowledge triple set and the technical knowledge triple set, graph knowledge representation learning is performed on entities and relations, mapping the entities and relations in the scientific knowledge triple set and the technical knowledge triple set to a unified semantic vector space, resulting in a scientific entity vector set and a technical entity vector set.

[0032] Based on the scientific entity vector set and the technical entity vector set, the cosine similarity score between cross-domain entity pairs is calculated, and the top-ranked attribute feature set and relation feature set of each candidate entity pair are extracted to obtain the entity candidate set to be disambiguated. Each record in the entity candidate set to be disambiguated contains an entity pair, a cosine similarity score, an attribute feature set, and a relation feature set.

[0033] Optionally, the dual-threshold pre-filtering includes:

[0034] Based on the candidate set of entities to be disambiguated, according to the cosine similarity score of each record in the candidate set, candidate entity pairs with cosine similarity scores not lower than a preset high threshold are assigned to the confirmation merge set, candidate entity pairs with cosine similarity scores lower than a preset low threshold are assigned to the exclusion set, and candidate entity pairs with cosine similarity scores between the preset low threshold and the preset high threshold are assigned to the candidate set to be inferred, thus obtaining the confirmation merge set, the exclusion set, and the candidate set to be inferred.

[0035] Optionally, the two-stage progressive large language model reasoning includes:

[0036] Based on the candidate set to be reasoned, the attribute feature set of each candidate entity pair in the candidate set to be reasoned is constructed as an attribute-aware prompt, which is input into the large language model to perform multiple rounds of parallel reasoning. Each candidate entity pair is aligned by majority voting. Candidate entity pairs that are aligned are included in the first stage confirmed alignment set, and candidate entity pairs that are not aligned are included in the second stage candidate set to be reasoned, thus obtaining the first stage confirmed alignment set and the second stage candidate set to be reasoned.

[0037] Based on the candidate set to be reasoned in the second stage, the relation feature set of each candidate entity pair in the candidate set to be reasoned in the second stage is constructed as a relation-aware prompt. The prompt is input into the large language model to perform parallel reasoning in the same number of rounds. Each candidate entity pair is initially aligned using the same majority voting method. The voting results are then validated using the cross-domain mapping relation set in the hierarchical heterogeneous ontology framework to obtain the second stage supplementary alignment result set.

[0038] Based on the first-stage confirmed alignment set and the second-stage supplementary alignment result set, the first-stage confirmed alignment set, the confirmed merged set, and the second-stage supplementary alignment result set are merged to obtain the cross-domain entity disambiguation alignment result set.

[0039] Optionally, the construction of a multi-granularity temporal hypergraph and the weighted integration of temporal information from each granularity layer via a cross-scale conflict sensing mechanism includes:

[0040] Based on the enhanced fusion triplet set and the semantic vectors of entities in each domain, an annual superlayer, a quarterly superlayer, and a citation-instantaneous superlayer are constructed according to three time granularities: year, quarter, and citation-instantaneous. Within each of the annual, quarterly, and citation-instantaneous superlayers, when the cosine similarity of the semantic vectors between a scientific entity and a technical entity exceeds a preset hyperedge construction threshold, and the document timestamps of the scientific entity and the technical entity both belong to the same time granularity interval, a hyperedge connection is established between the scientific entity and the technical entity within the superlayer, resulting in a multi-granularity temporal supergraph containing the annual, quarterly, and citation-instantaneous superlayers.

[0041] Based on the multi-granularity temporal hypergraph, it detects whether there are conflicts in the relationship annotations of the same entity pair among the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. It counts the number of conflicts for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. It calculates the corresponding fusion weight based on the number of conflicts for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. The fusion weight is inversely proportional to the number of conflicts for the corresponding hyperlayer. The fusion weights for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer are then weighted and fused with the temporal information of the corresponding hyperlayer to obtain the multi-granularity temporal attribute vector for each knowledge node.

[0042] This invention also provides a heterogeneous science and technology knowledge graph construction system, comprising:

[0043] The ontology framework construction module is used to analyze multi-source scientific and technological literature and domain expert interview data. It formally defines and verifies the differences in knowledge structure between scientific papers and technical patents through ontology modeling methods, and constructs a hierarchical heterogeneous ontology framework that includes scientific knowledge sub-ontologies, technical knowledge sub-ontologies and cross-domain mapping relationship sets.

[0044] The triplet extraction module is used to take the hierarchical heterogeneous ontology framework and the multi-source scientific and technological literature database as input, construct an entity-relation joint extraction framework through domain fine-tuning of the large language model, perform dual-channel retrieval enhanced triplet extraction in batch iteratively, and merge all batch results to obtain scientific knowledge triplet sets and technical knowledge triplet sets.

[0045] The cross-domain disambiguation alignment module is used to map entities and relations to a unified semantic vector space based on the scientific knowledge triple set and the technical knowledge triple set through graph knowledge representation learning. After cross-domain entity disambiguation alignment is completed through dual-threshold pre-filtering and two-stage progressive large language model inference, and after adding document metadata, an enhanced fusion triple set is obtained.

[0046] The knowledge graph persistence module is used to construct a multi-granularity temporal hypergraph based on the enhanced fusion triplet set and semantic vectors of entities in each domain, and to integrate the temporal information of each granularity layer through a cross-scale conflict perception mechanism. The resulting heterogeneous science-technology semantic knowledge graph is then persistently stored in a graph database to obtain a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

[0047] The present invention has the following beneficial technical effects:

[0048] 1. Addressing the problem of cross-source heterogeneous knowledge fusion: By providing a unified type constraint foundation through a hierarchical heterogeneous ontology framework, combined with the entity-relation joint extraction framework for the synchronous processing of two types of documents, and the combination mechanism of dual-channel retrieval enhancement and large language model illusion filtering, the ability to construct high-quality heterogeneous knowledge triples from scientific papers and technical patents in an integrated manner is realized, which significantly improves the problem of the separation of scientific and technological knowledge.

[0049] 2. Regarding the entity disambiguation problem: The dual-threshold pre-filtering mechanism compresses the scope of multi-round voting of the large language model to an uncertain range. The progressive decision chain from attribute perception to relation perception, combined with the legality verification of ontology mapping rules, systematically solves the disambiguation error caused by synonyms and variants of scientific and technical terms, while maintaining the computational cost that is feasible for engineering implementation.

[0050] 3. Addressing the issue of temporal representation: The multi-granularity temporal hypergraph combined with the cross-scale conflict perception weight allocation mechanism enables the knowledge graph to simultaneously track the macro-annual trends of scientific discoveries and the rapid iteration path of technology patents, supporting accurate querying of multi-granularity scientific and technological knowledge evolution paths, which is significantly better than the static annotation scheme of a single timestamp.

[0051] 4. The system is scalable: The batch iterative extraction architecture and intermediate incremental graph structure design enable the system to continuously update the graph as the literature database expands without the need for full reconstruction; the multi-granularity time index of the graph database supports efficient temporal queries and has engineering feasibility for deployment in industrial-grade scientific and technological intelligence platforms. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments 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.

[0053] Figure 1 This is a flowchart illustrating the heterogeneous science and technology knowledge graph construction method of the present invention;

[0054] Figure 2 This is a schematic diagram illustrating the process of constructing the hierarchical heterogeneous ontology framework of the present invention;

[0055] Figure 3 This is a schematic diagram of the heterogeneous science and technology knowledge graph construction system of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0057] Example 1

[0058] like Figure 1 As shown, this embodiment provides a method for constructing a heterogeneous science and technology knowledge graph, including:

[0059] Step S1: Analyze multi-source scientific and technological literature and domain expert interview data, and formally define and verify the differences in knowledge structure between scientific papers and technical patents through ontology modeling methods. Construct a hierarchical heterogeneous ontology framework that includes scientific knowledge sub-ontologies, technical knowledge sub-ontologies and cross-domain mapping relationship sets.

[0060] Specifically, the multi-source scientific and technological literature includes collections of academic papers and patent documents. The data from domain expert interviews comes from experts with relevant research backgrounds who provide systematic descriptions of the differences between scientific papers and technological patents in terms of knowledge organization, entity granularity, and relationship types. Through comprehensive analysis of these two types of data, the differences in knowledge structure between scientific papers, which emphasize theoretical logic and causal reasoning, and technological patents, which emphasize functional implementation and application scenarios, are identified, providing conceptual input for subsequent ontology modeling. The ontology modeling method employs formal ontology engineering, using ontology modeling tools (such as Protégé) to formally define entity types, relationship types, and attribute constraints, and performing consistency checks to ensure the ontology is logically consistent. The constructed hierarchical heterogeneous ontology framework includes scientific knowledge sub-ontologies, technological knowledge sub-ontologies, and a set of cross-domain mapping relationships connecting the two sub-ontologies. This framework serves as the unified type constraint basis for subsequent steps in knowledge extraction, entity disambiguation, and graph construction.

[0061] Step S2: Using the hierarchical heterogeneous ontology framework and the multi-source scientific and technological literature database as input, construct an entity-relation joint extraction framework through domain fine-tuning of the large language model, perform dual-channel retrieval enhanced triple extraction in batches, and merge all batch results to obtain the scientific knowledge triple set and the technical knowledge triple set.

[0062] Specifically, using the entity type set and relation type set defined in the hierarchical heterogeneous ontology framework as type constraints, the pre-trained large language model is fine-tuned to enable it to simultaneously identify relationships between scientific entities, technical entities, and knowledge entities, thus constructing an entity-relation joint extraction framework. The multi-source scientific literature database is divided into several batches according to semantic themes. For each batch, a dual-channel retrieval enhanced triple extraction is performed: on the one hand, the cross-batch semantic context of pre-entities is obtained through a text corpus backtracking retrieval channel; on the other hand, the neighborhood relation information of pre-entities is obtained from the intermediate incremental graph through a graph structure retrieval channel. The two sets of information are jointly input into the large language model for refined extraction. The extraction results of each batch are merged into the intermediate incremental graph to achieve cross-batch knowledge accumulation. After all batches are processed, the sets of scientific knowledge triples and technical knowledge triples are merged.

[0063] It should be noted that dividing the multi-source scientific literature database into multiple batches based on semantic topics directly impacts the quality of subsequent dual-channel retrieval and triple extraction. If highly semantically related documents are grouped into the same batch, the entity co-occurrence frequency among documents within the same batch is higher. This allows the large language model to obtain richer semantic cues from the context within the batch during the initial extraction stage, helping to improve the candidate pre-entity set E. pre The quality of the document clustering is crucial; conversely, if the semantic topics of the documents within a batch are too scattered, the large language model is more likely to produce relation type errors in candidate triples without sufficient cross-document context support. In practice, batch division can adopt topic clustering methods based on document titles and abstracts (such as LDA topic models or semantic clustering based on pre-trained language models), grouping documents with semantic similarity higher than a preset batch clustering threshold into the same batch. At the same time, upper and lower limits are set for the number of documents in each batch to ensure that the batch size is within the single processing capacity of the large language model. The selection of the batch size K is related to the total size of the document database and the context window length of the deployed large language model, typically using the reference that the total number of tokens in a single batch does not exceed 80% of the model's effective context window.

[0064] Step S3: Based on the scientific knowledge triplet set and the technical knowledge triplet set, entities and relations are mapped to a unified semantic vector space through graph knowledge representation learning. Cross-domain entity disambiguation and alignment are completed through dual-threshold pre-filtering and two-stage progressive large language model inference. After adding document metadata, an enhanced fusion triplet set is obtained.

[0065] Specifically, based on scientific knowledge triplets and technical knowledge triplets, entities and relations are mapped to a unified high-dimensional semantic vector space using graph knowledge representation learning methods (such as RotatE), resulting in a scientific entity vector set and a technical entity vector set. Cosine similarity scores between cross-domain entity pairs are calculated, and candidate entity pairs are divided into three categories—confirmed merging set, excluded set, and candidate set to be inferred—through dual-threshold pre-filtering. Only the candidate set to be inferred undergoes a two-stage progressive large language model inference: the first stage performs attribute-aware inference based on attribute features, and the second stage performs relation-aware inference based on relation features, with legality verification performed using the cross-domain mapped relation set, completing cross-domain entity disambiguation and alignment. Document citation relationships are transformed into association edges between knowledge nodes, and document metadata (timestamp, institution, domain tag) is appended as node attributes, resulting in an enhanced fusion triplet set.

[0066] Step S4: Based on the enhanced fusion triplet set and semantic vectors of entities in each domain, construct a multi-granularity temporal hypergraph and integrate the temporal information of each granularity layer through a cross-scale conflict perception mechanism. Persistently store the resulting heterogeneous science-technology semantic knowledge graph in the graph database to obtain a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

[0067] Specifically, based on enhanced fusion triplet sets and semantic vectors of entities in each domain, annual, quarterly, and citation-instantaneous superlayers are constructed according to three time granularities: year, quarter, and citation-instantaneous, forming a multi-granularity temporal supergraph. A cross-scale conflict awareness mechanism is used to detect the consistency of relational annotations for the same entity pair across different granularity layers. Fusion weights for each granularity layer are calculated based on the number of conflicts; the fusion weights are inversely proportional to the number of conflicts in the corresponding superlayer. The temporal information of each granularity layer is weighted and fused using these fusion weights to obtain a multi-granularity temporal attribute vector for each knowledge node, forming a heterogeneous science-technology semantic knowledge graph. The graph data is formatted into graph database query statements (such as Cypher statements) and imported in batches into a graph database (such as Neo4j). Node type indexes, relation type indexes, and time granularity indexes are established to obtain a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

[0068] Example 2

[0069] like Figure 2 As shown, in this embodiment, multi-source scientific literature and domain expert interview data are analyzed. The differences in knowledge structure between scientific papers and technological patents are formally defined and their consistency is checked using ontology modeling methods. A hierarchical heterogeneous ontology framework is constructed, comprising scientific knowledge sub-ontologies, technological knowledge sub-ontologies, and a cross-domain mapping relationship set, including:

[0070] Step S11: Based on multi-source scientific and technological literature and domain expert interview data, sort out the differences in knowledge structure between scientific papers and technical patents, and obtain a draft of the scientific-technical knowledge domain concept system that includes a set of entity types and a set of relation types.

[0071] Specifically, multi-source scientific and technological literature includes collections of papers and patents from different databases and disciplines. Domain expert interview data is obtained through structured interviews with experts possessing relevant research backgrounds. By systematically analyzing the differences between scientific papers and technological patents in terms of knowledge organization, entity granularity, and relational semantics, a set of entity types covering both types of literature (denoted as E) is summarized. raw ) and the set of relation types (denoted as R) raw This will lead to the drafting of a conceptual framework for the science and technology knowledge domain. raw The entity types in R cover the research subject, research object, experimental materials, etc. on the scientific paper side, as well as the technical means, functional objectives, application carriers, etc. on the technology patent side; raw The types of relationships in the database include causal relationships, dependency relationships, realization relationships, and application relationships.

[0072] Step S12: Based on the draft concept system of the scientific and technological knowledge domain, formally define the entity type set and the relation type set and supplement attribute constraints using an ontology modeling tool, perform consistency checks, and obtain an empirical scientific and technological knowledge ontology containing the entity type set, relation type set and attribute constraint set.

[0073] Specifically, the draft concept system (E) raw R raw Using E as input, the entity type set E is modeled using ontology modeling tools (such as Protégé). raw With relation type set R raw Formal definitions are performed, representing each entity type and relation type as an ontology class (OWL class) or object property, and attribute constraints such as data type constraints, cardinality constraints, and range constraints are added to each entity type and relation type to form an attribute constraint set (denoted as P). ont After completing the formal definition, an ontology inference engine (such as HermiT or Pellet) is invoked to perform a consistency check, verifying that there are no logical contradictions or circular definitions among the type definitions in the ontology. Based on the check results, any inconsistencies are corrected, ultimately resulting in a set of entity types E. ont Relation type set R ont With attribute constraint set P ont Empirical evidence-based science-technology knowledge ontology O.

[0074] Step S13: Based on the entity type set and relation type set in the verified scientific and technological knowledge ontology, define a scientific knowledge sub-ontology containing research questions, research methods, and experimental conclusions, and a technical knowledge sub-ontology containing technical efficacy, application scenarios, and invention methods. Based on the attribute constraint set, establish a cross-domain mapping relationship set between the scientific knowledge sub-ontology and the technical knowledge sub-ontology to obtain a hierarchical heterogeneous ontology framework that supports cross-domain alignment.

[0075] Specifically, based on the entity type set E in the validated science-technology knowledge ontology O. ont With relation type set R ont It is divided into two dedicated sub-ontologies: the scientific knowledge sub-ontology O. sci This includes definitions of entity types and their relation types for scientific papers, such as research questions, research methods, and experimental conclusions; and the technical knowledge sub-ontology O. tec This includes definitions of entity types and their relationships for technology patents, encompassing aspects such as technical efficacy, application scenarios, and inventive methods. Based on the attribute constraint set P... ont In O sci With O tec Establish a cross-domain mapping relationship set M between them ontM ont Each mapping rule explicitly describes the semantic equivalence or inclusion relationship between scientific entity types and technical entity types, such as the correspondence between the "research question" type and the "technical efficacy" type, and the association between the "research method" type and the "inventive means" type. (O) sci O tec M ont Together, they form a hierarchical heterogeneous ontology framework that supports cross-domain alignment, serving as a unified constraint basis for subsequent steps such as entity extraction, relation extraction, and cross-domain disambiguation. Meanwhile, M... ont It is used as a legality verification layer in the relation-aware reasoning stage.

[0076] A hierarchical heterogeneous ontology framework is an ontology engineering product that formally represents the knowledge concept systems of two types of documents—scientific papers and technical patents—using a hierarchical organizational structure. Essentially, it is a two-layer ontology structure that preserves the unique knowledge expression of each type of document while maintaining unified semantic constraints. Unlike a single-layer unified ontology (which indiscriminately mixes and defines scientific and technical entity types within the same flat ontology structure), the hierarchical heterogeneous ontology framework explicitly separates scientific knowledge sub-ontologies O... sci With technical knowledge sub-ontology O tec They are placed in relatively independent domains, and then mapped to a cross-domain mapping set M. ont As a semantic bridge layer connecting the two sub-ontologies, a two-layer architecture of "independent definition - explicit bridging" is formed. The core value of this layered design lies in the following: In the knowledge extraction stage, each sub-ontology provides entity type labels and relation type constraints specific to its document type, enabling the large language model to perform extraction at a semantic granularity that conforms to the knowledge expression habits of various document types, without having to accommodate the expression differences between the two types of documents under the same constraint system; In the cross-domain disambiguation stage, M... ont The explicitly defined cross-domain mapping rules serve directly as a legality verification layer, providing ontology-level semantic constraint filtering for the reasoning results of large language models.

[0077] The technical principle behind the superiority of a hierarchical heterogeneous ontology framework over a single-layer unified ontology is as follows: In scientific papers, the granularity of knowledge entities tends to be at the conceptual level (e.g., "research question" typically describes a scientific hypothesis to be verified, with abstract and broad semantic connotations); in technical patents, the granularity of knowledge entities tends to be at the functional implementation level (e.g., "inventive means" typically describes a specific technical operation process, with specific and clear semantic connotations). Forcibly unifying these two types of entity types into a single-layer ontology requires introducing an artificial, mandatory equivalence assumption between entity granularities, leading to type constraint mismatch—labeling "inventive means" in a patent with the granularity of "research question" will result in overgeneralization, while labeling "research question" in a paper with the granularity of "inventive means" will result in semantic fragmentation. The hierarchical design maintains type granularity self-consistency within each sub-ontology, in M... ont The layer uses weak constraint relationships such as semantic equivalence or semantic support to perform cross-layer mapping, which fundamentally avoids the above-mentioned granularity mismatch problem.

[0078] In practical applications, the three-layer structure of the hierarchical heterogeneous ontology framework (O sci O tec M ont ) plays different constraining roles at different stages of the overall knowledge graph construction process: In the knowledge extraction stage, O sci With O tec These serve as type constraint prompts for the extraction tasks of academic papers and patent documents, respectively, ensuring that the type labels of the extracted results do not exceed the definition range of each sub-ontology; during the entity disambiguation stage, M... ont In the second stage of the two-stage progressive reasoning process (relationship-aware reasoning), it serves as a legality verification layer, filtering cross-type misalignment; in the knowledge graph construction stage, O sci With O tec The type labels serve as entity affiliation identifiers for hyper-edge connections within the three hyper-layers: year, quarter, and reference time, respectively, supporting temporal queries and evolutionary analysis based on sub-ontology types.

[0079] Taking the processing of scientific and technological literature in the field of artificial intelligence as an example, this illustrates the practical effect of a hierarchical heterogeneous ontology framework: Assume a multi-source scientific and technological literature database contains an academic paper on "attention mechanisms" (including entities "self-attention," the relation "improved," and the entity "sequence modeling accuracy") and a patent document on "a Transformer-based natural language processing system" (including entities "multi-head attention module," the relation "implemented," and the entity "machine translation task"). Under the hierarchical heterogeneous ontology framework, the "self-attention" in the paper is... sci The "multi-head attention module" in the patent, labeled as an entity of type "research method," is O. tec Entities labeled as "inventive means", M ontThe existence of mapping rules ("research methods", "refinement", "inventive means") enables the cross-domain disambiguation stage to identify the semantic inheritance relationship between the two, thereby establishing an evolutionary path connection from the "self-attention" of scientific discovery to the "multi-head attention module" of technological application in the knowledge graph, effectively revealing the knowledge dissemination trajectory of the transformation of scientific research results into engineering technology.

[0080] Among them, the cross-domain mapping relationship set M ont Each mapping rule in the code is represented as a triple (scientific entity type t) s Mapping relationship type m, technical entity type t t The mapping relationship type m takes values ​​from a predefined mapping relationship vocabulary, including semantically equivalent to (meaning that two types of entities are interchangeable at the semantic level, such as the scientific entity type "research problem" and the technical entity type "technical efficacy" being equivalent at a certain semantic granularity), refined to (meaning that the technical entity type is a specific instance of the scientific entity type, such as "experimental conclusion" being refined to "inventive means" in the patent field), and supporting (meaning that the scientific entity type provides a theoretical basis for the technical entity type, such as "research method" supporting the proposal of "inventive means") and other semantic mapping categories. ont The completeness of M is guaranteed through an iterative review mechanism with domain experts: during the initial establishment of M... ont Then, a certain proportion of cross-domain entity disambiguation samples are randomly selected, and domain experts judge the disambiguation results for factors M. ont Validation failures due to missing rules will be identified, and the corresponding mapping rules will be added to M. ont After several rounds of iterative review, M ont To achieve sufficient coverage of cross-domain mapping scenarios within the domain of the literature being processed.

[0081] M ont In the validity verification phase of relation-aware reasoning, the usage is as follows: Candidate entity pairs (e) that are voted as "aligned" by the large language model. s e t ), with e s entity type t s With e t entity type t t As the search key, in M ont In the query, does it exist with the t? s and t t Mapping rules for endpoints (t) s ,m,t t If the query result is not empty, the candidate entity pair passes the validity check and is retained as the aligned result; if the query result is empty, it indicates that e s With e tThere is no predefined cross-domain mapping semantic relationship between the entity types in the ontology framework. The alignment judgment of the large language model contradicts the ontology constraints. The candidate entity pair is removed from the alignment results to prevent the erroneous merging of cross-type entities.

[0082] Example 3

[0083] In this embodiment, an entity-relation joint extraction framework is constructed by fine-tuning a large language model in the domain, including:

[0084] Based on the hierarchical heterogeneous ontology framework and the multi-source scientific and technological literature database, the knowledge entities and the relationships between knowledge entities in the text of the multi-source scientific and technological literature database are manually annotated, and the multi-source scientific and technological literature database is divided into multiple batches according to semantic topics to obtain an annotated sample set and a batched literature index table.

[0085] Specifically, using a hierarchical heterogeneous ontology framework (O sci O tec M ont The entity type set E defined in ) ont With relation type set R ont As a standard for annotation, it is compiled by domain experts for multi-source scientific and technological literature databases (including paper collections D). sci With Patent Document Collection D tec The text fragments in the document are manually annotated, including the knowledge entities appearing in the text and their respective entity types, the relationships between knowledge entities and their respective relationship types, forming a high-quality annotated sample set S. ann S ann Each labeled record contains the original text fragment, entity boundary annotations, entity type labels, and relation type labels, covering both scientific and technical entities. Simultaneously, the entire literature database from multiple sources is divided into K batches {Batch1, Batch2, ..., Batch...} according to semantic themes (e.g., research direction, technology category). K Generate batched literature index table I batch I batch Record the list of document identifiers contained in each batch and their processing order for use in subsequent iterations.

[0086] Based on the labeled sample set, the entity type set and relation type set defined in the hierarchical heterogeneous ontology framework are used as type constraints to fine-tune the pre-trained large language model in the domain. The batch iterative extraction order of the fine-tuned large language model is set according to the batch literature index table to obtain an entity-relation joint extraction framework that can simultaneously identify the relationship between scientific entities, technical entities and knowledge entities.

[0087] Specifically, based on the labeled sample set, the entity type set and relation type set defined in the hierarchical heterogeneous ontology framework are used as type constraints to fine-tune the pre-trained large language model for domain purposes. The batch iteration extraction order of the fine-tuned large language model is then set according to the batched literature index table. The specific method for injecting type constraints is as follows: the entity type set E... ont All entity type names and relation type set R defined in [the document / context] ont All relation type names defined in the framework are concatenated into the input of each training sample in the form of structured prompts. This instructs the model to select only valid entity type labels and relation type labels from the predefined type set during the output phase, prohibiting the model from generating free labels that exceed the range of the type set. This constrains the output space of the model within the semantic range defined by the hierarchical heterogeneous ontology framework.

[0088] During the supervised fine-tuning phase, the input of each training sample consists of three parts: the first part is the type constraint cue area, which contains E ont With R ont The first part is a complete list of types and a brief semantic description of each type; the second part is the text area to be extracted, containing data from the labeled sample set S. ann The first part is the original scientific and technological literature text fragment; the second part is the task instruction area, which explicitly instructs the model to synchronously output the entities appearing in the text and their type labels, the relationships between entity pairs and their relationship type labels using a joint extraction method. The corresponding training objective output is the labeled sample set S. ann The system already includes the labeled entity boundaries, entity type labels, and relation type label sequences. Through this method, the fine-tuned large language model can automatically identify scientific and technical entities simultaneously, and jointly extract the relationships between the two types of entities, based on the type system defined by the hierarchical heterogeneous ontology framework, when processing unseen text, without needing to switch between different extraction models for academic papers and patent documents.

[0089] Using labeled sample set S ann For training data, the entity type set E defined in the hierarchical heterogeneous ontology framework is used. ont With relation type set R ont Input is embedded as prompts to serve as type constraints for the model's output, enabling domain-specific fine-tuning of the pre-trained large language model. The fine-tuning process employs a supervised fine-tuning (SFT) strategy, allowing the model to learn to simultaneously identify scientific entities (O) from scientific literature texts under given type constraints. sci (within the scope) and technical entities (O) tec This method leverages the joint extraction capability of data within a given scope and the relationships between them, avoiding the limitations of traditional methods that model papers and patents separately. After fine-tuning, it is based on the batched literature index table I.batch Set the batch iteration extraction order for the fine-tuned large language model (from Batch1 to Batch2). K (Executing in sequence) yields the entity-relationship joint extraction framework F. ext F ext This serves as the core extraction engine for subsequent dual-channel retrieval and enhanced triple extraction.

[0090] Furthermore, the specific training settings for domain fine-tuning of the pre-trained large language model are as follows: In terms of base model selection, pre-trained large language models with strong instruction following ability and structured output ability are preferred to ensure that the fine-tuned model can stably output content that conforms to the entity type set E. ont With relation type set R ont The structured extraction results are constrained by format; regarding the loss function, the cross-entropy loss function is used in the fine-tuning stage to select the labeled sample set S. ann The golden label sequence (entity type label and relation type label sequence) in the model is used as the supervision target to optimize the conditional probability of the model generating correct label output under given type constraint cues.

[0091] It is worth noting that, due to the labeled sample set S ann Sourced from Paper Collection D sci With Patent Document Collection D tec The two types of heterogeneous documents differ significantly in writing style, terminology density, and knowledge expression patterns. If they are directly mixed for training without balancing, the model may overfit to the type of document with the larger sample size. Therefore, when constructing the training batch, S should be considered. ann From D sci Annotated samples and those from D tec The labeled samples are sampled proportionally to ensure that the ratio of the two types of documents in each training batch remains within a reasonable range (refer to the actual ratio of the two types of documents in the multi-source scientific literature database), so that the fine-tuned entity-relation joint extraction framework F... ext It exhibits stable extraction performance when processing both scientific papers and technical patents, rather than performing well only on a particular type of document. After fine-tuning, it can be extracted from the retention set (from S). ann On the proportionally partitioned validation set, F ext The model's precision and recall rates for extracting scientific entities, technical entities, and their corresponding relationship types were evaluated separately, serving as a quantitative assessment of the model's quality.

[0092] The entity-relation joint extraction framework refers to an extraction architecture that merges named entity recognition and relation extraction tasks into a single model for a unified reasoning process. Unlike the traditional two-step pipeline method of "first identifying entities, then identifying relations", the joint extraction framework is driven by the generative capabilities of a large language model. Under the condition of given type constraint prompts, it outputs a complete structured extraction result containing entity boundaries, entity type labels, and relation type labels in one go from the input text. Entity recognition and relation recognition are completed collaboratively in the same forward reasoning process.

[0093] The joint extraction framework's technical advantages over the traditional two-step pipeline method are reflected in the following two aspects: First, in scientific literature processing scenarios, the correct identification of entities often depends on the relational context in which they participate (for example, the word "mechanism" refers to entity types of different granularities in the relational contexts of "attention mechanism" and "regulation mechanism"). Traditional pipeline methods cannot utilize relational context information in the entity identification stage, resulting in a high error rate in entity type labeling for synonymous and heterogeneous terms. The joint extraction framework, by simultaneously modeling entities and relations in a single inference, enables the model to disambiguate entity types using relational context, significantly reducing the entity identification error rate. Second, when processing cross-source heterogeneous literature (scientific papers and technical patents), traditional methods typically establish independent extraction models for the two types of literature, resulting in incompatibility between the extraction results produced by the two models in terms of entity granularity and relational semantics, requiring additional post-processing alignment steps. The joint extraction framework, through a unified type set E defined in a hierarchical heterogeneous ontology framework... ont With R ont As a type constraint, the extraction tasks of two types of documents are processed simultaneously within a single model. The output scientific entities and technical entities naturally share the same type label space, without the need for additional alignment processing.

[0094] In practical applications, the entity-relationship joint extraction framework F... extThe output format for a scientific paper (e.g., "This study proposes a molecular property prediction method based on graph neural networks, and the effectiveness of the method has been verified on multiple benchmark datasets") is as follows: output scientific entities "graph neural network" (type: research method), scientific entities "molecular property prediction" (type: research question), scientific entities "benchmark datasets" (type: experimental materials), and relational triples ("graph neural network", "applied to", "molecular property prediction") and ("graph neural network", "verified on", "benchmark datasets"). The output format for a technical patent (e.g., "This invention achieves rapid screening of the bioactivity of candidate compounds by introducing a graph convolution module to extract features from the molecular graph structure") is as follows: output technical entities "graph convolution module" (type: invention means), technical entities "molecular graph structure" (type: application scenario), technical entities "candidate compound bioactivity screening" (type: technical efficacy), and relational triples ("graph convolution module", "achieved", "candidate compound bioactivity screening"). The extraction results of the two texts are generated within the same framework, and their entity type labels are all from the unified type set of the hierarchical heterogeneous ontology framework, which provides a direct input basis for subsequent cross-domain entity disambiguation alignment (identifying the semantic inheritance relationship between "graph neural network" and "graph convolution module").

[0095] Example 4

[0096] In this embodiment, batch iterative execution of dual-channel retrieval enhanced triple extraction includes:

[0097] Based on the entity-relation joint extraction framework, the current batch of documents is initially extracted to obtain a candidate pre-entity set and a candidate triple set, and the intermediate incremental graph up to the current batch is loaded.

[0098] Specifically, for the current batch of documents... i (i starts from 1, based on the batched literature index table I) batch (Execute sequentially), invoking the entity-relationship joint extraction framework F. ext Perform preliminary extraction on all text within the batch and output candidate pre-entity set E. pre (i) (including candidates for scientific and technological entities) and candidate triplet set T pre (i) (including the head entity, relation type, and tail entity of candidate triples). Simultaneously, load the intermediate incremental graph G saved up to the completion of the current batch i-1. inc (When i=1, G) inc For an empty hypergraph, the system operates in single-channel mode, and the graph structure retrieval channel returns an empty set. G incIt is stored in a hypergraph data structure, where nodes are entities confirmed in historical batches and edges are confirmed triple relationships. It supports querying neighborhood relationship information by entity identifier.

[0099] Based on the candidate pre-entity set and the intermediate incremental graph, each pre-entity in the candidate pre-entity set is used as a search query term to retrieve cross-batch text fragments in the full document library of multi-source scientific and technological literature, thereby obtaining the semantic context corresponding to each pre-entity.

[0100] Specifically, with E pre (i) The entity name or identifier of each candidate pre-entity is used as the search query term. Cross-batch text retrieval is performed in the full document library of the multi-source scientific and technological literature database (including all documents in the current batch and all documents in the historical batch). The retrieval method adopts a hybrid strategy of semantic retrieval (approximate nearest neighbor retrieval based on dense vector representation) and keyword retrieval (based on sparse retrieval algorithms such as BM25) to obtain the cross-batch text fragments of each pre-entity appearing in the full document library, forming the semantic context C corresponding to each pre-entity. text C text It includes contextual descriptions of pre-entities in different batches of documents, used to supplement the cross-document semantic information missing in the current batch of text.

[0101] The fusion method of semantic retrieval and keyword retrieval results in the hybrid retrieval strategy is as follows: For each candidate pre-entity E pre The semantic retrieval channel is based on E pre The dense vector representation (encoding entity names or identifiers into high-dimensional vectors through a pre-trained text encoder) performs approximate nearest neighbor retrieval in a pre-built vector index of the entire document, returning the text segments with the highest semantic similarity, each accompanied by its semantic similarity score; the keyword retrieval channel is based on E pre The entity name or its known aliases (if an alias mapping table exists, aliases are retrieved synchronously) are used to perform a lexical matching search in the inverted index of the entire document using sparse search algorithms such as BM25. This returns the text fragments with the highest lexical relevance, each accompanied by its BM25 relevance score. The two search results are then unified and merged using the Reciprocal Rank Fusion (RRF) method. The text fragments with the highest rank after fusion are selected as candidate pre-entity E. pre The corresponding semantic context C text .

[0102] It should be understood that the advantage of the semantic retrieval channel lies in its ability to recall text fragments that are semantically similar to the pre-entity but have different lexical expressions (such as abbreviations and full names, synonyms, etc.), while the advantage of the keyword retrieval channel lies in its high precision in accurately matching the pre-entity name. The fusion of the two retrieval results enhances the semantic context C.text It combines the advantages of broad semantic coverage and high vocabulary matching accuracy, providing more comprehensive cross-document knowledge clues for subsequent refinement extraction by large language models. The number of text fragments retrieved by each pre-entity can be dynamically adjusted according to the context window length of the deployed large language model, ensuring that the total number of tokens in the dual-channel augmented context is within the effective processing range of the model.

[0103] Based on the candidate pre-entity set and the intermediate incremental graph, the neighborhood relationship information of each pre-entity is retrieved from the intermediate incremental graph to obtain the graph structure information corresponding to each pre-entity.

[0104] Specifically, with E pre In (i), the entity identifier of each candidate pre-entity is the retrieval key, and in the intermediate incremental graph G inc The system queries the one-hop and multi-hop neighborhood relationship information of the pre-entity, including the set of neighboring entities directly connected to the pre-entity, the connection relationship type, and the relationship confidence score, forming the graph structure information G corresponding to each pre-entity. struct When i=1, G inc For an empty graph, G struct An empty set is returned for all pre-entities. The system operates in single-channel mode for this batch, relying only on the semantic context C. text Further refinement will be carried out. As batch iterations progress, G... inc Gradually accumulate high-quality triplet sets from historical batches, G struct The graph structure information carried is gradually enriched, and the effect of cross-batch knowledge accumulation is gradually enhanced.

[0105] The semantic context and the graph structure information are combined according to the pre-entity alignment to obtain the dual-channel enhanced context corresponding to each pre-entity.

[0106] Specifically, for E pre For each candidate pre-entity in (i), its corresponding semantic context C text With graph structure information G struct Align and combine the identifiers of the pre-entities to form a dual-channel enhancement context (C) corresponding to each pre-entity. text G struct The alignment and combination method is as follows: using the pre-entity identifier as the key, C... text Cross-batch text fragments and G struct The neighborhood relation list in G is concatenated into a structured enhanced context representation for subsequent joint input into a large language model. struct When the set is empty, the dual-channel augmented context degenerates to contain only C. text Single-channel enhanced context.

[0107] The dual-channel enhanced context and the candidate triple set are jointly input into the large language model in the entity-relation joint extraction framework to obtain the refined triple set of the current batch. The refined triple set is then merged into the intermediate incremental graph to update the intermediate incremental graph. This process is iterated until all batches are processed. Finally, the refined triple sets of all batches are merged to obtain the scientific knowledge triple set and the technical knowledge triple set.

[0108] Specifically, the dual-channel enhancement context (C) corresponding to each pre-entity is... text G struct ) and candidate triple set T pre (i) Jointly constructing input prompts, input entity-relation joint extraction framework F ext The large language model in the middle, based on the dual-channel enhanced context, provides T pre (i) refines the head entity, relation type, and tail entity of each candidate triple, and outputs the refined triple set T for the current batch. batch (i). T batch All ternary combinations in (i) are merged into the intermediate increment graph G. inc Update G inc The node and edge sets provide a richer accumulation of historical knowledge for the next batch of graph structure retrieval. Based on the batched literature index table I... batch sequentially from Batch1 to Batch K Execute the above iterative process until all K batches have been processed. Finally, merge the refined triplet set {T} from all batches. batch (1), T batch (2), ..., T batch (K)}, according to the sub-ontology (O) to which the entity in the triple belongs. sci Or O tec By classifying and summarizing, we obtain the scientific knowledge triplet set T. sci With the technical knowledge triple set T tec Both triplet sets contain document metadata attributes (timestamp, source document identifier, domain tag, etc.).

[0109] Among them, the dual-channel retrieval-enhanced triple extraction mechanism refers to an iterative extraction architecture that uses candidate pre-entities as retrieval anchors, simultaneously triggers two information acquisition channels—text corpus backtracking retrieval and graph structure retrieval—and jointly inputs the dual-channel information into the large language model for triple refinement. Its essence is to introduce the external retrieval-enhanced generation (RAG) paradigm into the knowledge extraction process of the large language model to make up for the information limitations of the large language model when processing cross-document and cross-batch knowledge associations.

[0110] The text corpus backtracking retrieval channel and the graph structure retrieval channel are complementary in terms of information dimension. Their technical principles are as follows: Text corpus backtracking retrieval channel (C... text This provides information on the semantic diversity of entities in different document contexts, revealing instances where the same entity expresses the same semantic connotation in different documents using different wording, thus providing rich semantic ambiguity resolution basis for large language models; Graph structure retrieval channel (G struct This provides the structural location information of an entity within the confirmed knowledge network of historical batches, namely, the combination of the entity's known neighbor entity types and relation types. This reveals the semantic role the entity plays in the knowledge graph topology (e.g., if an entity consistently appears as a neighbor of "research methods" in the known knowledge network, the relation type extraction for that entity in the new batch can refer to this structural prior), providing the large language model with a basis for relation type inference based on the knowledge graph topology. The two channels supplement the large language model with cross-document and cross-batch knowledge information missing in the current batch of text from two independent dimensions: "textual semantics" and "graph structure," respectively, jointly constituting a dual verification mechanism for candidate triple relation types.

[0111] Intermediate Incremental Plot G inc The enhancement effect in the iterative extraction process is reflected in the gradual quality improvement as the batch progresses: in the first batch (i=1), G inc If the graph is empty, the graph structure retrieval channel returns an empty set. The system operates in single-channel mode, relying solely on information from the text corpus backtracking retrieval channel for refined extraction. At this stage, the extraction quality depends on the richness of semantic information within the first batch of documents. As the batch progresses, G... inc By accumulating high-quality triples filtered through illusions in batches, the graph structure retrieval channel can provide increasingly richer prior neighborhood relationships for pre-entities in subsequent batches, resulting in a systematic improvement in extraction accuracy for later batches compared to earlier batches. This incremental learning effect enables the overall extraction architecture to adaptively improve as the document database expands, without requiring a full reprocessing of the extracted results when new documents are added.

[0112] The LLM-Judge illusion filtering mechanism acts as a quality gatekeeper in dual-channel retrieval augmented extraction. While large language models can improve the accuracy of relation type judgment by utilizing dual-channel augmented context during the refinement extraction stage, they struggle in situations where there are semantic contradictions between the dual-channel information and candidate triples (such as semantic context C). text With graph structure information G structEven when conflicting semantic cues are given for the same relation, large language models may still generate illusory triples that lack direct support from the original text. By performing independent semantic consistency scoring on each candidate triple in the refined triple set, the illusion filtering mechanism can filter triples with semantic consistency confidence scores below a preset threshold θ. conf The candidate triplet from the G inc Remove from the results to prevent illusory knowledge from being transmitted through G during the iteration process. inc This could cause contamination and spread to subsequent batches of graph structure retrieval channels, ensuring G... inc Maintain a consistently high-quality knowledge base. In practical applications, θ conf The settings should be adjusted to filter precision (to prevent illusory knowledge from entering the G-factor). inc A balance can be struck between θ and filtering recall (avoiding the misfiltering of valid triples with implicit semantic support), which can be achieved by evaluating different θ values ​​on an labeled validation set. conf The optimal threshold is determined by the precision and recall curves under the given values.

[0113] Example 5

[0114] In this embodiment, the hallucination filtering in dual-channel retrieval enhanced triple extraction includes:

[0115] Based on each candidate triple in the refined triple set and the original text of the current batch of documents, the candidate triple and the original text are used to construct a verification prompt, which is then input into a large language model. The large language model scores the semantic consistency between the candidate triple and the original text to obtain the semantic consistency confidence score for each candidate triple.

[0116] Specifically, for the refined triplet set T batch For each candidate triple (head entity, relation type, tail entity) in (i), match the triple with the current batch of documents. i The text fragments corresponding to the original text are jointly constructed to form a verification prompt. The format of the verification prompt is as follows: the semantic content of the candidate triple is expressed in natural language form, and the original text fragment is attached. The large language model is required to judge whether the knowledge content described by the triple is consistent with the semantics of the original text, and output a normalized confidence score (within the range of [0, 1]). The large language model independently scores the semantic consistency between each candidate triple and the original text based on the verification prompt, and obtains the semantic consistency confidence score corresponding to each candidate triple. conf This score reflects the confidence level of the large language model in determining whether the triple is a phantom triple (i.e., a triple generated by the model out of thin air and without semantic support from the original text).

[0117] The verification hint is constructed by combining candidate triples with the original text, and the specific construction rules are as follows: The verification hint is composed of three semantic blocks concatenated sequentially. The first block is the triple statement area, which expresses the candidate triples (head entity, relation type, tail entity) in the form of natural language statements. For example, the triples ("attention mechanism", "applied to", "text classification") are transformed into "attention mechanism is applied to text classification tasks". The second block is the original text evidence area, which is derived from the current batch of documents. i The first section extracts the original text fragments related to the head or tail entity as the basis for semantic consistency judgment; the second section is the scoring instruction area, which requires the large language model to output a confidence score with a continuous value between 0 and 1. The higher the score, the more semantically consistent the knowledge content of the triple is with the original text, and the lower the score, the more likely the triple is a phantom triple generated by the model out of thin air.

[0118] It should be noted that the semantic consistency confidence score is... conf This reflects the comprehensive judgment of the large language model on the authenticity of candidate triples based on verification prompts, rather than simply the word overlap rate. During the judgment process, the large language model can identify implicit semantic support in the text; that is, although the relationship described by the candidate triples does not appear in the original text in the exact same way, its semantic essence is supported by the contextual logic of the original text. Such candidate triples can still obtain a high score. conf This avoids false filtering due to diversity of expressions. The preset confidence threshold θ is used to achieve this. conf The illusion filtering mechanism will remove candidate triples that lack semantic support from the original text from the refined triples set, ensuring that they are incorporated into the intermediate incremental graph G. inc All triples have traceable literature sources, preventing the cumulative spread of illusory knowledge during cross-batch iteration.

[0119] Based on the semantic consistency confidence score, candidate triples below the preset confidence threshold are filtered out from the refined triple set to obtain the filtered triple set for the current batch. The filtered triple set is then merged into the intermediate incremental graph to update the intermediate incremental graph. This process is iterated until all batches are processed. Finally, the filtered triple sets from all batches are merged to obtain the scientific knowledge triple set and the technical knowledge triple set.

[0120] Specifically, let the pre-set confidence threshold be θ. conf , for T batch For each candidate triple in (i), if its semantic consistency confidence score is... conf Below θ conf Then the triple is determined to be an illusion triple, from T batch(i) filter out all scores; conf Not less than θ conf The candidate triples are retained to form the filtered triple set T for the current batch. filter (i). T filter (i) Merge into intermediate increment graph G inc To update G inc This provides high-quality knowledge accumulation filtered by illusion for the next batch of graph structure retrieval, avoiding the influence of illusion triples on G during the iteration process. inc Pollution. Based on batched literature index table I batch The above illusion filtering and G are performed sequentially on all K batches. inc Update the process until all batches are processed. Finally, merge the filtered triples {T} from all batches. filter (1), T filter (2), ..., T filter (K)}, categorized and summarized according to the sub-ontologies to which the entities belong, yields the scientific knowledge triplet set T. sci With the technical knowledge triple set T tec .

[0121] Example 6

[0122] In this embodiment, entities and relations are mapped to a unified semantic vector space through graph knowledge representation learning, including:

[0123] Based on the scientific knowledge triple set and the technical knowledge triple set, graph knowledge representation learning is performed on entities and relations, mapping the entities and relations in the scientific knowledge triple set and the technical knowledge triple set to a unified semantic vector space, resulting in a scientific entity vector set and a technical entity vector set.

[0124] Specifically, using the scientific knowledge triplet T sci With the technical knowledge triple set T tec The union of T is used as the training input for graph knowledge representation learning. Graph knowledge representation learning methods (such as RotatE, TransE, or ComplEx) are then applied to T. sci ∪T tec The system performs vectorization representation learning on all entities and relations, mapping them to low-dimensional dense vectors in the same high-dimensional semantic vector space. This ensures that semantically similar entities are close in distance within the vector space, and relations with similar patterns are oriented similarly. After learning, the vector representations of scientific entities are extracted according to their respective sub-ontologies, and then aggregated into a scientific entity vector set V. sci Extract the vector representations of the technical entities and summarize them into a technical entity vector set V. tec V sci With V tecBoth belong to the same semantic vector space, supporting the direct calculation of cross-domain semantic similarity between scientific entities and technical entities.

[0125] It should be noted that the scientific knowledge triplet T sci With the technical knowledge triple set T tec The design of using the union of T as a unified training input for graph knowledge representation learning is motivated by leveraging the natural connectivity of shared entities and relations between two types of triplet sets to simultaneously learn vector representations of scientific and technical entities in a single vector space. sci With T tec Performing independent graph knowledge representation learning separately will result in independent vector spaces from the two training sessions. A subsequent linear alignment transformation alone cannot guarantee the accuracy of cross-domain semantic similarity calculation; however, using T... sci ∪T tec To unify the training input, in the cross-domain mapping relation set M ont Under the constraints, the aligned entity pairs (by M) ont The "semantic equivalence" mapping rule (as determined in the training process) allows for the sharing of gradient propagation paths, enabling semantically equivalent scientific and technical entities to converge to similar positions in a unified vector space. This provides a vector representation basis with cross-domain comparability for subsequent cosine similarity calculations.

[0126] The selection of graph knowledge representation learning methods should prioritize those that address relation diversity (including M). ont The expressive power of the cross-domain mapping relation types defined in the standard: The RotatE method represents relations as rotational transformations in complex vector space, effectively modeling various relation patterns such as symmetric, antisymmetric, and combinatorial, making it suitable for handling O(n) domain mappings. sci With O tec The TransE method defines multiple types of cross-domain mapping relationships; it assumes that relationships are vector translations, resulting in a simple structure but limited ability to express complex relationship schemas, making it suitable for scenarios with relatively simple relationship types. In practical implementation, the relationship type set R in the hierarchical heterogeneous ontology framework can be used as a basis. ont Choose an appropriate graph knowledge representation learning method based on the complexity, and then... sci ∪T tec The quality of the learned vector space is verified by evaluating link prediction metrics (such as MRR and Hits@N) on a set of validation triples.

[0127] Based on the scientific entity vector set and the technical entity vector set, the cosine similarity score between cross-domain entity pairs is calculated, and the top-ranked attribute feature set and relation feature set of each candidate entity pair are extracted to obtain the entity candidate set to be disambiguated. Each record in the entity candidate set to be disambiguated contains an entity pair, a cosine similarity score, an attribute feature set, and a relation feature set.

[0128] Specifically, for V sci Each scientific entity vector in the vector is related to V tec For each technical entity vector, calculate the cosine similarity score between them. The formula for calculating cosine similarity is score(e...). s , e t ) = (v s · v t ) / (‖v s ‖ · ‖v t ‖), where v s With v t They are scientific entities e s With technology entity e t The semantic vector. For entity pairs with a cosine similarity score exceeding the minimum candidate threshold, the semantic vector is calculated from T, unit by unit, for each candidate entity pair. sci With T tec Extract the top K (preset number) attribute features (entity attribute dimensions, such as entity type label, literature domain label, definition description text, etc.) of each entity in the entity pair to form an attribute feature set A. feat Simultaneously extract each entity in T sci or T tec The top K neighboring relationships in the graph (graph neighboring relationship dimension, such as the relationship type directly connected to the entity and the neighboring entity type) form a relationship feature set R. feat Finally, construct the candidate set P of entities to be disambiguated. cand P cand Each record contains: candidate entity pairs (e s e t ), cosine similarity score, attribute feature set A feat With relation feature set R feat Four items, while retaining V sci With V tec For use in subsequent hyperedge construction.

[0129] Example 7

[0130] In this embodiment, the dual-threshold pre-filtering includes:

[0131] Based on the candidate set of entities to be disambiguated, according to the cosine similarity score of each record in the candidate set, candidate entity pairs with cosine similarity scores not lower than a preset high threshold are assigned to the confirmation merge set, candidate entity pairs with cosine similarity scores lower than a preset low threshold are assigned to the exclusion set, and candidate entity pairs with cosine similarity scores between the preset low threshold and the preset high threshold are assigned to the candidate set to be inferred, thus obtaining the confirmation merge set, the exclusion set, and the candidate set to be inferred.

[0132] Specifically, let the preset high threshold be θ. high (Example value: 0.95), preset low threshold is θ low (Example value: 0.65) To disambiguate the candidate set P cand For each record, perform the following three-way classification based on the cosine similarity score:

[0133] (1) If score ≥ θ high If the candidate entity is determined to be highly semantically similar, it will be directly included in the Align confirmation set. direct Without the need for large language model reasoning, the entity pair can directly perform entity merging operations in subsequent steps;

[0134] (2) If score < θ low If the candidate entity pair is semantically significantly different, it will be directly included in the exclusion set Reject and excluded from the cross-domain alignment candidate range.

[0135] (3) If θ low ≤ score < θ high If the candidate entity pair is in the uncertainty interval, it is determined that it is included in the candidate set P to be reasoned. uncertain Then, proceed to the next two stages of progressive large language model reasoning process.

[0136] Through the above dual-threshold pre-filtering, P uncertain The size is approximately P cand Reducing the number of candidate pairs by 20% to 30% effectively compresses the subsequent large language model calls to a reasonable scale, thus reducing computational costs while ensuring alignment quality.

[0137] The design basis of the dual-threshold pre-filtering mechanism is as follows: In large-scale cross-domain entity disambiguation scenarios, the candidate set of entities to be disambiguated, P cand The number of candidate entity pairs increases quadratically with the size of the literature database. Calling a large language model to perform inference on all candidate pairs would result in unacceptable computational costs. The proposed dual-threshold pre-filtering mechanism utilizes the discriminative power of semantic vectors learned from graph knowledge representations to divide candidate entity pairs into three confidence intervals based on their cosine similarity scores: a high similarity interval (score ≥ θ). high This corresponds to entity pairs with highly overlapping semantics. These entity pairs are extremely close in the semantic vector space, providing a sufficient basis for direct merging. Low similarity intervals (score < θ) low ) Entity pairs with clearly different semantics are directly excluded without further reasoning; uncertain intervals (θ) low ≤score<θ highThis refers to entity pairs that are semantically similar but difficult to accurately judge based solely on vector similarity. These entity pairs contain a large number of synonymous but different terms in the field of science and technology (such as abbreviations and full names, the same concept expressed in different languages, etc.), and require deep reasoning by leveraging the semantic understanding capabilities of large language models.

[0138] After dual-threshold pre-filtering, the candidate set P to be inferred is entered. uncertain The size of the candidate entity pairs is approximately 20% to 30% of the total number of candidate entity pairs. Achieving this compression ratio relies on a preset high threshold θ. high With preset low threshold θ low Reasonable setting: θ high Setting it too high can force some actually aligned entity pairs into the inference stage, increasing unnecessary computation; θ low Setting θ too low can lead to some actually misaligned entity pairs being mistakenly included in the reasoning stage, reducing filtering efficiency. In practical applications, θ can be adjusted based on the proportion of entity synonyms and heterogeneous terms in the domain being processed. high With θ low Adaptive adjustments are made to achieve an optimal balance between aligned recall and computational cost. The candidate entity set P to be disambiguated... cand The set of attribute features carried by each record in the middle. feat With relation feature set R feat In the pre-filtering stage, it does not participate in classification decision-making, but is used as construction material for attribute-aware prompts and relation-aware prompts in the subsequent two progressive large language model reasoning stages.

[0139] The dual-threshold pre-filtering combined with attribute-relation two-stage progressive entity disambiguation method is a cross-domain entity alignment method that organically connects coarse-grained semantic vector screening with fine-grained reasoning of large language models, and progressively allocates computational resources according to decision difficulty. Essentially, it is an optimization of the computational resource reallocation of the traditional entity alignment task's "treating all candidate pairs equally" approach. In heterogeneous knowledge fusion scenarios in the technology field, the candidate set P to be disambiguated... cand The candidate entity pairs in the data exhibit a highly uneven distribution according to the difficulty of disambiguation: a considerable proportion of candidate pairs are highly clustered in the semantic vector space (corresponding to simple and clear synonymous entity pairs, such as "neural network" and "artificial neural network"), which can be directly judged by vector similarity without reasoning; another considerable proportion of candidate pairs are clearly separated in the semantic vector space (corresponding to semantically unrelated entity pairs), which can be directly eliminated without reasoning; only a small number of candidate pairs fall into the uncertain range of semantic vector similarity (corresponding to synonymous terms in the field of science and technology, such as abbreviations and full names, and common expressions in different fields). These boundary cases are the real challenge of cross-domain disambiguation.

[0140] The two-stage progressive reasoning design based on attribute awareness and relation awareness is based on the complementary coverage of entity semantics by two types of features: attribute feature set A. feat (Entity type labels, domain labels, definition descriptions, etc.) reflect the static semantic attributes of the entity itself. When an entity has clear attribute definitions (such as an entity name with a clear type description or domain label), the attribute features are sufficient to support a large language model in making accurate alignment judgments, without the need to introduce graph structure information; relation feature set R feat (The type of neighborhood relationship between an entity and the types of neighboring entities in a knowledge graph) reflects the structural role of an entity in the knowledge network. For entities with sparse attribute descriptions or ambiguous semantics (such as technical terms that appear only in abbreviations or lack complete attribute definitions), the type of entity they are connected to in the knowledge graph and the nature of their connections can often reveal their true semantics and compensate for the lack of attribute features. Let's start with A... feat Perform attribute-aware inference (low-cost, candidate pairs with clear attribute coverage), then use R... feat The system performs relation-aware reasoning (high-cost, sparsely covering boundary candidate pairs). Its two-stage progressive design systematically covers two types of disambiguation scenarios, achieving dual optimization of computational efficiency and disambiguation accuracy, provided that the overall reasoning cost does not exceed that of a single round of full attribute reasoning.

[0141] Taking a specific application scenario as an example, in a cross-domain entity disambiguation process, candidate entity pairs ("CNN", "convolutional neural network") enter the candidate set P to be inferred. uncertain (The cosine similarity score is 0.82, falling within the uncertain interval). In the first stage of attribute-aware reasoning, the large language model is based on A. feat The attribute information of "CNN" (domain label: computer vision; type label: research method) and the attribute information of "convolutional neural network" (domain label: deep learning; type label: research method) are used to construct attribute-aware cues, which are then processed by N. vote After three rounds of parallel reasoning, all three votes unanimously determined the pair to be "aligned," and the entity pair was included in the first-stage confirmed alignment set, Align. attr The first candidate entity pair ("XGBoost", "gradient boosting decision tree") did not need to proceed to the second stage; however, the other candidate entity pair ("XGBoost", "gradient boosting decision tree") was included in the second stage candidate set P because its attribute feature description was relatively vague (the domain labels of the two overlapped but were not completely consistent). The voting result in the first stage was 2 to 1, which did not reach the majority vote threshold. uncertain2 In the second stage, relationship-aware prompts are constructed based on relational features (both entities in the knowledge graph exhibit numerous technical entities such as "feature engineering" and "ensemble learning"), and then inferred and M... ont After legality verification and alignment confirmation, the alignment result is included in the second-stage supplementary alignment result set, Align. relThe above example illustrates that the two-stage progressive design can adaptively select inference strategies based on the richness of attribute information of different candidate entity pairs, minimizing the total number of calls to the large language model while ensuring disambiguation accuracy.

[0142] Example 8

[0143] In this embodiment, the two-stage progressive large language model reasoning includes:

[0144] Based on the candidate set to be reasoned, the attribute feature set of each candidate entity pair in the candidate set to be reasoned is constructed as an attribute-aware prompt, which is input into the large language model to perform multiple rounds of parallel reasoning. Each candidate entity pair is aligned by majority voting. Candidate entity pairs that are aligned are included in the first stage confirmed alignment set, and candidate entity pairs that are not aligned are included in the second stage candidate set to be reasoned, thus obtaining the first stage confirmed alignment set and the second stage candidate set to be reasoned.

[0145] Specifically, regarding the candidate set P for reasoning uncertain Each candidate entity pair (e) s e t ), and its attribute feature set A feat The algorithm constructs an attribute-aware prompt, which includes attribute descriptions (top-K attributes such as type labels, domain labels, and definition text) for each of the two entities. The large language model is required to determine whether the two entities refer to the same semantic concept and output a binary judgment result (aligned / unaligned). For each candidate entity pair, N operations are performed. vote Wheel (N) vote The value range is 3 to 5 rounds) of parallel reasoning, with each round of reasoning outputting a judgment result independently, and the majority voting method is used (the passing threshold for voting is ⌈N). vote / 2⌉+1 votes) against N vote The results of the rounds are aggregated: if the voting results reach the majority threshold and are determined to be "aligned", then the candidate entity pair is included in the first-stage confirmed alignment set Align. attr Otherwise, it will be included in the second-stage candidate set P for further reasoning. uncertain2 Then, it enters the relation-aware reasoning stage. The voting aggregation mechanism of multi-round parallel reasoning can effectively reduce the impact of the randomness of a single reasoning in a large language model on the judgment result.

[0146] The design basis for the multi-round parallel inference mechanism is as follows: When performing alignment judgment, the large language model is affected by factors such as temperature parameters and sampling strategies, and different inference rounds for the same input may produce different judgment results. This randomness of single-round inference is particularly significant in boundary cases (i.e., candidate entities that are semantically similar but not completely equivalent). By executing N... voteEach round of independent parallel reasoning (each round of reasoning independently calls the large language model with the same prompt words, and the reasoning results do not interfere with each other), and N is selected by majority voting. vote Aggregating rounds of results can reduce the impact of randomness in a single inference on the final judgment from probability error to majority probability error, thereby improving the robustness of alignment judgment.

[0147] N vote The value of N is related to the consistency of the expected judgment of the large language model on the candidate entity pairs processed: vote When N=3, the voting threshold is 2 votes, which is suitable for scenarios where the large language model has high consistency (such as candidate pairs with highly defined attribute features); vote When the threshold is 5, the voting pass threshold is 3 votes, which is suitable for boundary case scenarios requiring higher robustness guarantees (such as second-stage candidate pairs with relatively ambiguous relationship characteristics). Optionally, in actual deployment, the voting results of the first stage can also be used as the basis (e.g., N). vote When the threshold is 3, a split vote of 2 to 1 occurs. The system dynamically determines whether to add additional inference rounds in the current stage to achieve a dynamic balance between judgment confidence and computational cost. It is worth noting that multi-round parallel inference is not executed sequentially; N... vote Round-by-round inference tasks can be submitted concurrently to the large language model server. In engineering implementation, this does not linearly increase response latency, thus maintaining the overall processing efficiency of the system.

[0148] Based on the candidate set to be reasoned in the second stage, the relation feature set of each candidate entity pair in the candidate set to be reasoned in the second stage is constructed as a relation-aware prompt. The prompt is input into the large language model to perform parallel reasoning in the same number of rounds. Each candidate entity pair is initially aligned using the same majority voting method. The voting results are then validated using the cross-domain mapping relation set in the hierarchical heterogeneous ontology framework to obtain the second stage supplementary alignment result set.

[0149] Specifically, for the candidate set P to be inferred in the second stage uncertain2 Each candidate entity pair (e) s e t ), and set its relational feature set R feat The construct is a relation-aware prompt, which includes the Top-K neighborhood relation types of each entity in the knowledge triples and descriptions of the neighboring entity types. The large language model is required to determine whether two entities refer to the same semantic concept based on the graph neighborhood relation patterns of the entities. Similarly, N is executed. vote Rounds of parallel reasoning, using a majority vote (the threshold for passing a vote is ⌈N). voteEach candidate entity pair is initially aligned using 2 votes (+1 votes) to obtain preliminary voting results. After obtaining the preliminary voting results, the cross-domain mapping relationship set M in the hierarchical heterogeneous ontology framework is used. ont As a rule constraint layer, the legitimacy of candidate entity pairs judged as "aligned" in the preliminary voting results is verified: This verifies whether there exists an M-type relationship between the scientific entity type and the technical entity type in the candidate entity pair. ont Cross-domain mapping relationships defined in the code are validated and retained as alignment results if they exist; otherwise, validation fails and they are removed from the alignment results. Finally, candidate entity pairs that pass validation are aggregated into the second-stage supplementary alignment result set, Align. rel .

[0150] Based on the first-stage confirmed alignment set and the second-stage supplementary alignment result set, the first-stage confirmed alignment set, the confirmed merged set, and the second-stage supplementary alignment result set are merged to obtain the cross-domain entity disambiguation alignment result set.

[0151] Specifically, the alignment set Align, which was confirmed in the first phase, will be... attr In Example 7, the Align acknowledgment merge set directly obtained from dual-threshold pre-filtering direct And the second-stage supplementary alignment result set Align rel Perform a set union operation to obtain the complete cross-domain entity disambiguation and alignment result set, Align. final = Align direct ∪ Align attr ∪ Align rel Regarding Align final Each aligned entity pair undergoes an entity merging operation, combining the two entities into a unified node. During the merging process, if two entities have conflicting relationships (i.e., different attribute values ​​for the same relationship), the attribute value from the source with higher confidence is used. After entity merging, document citation relationships are transformed into "support / citation" association edges between knowledge nodes, and document metadata (timestamp, institution name, domain tag, etc.) is appended as node attributes to the corresponding knowledge nodes, resulting in an enhanced fusion triplet set T with metadata and citation relationships. enhanced .

[0152] Example 9

[0153] In this embodiment, a multi-granularity temporal hypergraph is constructed and the temporal information of each granularity layer is weighted and integrated through a cross-scale conflict sensing mechanism, including:

[0154] Based on the enhanced fusion triplet set and the semantic vectors of entities in each domain, an annual superlayer, a quarterly superlayer, and a citation-instantaneous superlayer are constructed according to three time granularities: year, quarter, and citation-instantaneous. Within each of the annual, quarterly, and citation-instantaneous superlayers, when the cosine similarity of the semantic vectors between a scientific entity and a technical entity exceeds a preset hyperedge construction threshold, and the document timestamps of the scientific entity and the technical entity both belong to the same time granularity interval, a hyperedge connection is established between the scientific entity and the technical entity within the superlayer, resulting in a multi-granularity temporal supergraph containing the annual, quarterly, and citation-instantaneous superlayers.

[0155] Specifically, to enhance the fusion triplet set T enhanced With the semantic vectors of entities in each domain (V) sci V tec As a joint input, superlayers are constructed based on three time granularities: year, quarter, and reference time.

[0156] Annual superlayer: Using the calendar year (e.g., 2020, 2021, etc.) as the time granularity interval, it applies to T... enhanced Each pair of scientific entities e s With technology entity e t If the cosine similarity of the semantic vectors of the two is sim(v s , v t ) > θ hyper (θ) hyper (Construct a threshold for the preset superedge), and e s Document timestamp year(e) s ) and e t Document timestamp year(e) t If they belong to the same calendar year, then they are represented as e within the year superlayer. s With e t Establish a hyperedge connection;

[0157] Quarterly superlayer: Using natural quarters (e.g., 2020Q1, 2020Q2, etc.) as the time granularity interval, it assigns values ​​to elements that satisfy sim(v... s ,v t ) > θ hyper And quarter(e) s ) = quarter(e t Establish hyperedge connections between entity pairs that meet certain conditions;

[0158] Citation Immediate Hyperlayer: Using citation batches (i.e., citation dissemination batches after paper publication or patent authorization) as the time granularity interval, it defines the citations that satisfy sim(v) s , v t ) > θ hyperAnd citation batch (e s ) = citation batch (e t Establish hyperedge connections between entity pairs that meet certain conditions.

[0159] The three superlayers together constitute a multi-granularity temporal supergraph G. hyper Each hyperedge contains the identifiers of all entity nodes within the hyperedge, the temporal granularity interval label corresponding to the hyperedge, and the semantic similarity weight of the hyperedge.

[0160] Based on the multi-granularity temporal hypergraph, it detects whether there are conflicts in the relationship annotations of the same entity pair among the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. It counts the number of conflicts for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. It calculates the corresponding fusion weight based on the number of conflicts for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. The fusion weight is inversely proportional to the number of conflicts for the corresponding hyperlayer. The fusion weights for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer are then weighted and fused with the temporal information of the corresponding hyperlayer to obtain the multi-granularity temporal attribute vector for each knowledge node.

[0161] Specifically, based on the multi-granularity temporal hypergraph G hyper , for G hyper For the same entity pair contained in the annual, quarterly, and reference-based instant hyperlayers, check for conflicts in the relation annotations (relation type and direction) of the three hyperlayers for this entity pair; that is, whether different granularity layers have annotated the same entity pair with semantically inconsistent relation types. Count the number of conflicts C in the annual hyperlayer. year The number of conflicts in the quarterly superlayer C quarter The number of conflicts when referencing an instant hyperlayer (C) instant The corresponding blending weight is calculated based on the inverse relationship between the number of conflicts in each superlayer:

[0162] W year = (1 / C year ) / (1 / C year + 1 / C quarter + 1 / C instant )

[0163] W quarter = (1 / C quarter ) / (1 / C year + 1 / C quarter + 1 / C instant )

[0164] W instant= (1 / C instant ) / (1 / C year + 1 / C quarter + 1 / C instant )

[0165] When the number of conflicts in a certain superlayer is zero, the corresponding weight is smoothed with a minimum value to avoid division by zero anomalies. For G... hyper For each knowledge node, extract its temporal information vector (including hyperedge connection patterns, hyperedge density within the time granularity interval, and neighborhood entity evolution trends, etc.) in the annual, quarterly, and referenced instantaneous hyperlayers, respectively, using W... year W quarter W instant The temporal information vectors of the three granularity layers are weighted and fused to obtain a multi-granularity temporal attribute vector for each knowledge node. This multi-granularity temporal attribute vector is then appended to the attributes of the corresponding knowledge node, forming a heterogeneous science-technology semantic knowledge graph G with multi-granularity temporal attributes. kgraph G kgraph The graph data was formatted as Cypher query statements, batch-imported into the Neo4j graph database, and indexes were created for node type, relationship type, and three-level time granularity (year / quarter / reference) to obtain a persistent heterogeneous science-technology knowledge graph database DB that supports multi-granularity temporal queries. neo4j .

[0166] Among them, the judgment criteria for the "semantic inconsistency" of relation annotations for the same entity pair between different granularity layers are as follows: If any two superlayers in the annual superlayer, quarterly superlayer, and referenced instant superlayer have conflicting relationships for the same entity pair (e... s e t The superedge relation type label marked by ) is in relation type set R ont If the entity pair belongs to different relation types, it is determined that there is a relation annotation conflict between the two granularity layers; if there is no hyperedge connection between the entity pair in a certain hyperlayer (i.e., the semantic similarity threshold θ is not met), it is considered that there is a relation annotation conflict between the two granularity layers. hyper If there is a dual constraint of the same granularity interval as the timestamp, then the superlayer will not include the entity pair in the conflict statistics and will be regarded as a missing label rather than a conflict label.

[0167] It should be noted that the minimum value smoothing in the fusion weight calculation formula when the number of conflicts of a certain superlayer C=0 can be achieved by replacing C with ε (ε is a preset minimum positive value, with an example value of 1×10⁻). 6This is achieved by using a method that avoids numerical calculation errors caused by division by zero, while ensuring that the superlayer with zero conflict (i.e., the superlayer with completely consistent internal relation labels) receives the highest relative weight in the fusion weight allocation, reflecting its status as the most reliable temporal granularity layer. In actual calculations, if the number of conflicts in all three superlayers is zero (i.e., the internal relation labels of each granularity layer are completely consistent), then the fusion weights of the three granularities are smoothed and taken as equal values ​​(approximately 1 / 3 each), degenerating into an equal-weighted average fusion. This situation reflects the high consistency of temporal information among the three granularities and can be regarded as a high-confidence joint representation of the temporal attributes of knowledge nodes.

[0168] Multi-granularity temporal hypergraphs refer to a knowledge graph temporal representation structure that uses hypergraphs as its mathematical foundation, multiple temporal granularity layers as its organizational dimension, and entity semantic vector similarity and timestamp attribution intervals as dual constraints for establishing hyperedges. Its core difference from traditional scientific and technological knowledge graphs lies in the following: traditional knowledge graphs only attach a single timestamp attribute (such as the year of publication) to knowledge nodes, storing temporal information in isolation as node attributes, and the temporal relationships between nodes require subsequent queries and calculations to reflect them; while multi-granularity temporal hypergraphs directly encode the temporal relationships between knowledge entities into a hyperedge connection structure, and simultaneously construct independent hyperlayers at three granularity dimensions: year, quarter, and citation time. This allows the evolutionary patterns of knowledge at different time scales to be expressed and queried in parallel within the same graph structure.

[0169] The semantic vector cosine similarity threshold θ in the hyperedge construction rule hype The technical significance of the dual constraints of r and timestamp granularity interval is as follows: The cosine similarity constraint ensures sufficient semantic relevance between the scientific and technological entities connected by the hyperedge. That is, the hyperedge represents, at the semantic level, "the potential impact path of scientific discovery on technological application," rather than noisy connections between random entity pairs. The timestamp granularity interval constraint ensures, at the temporal level, that the hyperedge represents "the co-occurrence association of scientific and technological knowledge within the same time window." That is, the two connected entities are simultaneously active within the corresponding time granularity interval (published in the same year, the same quarter, or in the same citation dissemination batch), thus guaranteeing the temporal validity of the temporal information carried by the hyperedge. The combined effect of these two constraints allows the hyperedge connection to simultaneously filter meaningful scientific-technological knowledge associations from both temporal and semantic dimensions, effectively suppressing the introduction of noisy hyperedges with mismatched time windows or irrelevant semantics.

[0170] The technical complementarity of parallel modeling at three granularities—annual, quarterly, and citation-instantaneous—is reflected in its adaptive coverage of the evolutionary patterns of different types of scientific and technological knowledge: the annual superlayer captures the long-term evolutionary pattern of the diffusion of macro-research trends in the scientific field to the field of technological applications with a coarser time granularity, which is suitable for analyzing the macro-path of "a certain basic scientific theory (such as the deep learning attention mechanism) gradually penetrating into multiple technological application directions over a span of many years"; the quarterly superlayer captures the synergistic evolutionary pattern of scientific research hotspots and technological innovation waves within a specific period with a medium time granularity, which is suitable for analyzing "the impact of a quarterly breakthrough in a certain research field (such as the publication of a new neural network architecture) on the distribution of patent application hotspots in the same period"; and the citation-instantaneous superlayer captures the immediate effect of citation dissemination after the publication of scientific papers or the granting of patents with the finest time granularity, which is suitable for analyzing "the rapid diffusion path of a certain breakthrough scientific achievement being intensively cited in a short period of time and giving rise to technological patents". The parallel existence of the three granularity layers enables the heterogeneous science and technology knowledge graph to return evolution path information at different resolutions according to the specified time granularity when users query, thus meeting the differentiated query needs of different application scenarios such as science and technology intelligence analysis (annual trends), R&D decision support (quarterly hot topics), and patent early warning (instant citation).

[0171] Taking a specific application scenario for science and technology intelligence analysis as an example, this illustrates the practical effect of multi-granularity temporal hypergraphs: A science and technology intelligence agency aims to analyze the multi-scale evolutionary path of the transformation of scientific research results from "graph neural networks" into industrial technological applications. This is achieved by querying the persistent heterogeneous science and technology knowledge graph database DB. neo4j The annual superlayer query results show that from 2018 to 2022, the density of hyperedge connections between scientific entities related to "graph neural networks" and technical entities such as "molecular structure analysis," "recommendation systems," and "knowledge graph reasoning" continued to increase, reflecting the macro trend of graph neural networks spreading from basic research to multiple technological applications. The quarterly superlayer query results show a concentrated burst of hyperedge connections in the third quarter of 2020, which coincides with the historical event of a benchmark paper on graph neural networks being published in that quarter and triggering a large number of patent applications. The citation instant superlayer query results show that in the first batch of citations after the publication of the benchmark paper, technical entities from certain institutions appeared densely in the hyperedge connections, revealing the organizational entities and dissemination paths of technology rapidly following scientific breakthroughs. The query results of the three granularities complement each other, jointly constituting a multi-dimensional and accurate characterization of the evolution path of scientific and technological knowledge of "graph neural networks," far exceeding the temporal information depth provided by a single timestamp annotation scheme.

[0172] Example 10

[0173] like Figure 3 As shown, this embodiment provides a heterogeneous science and technology knowledge graph construction system, including:

[0174] The ontology framework construction module 301 is used to analyze multi-source scientific and technological literature and domain expert interview data. It formally defines and verifies the differences in knowledge structure between scientific papers and technical patents through ontology modeling methods, and constructs a hierarchical heterogeneous ontology framework that includes scientific knowledge sub-ontologies, technical knowledge sub-ontologies and cross-domain mapping relationship sets.

[0175] The triplet extraction module 302 is used to take the hierarchical heterogeneous ontology framework and the multi-source scientific and technological literature database as input, construct an entity-relation joint extraction framework through domain fine-tuning of the large language model, perform dual-channel retrieval enhanced triplet extraction in batches, and merge all batch results to obtain a set of scientific knowledge triplets and a set of technical knowledge triplets.

[0176] The cross-domain disambiguation alignment module 303 is used to map entities and relations to a unified semantic vector space based on the scientific knowledge triple set and the technical knowledge triple set through graph knowledge representation learning. After cross-domain entity disambiguation alignment is completed through dual-threshold pre-filtering and two-stage progressive large language model inference, and after adding document metadata, an enhanced fusion triple set is obtained.

[0177] The knowledge graph persistence module 304 is used to construct a multi-granularity temporal hypergraph based on the enhanced fusion triplet set and the semantic vectors of entities in each domain, and to integrate the temporal information of each granularity layer through a cross-scale conflict perception mechanism. The resulting heterogeneous science-technology semantic knowledge graph is then persistently stored in a graph database to obtain a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

[0178] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of this application are not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0179] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0180] Similarly, it should be understood that, in order to simplify this application and aid in understanding one or more of the various aspects of the invention, in the above description of exemplary embodiments of this application, various features of the embodiments of this application are sometimes grouped together into a single embodiment, figure, or description thereof.

[0181] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying abstract and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying abstract and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0182] It should be noted that the above embodiments are illustrative of this application and not restrictive, and those skilled in the art can design alternative embodiments without departing from the scope. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

[0183] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a heterogeneous scientific-technological knowledge graph, characterized in that, include: We analyze multi-source scientific literature and domain expert interview data, and formally define and verify the differences in knowledge structure between scientific papers and technical patents through ontology modeling methods. We construct a hierarchical heterogeneous ontology framework that includes scientific knowledge sub-ontologies, technical knowledge sub-ontologies and cross-domain mapping relationship sets. Using the aforementioned hierarchical heterogeneous ontology framework and multi-source scientific and technological literature database as input, an entity-relation joint extraction framework is constructed through a domain-fine-tuned large language model. Dual-channel retrieval is performed iteratively in batches to enhance triple extraction. All batch results are merged to obtain a set of scientific knowledge triples and a set of technical knowledge triples. Based on the aforementioned scientific knowledge triplet and technical knowledge triplet, entities and relations are mapped to a unified semantic vector space through graph knowledge representation learning. Cross-domain entity disambiguation and alignment are completed through dual-threshold pre-filtering and two-stage progressive large language model inference. After adding document metadata, an enhanced fusion triplet is obtained. The dual-threshold pre-filtering and the two-stage progressive large language model inference include: based on the candidate entity set to be disambiguated, according to the cosine similarity score of each record in the candidate entity set to be disambiguated, candidate entity pairs with cosine similarity scores not lower than a preset high threshold are included in the confirmation merging set; candidate entity pairs with cosine similarity scores lower than a preset low threshold are included in the exclusion set; and candidate entity pairs with cosine similarity scores between the preset low threshold and the preset high threshold are included in the candidate set to be inferred, thus obtaining the confirmation merging set, the exclusion set, and the candidate set to be inferred; based on the candidate set to be inferred, the attribute feature set of each candidate entity pair in the candidate set to be inferred is constructed as an attribute-aware prompt, which is then input into the large language model. The model performs multiple rounds of parallel inference, using majority voting to align each candidate entity pair. Aligned candidate entity pairs are added to the first-stage confirmed alignment set, while misaligned candidate entity pairs are added to the second-stage inference candidate set, resulting in the first-stage confirmed alignment set and the second-stage inference candidate set. Based on the second-stage inference candidate set, relational feature sets for each candidate entity pair are constructed as relation-aware prompts. These prompts are input into the large language model for the same number of rounds of parallel inference, using the same majority voting method to perform preliminary alignment judgments on each candidate entity pair. The voting results are then validated using the cross-domain mapping relation set in the hierarchical heterogeneous ontology framework, resulting in the second-stage supplementary alignment result set. Based on the first-stage confirmed alignment set and the second-stage supplementary alignment result set, the first-stage confirmed alignment set, the merged confirmed set, and the second-stage supplementary alignment result set are merged to obtain the cross-domain entity disambiguation alignment result set. Based on the enhanced fusion triplet set and semantic vectors of entities in each domain, a multi-granularity temporal hypergraph is constructed and the temporal information of each granularity layer is weighted and integrated through a cross-scale conflict perception mechanism. The resulting heterogeneous science-technology semantic knowledge graph is persistently stored in a graph database, resulting in a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

2. The method for constructing a heterogeneous science and technology knowledge graph according to claim 1, characterized in that, The analysis of multi-source scientific literature and domain expert interview data, using ontology modeling methods to formally define and verify the differences in knowledge structure between scientific papers and technological patents, constructs a hierarchical heterogeneous ontology framework comprising scientific knowledge sub-ontologies, technological knowledge sub-ontologies, and cross-domain mapping relationship sets, including: Based on multi-source scientific and technological literature and interview data with domain experts, we sorted out the differences in knowledge structure between scientific papers and technology patents, and obtained a draft of the concept system of scientific and technological knowledge domain that includes a set of entity types and a set of relation types. Based on the aforementioned draft concept system of scientific and technological knowledge domain, the entity type set and the relation type set are formally defined and attribute constraints are added using ontology modeling tools. Consistency checks are then performed to obtain an empirical scientific and technological knowledge ontology that includes the entity type set, relation type set, and attribute constraint set. Based on the entity type set and relation type set in the verified scientific and technological knowledge ontology, a scientific knowledge sub-ontology containing research questions, research methods, and experimental conclusions, and a technical knowledge sub-ontology containing technical efficacy, application scenarios, and invention methods are defined. A cross-domain mapping relation set between the scientific knowledge sub-ontology and the technical knowledge sub-ontology is established based on the attribute constraint set, resulting in a hierarchical heterogeneous ontology framework that supports cross-domain alignment.

3. The method for constructing a heterogeneous science and technology knowledge graph according to claim 1, characterized in that, The framework for constructing a joint entity-relation extraction by fine-tuning a large language model in a domain includes: Based on the hierarchical heterogeneous ontology framework and multi-source scientific and technological literature database, the knowledge entities and relationships between the texts in the multi-source scientific and technological literature database are manually annotated, and the multi-source scientific and technological literature database is divided into multiple batches according to semantic topics to obtain an annotated sample set and a batched literature index table. Based on the labeled sample set, the entity type set and relation type set defined in the hierarchical heterogeneous ontology framework are used as type constraints to fine-tune the pre-trained large language model in the domain. The batch iterative extraction order of the fine-tuned large language model is set according to the batch literature index table to obtain an entity-relation joint extraction framework that can simultaneously identify the relationship between scientific entities, technical entities and knowledge entities.

4. The method for constructing a heterogeneous science and technology knowledge graph according to claim 3, characterized in that, The batch-iterative dual-channel retrieval enhanced triple extraction includes: Based on the entity-relation joint extraction framework, the current batch of documents is initially extracted to obtain a candidate pre-entity set and a candidate triple set, and the intermediate incremental graph up to the current batch is loaded. Based on the candidate pre-entity set and the intermediate incremental graph, each pre-entity in the candidate pre-entity set is used as a search query term to retrieve cross-batch text fragments in the full document of the multi-source scientific and technological literature database, so as to obtain the semantic context corresponding to each pre-entity. Based on the candidate pre-entity set and the intermediate incremental graph, the neighborhood relationship information of each pre-entity is retrieved from the intermediate incremental graph to obtain the graph structure information corresponding to each pre-entity; The semantic context and the graph structure information are combined according to the pre-entity alignment to obtain the dual-channel enhanced context corresponding to each pre-entity. The dual-channel enhanced context and the candidate triple set are jointly input into the large language model in the entity-relation joint extraction framework to obtain the refined triple set of the current batch. The refined triple set is then merged into the intermediate incremental graph to update the intermediate incremental graph. This process is iterated until all batches are processed. Finally, the refined triple sets of all batches are merged to obtain the scientific knowledge triple set and the technical knowledge triple set.

5. The method for constructing a heterogeneous science and technology knowledge graph according to claim 4, characterized in that, The dual-channel retrieval enhancement triplet extraction hallucination filtering includes: Based on each candidate triplet in the refined triplet set and the original text of the current batch of documents, the candidate triplet and the original text are used to construct a verification prompt, which is then input into a large language model. The large language model scores the semantic consistency between the candidate triplet and the original text to obtain the semantic consistency confidence score corresponding to each candidate triplet. Based on the semantic consistency confidence score, candidate triples that are below the preset confidence threshold are filtered out from the refined triple set to obtain the filtered triple set for the current batch. The filtered triple set is then merged into the intermediate incremental graph to update the intermediate incremental graph. This process is iterated until all batches are processed. Finally, the filtered triple sets from all batches are merged to obtain the scientific knowledge triple set and the technical knowledge triple set.

6. The method for constructing a heterogeneous science and technology knowledge graph according to claim 1, characterized in that, The method of mapping entities and relations to a unified semantic vector space through graph knowledge representation learning includes: Based on the scientific knowledge triple set and the technical knowledge triple set, graph knowledge representation learning is performed on entities and relations, mapping the entities and relations in the scientific knowledge triple set and the technical knowledge triple set to a unified semantic vector space, resulting in a scientific entity vector set and a technical entity vector set. Based on the scientific entity vector set and the technical entity vector set, a cosine similarity score is calculated between cross-domain entity pairs. For each candidate entity pair, a top-preset set of attribute features and a set of relation features are extracted to obtain a candidate entity set to be disambiguated. Each record in the candidate entity set contains an entity pair, a cosine similarity score, an attribute feature set, and a relation feature set. The attribute feature set is obtained by extracting the top-preset set of entity attributes from the scientific knowledge triple set and the technical knowledge triple set, on a per-candidate entity-pair basis. The entity attributes are formed in the form of attribute features, including one or more of entity type labels, document domain labels, and definition description text; the relationship feature set is formed by sorting the neighborhood relationship information of each entity in the candidate entity pair according to the relationship confidence score, taking the candidate entity pair as the unit, and extracting the neighborhood relationship information of the top preset number of positions as the relationship features. The neighborhood relationship information includes the relationship type directly connected to each entity in the candidate entity pair and the neighbor entity type. The relationship confidence score is the confidence score of the connection relationship that constitutes the neighborhood relationship information.

7. The method for constructing a heterogeneous science and technology knowledge graph according to claim 1, characterized in that, The construction of a multi-granularity temporal hypergraph and the weighted integration of temporal information from each granularity layer via a cross-scale conflict sensing mechanism include: Based on the enhanced fusion triplet set and the semantic vectors of entities in each domain, an annual superlayer, a quarterly superlayer, and a citation-instantaneous superlayer are constructed according to three time granularities: year, quarter, and citation-instantaneous. Within each of the annual, quarterly, and citation-instantaneous superlayers, when the cosine similarity of the semantic vectors between a scientific entity and a technical entity exceeds a preset hyperedge construction threshold, and the document timestamps of the scientific entity and the technical entity both belong to the same time granularity interval, a hyperedge connection is established between the scientific entity and the technical entity within the superlayer, resulting in a multi-granularity temporal supergraph containing the annual, quarterly, and citation-instantaneous superlayers. Based on the multi-granularity temporal hypergraph, it detects whether there are conflicts in the relationship annotations of the same entity pair among the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. It counts the number of conflicts for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. It calculates the corresponding fusion weight based on the number of conflicts for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer. The fusion weight is inversely proportional to the number of conflicts for the corresponding hyperlayer. The fusion weights for each of the annual hyperlayer, the quarterly hyperlayer, and the referenced instant hyperlayer are then weighted and fused with the temporal information of the corresponding hyperlayer to obtain the multi-granularity temporal attribute vector for each knowledge node.

8. A heterogeneous science and technology knowledge graph construction system, using the heterogeneous science and technology knowledge graph construction method as described in any one of claims 1-7, characterized in that, include: The ontology framework construction module is used to analyze multi-source scientific and technological literature and domain expert interview data. It formally defines and verifies the differences in knowledge structure between scientific papers and technical patents through ontology modeling methods, and constructs a hierarchical heterogeneous ontology framework that includes scientific knowledge sub-ontologies, technical knowledge sub-ontologies and cross-domain mapping relationship sets. The triplet extraction module is used to take the hierarchical heterogeneous ontology framework and the multi-source scientific and technological literature database as input, construct an entity-relation joint extraction framework through domain fine-tuning of the large language model, perform dual-channel retrieval enhanced triplet extraction in batch iteratively, and merge all batch results to obtain scientific knowledge triplet sets and technical knowledge triplet sets. The cross-domain disambiguation alignment module is used to map entities and relations to a unified semantic vector space based on the scientific knowledge triple set and the technical knowledge triple set through graph knowledge representation learning. After cross-domain entity disambiguation alignment is completed through dual-threshold pre-filtering and two-stage progressive large language model inference, and after adding document metadata, an enhanced fusion triple set is obtained. The knowledge graph persistence module is used to construct a multi-granularity temporal hypergraph based on the enhanced fusion triplet set and semantic vectors of entities in each domain, and to integrate the temporal information of each granularity layer through a cross-scale conflict perception mechanism. The resulting heterogeneous science-technology semantic knowledge graph is then persistently stored in a graph database to obtain a persistent heterogeneous science-technology knowledge graph database that supports multi-granularity temporal queries.

Images