Multi-source heterogeneous scientific and technological data fusion collection method and system
By constructing a knowledge graph in the field of science and technology innovation and performing time-series fusion and graph neural network reasoning, the problem of integrating multi-source heterogeneous science and technology innovation data was solved, and the automatic mining of explicit and implicit associations was realized, improving the quality of data fusion and collection and the accuracy of query.
Patent Information
- Application Number
- CN202610462946.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-03
Smart Images
Figure CN122332576A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data fusion and acquisition technology, and more specifically, to a method and system for fusion and acquisition of multi-source heterogeneous scientific and technological innovation data. Background Technology
[0002] Data fusion acquisition refers to the process of taking raw data from multiple heterogeneous data sources and processing it through a series of steps such as acquisition, cleaning, transformation, alignment and integration to form a unified, complete, accurate and more informational dataset.
[0003] Currently, the sources of data in the field of science and technology innovation are extensive and diverse, encompassing multi-source heterogeneous data such as structured project information, semi-structured annual report materials, and unstructured text. This data involves various science and technology innovation entities, including projects, institutions, personnel, achievements, and policies. The data is stored in a scattered manner and lacks standardized representation. Traditional collection methods primarily rely on static extraction and single-dimensional organization, failing to construct a unified knowledge framework for the science and technology innovation field to standardize and integrate the data. Furthermore, they generally ignore the dynamic characteristics of the evolution of entities and their relationships over time, resulting in fragmented and one-sided entity relationship identification. This approach can only uncover superficially observable relationships, failing to discover potential implicit connections between entities. When users perform interactive queries with time conditions, it is difficult to accurately match effective related content at the corresponding time point. Consequently, the completeness, timeliness, and accuracy of science and technology innovation data fusion and collection are insufficient, failing to meet the needs of refined and dynamic science and technology innovation data management and query analysis. Therefore, how to automatically mine and output explicit and implicit relationships in science and technology innovation from dynamically evolving data has become a challenge for the industry. Summary of the Invention
[0004] This application provides a method and system for the fusion acquisition of multi-source heterogeneous science and technology innovation data, which can automatically mine and output the explicit and implicit correlations of science and technology innovation from dynamically evolving science and technology innovation data.
[0005] Firstly, this application provides a method for the fusion and acquisition of multi-source heterogeneous scientific and technological innovation data, comprising the following steps:
[0006] An initial framework for constructing a knowledge graph in the field of science and technology innovation;
[0007] Acquire multi-source heterogeneous science and technology innovation data, perform entity association identification on the multi-source heterogeneous science and technology innovation data, and generate multiple fact triples of the multi-source heterogeneous science and technology innovation data;
[0008] Based on the initial framework of the knowledge graph in the field of science and technology innovation, the fact triples are fused in time to form the time-series knowledge graph of the multi-source heterogeneous science and technology innovation data.
[0009] The time tag of the user’s interactive query on the multi-source heterogeneous science and technology innovation data is obtained, and the graph neural network reasoning under time constraints is performed on the time-series knowledge graph based on the time tag to generate the feature representation of each entity node in the time-series knowledge graph at the user’s query time point.
[0010] Based on the aforementioned feature representation, the fusion collection results of the multi-source heterogeneous science and technology innovation data are output, and the fusion collection results include explicit and implicit associations between entities.
[0011] In some embodiments, the initial framework includes entity types, relationship types, and time dimension constraints, wherein the entity types include at least project entities, institutional entities, personnel entities, outcome entities, and policy entities, the relationship types include at least collaborative relationships, funding relationships, output relationships, and association relationships, and the time dimension constraints are used to record the effective time interval of entities or relationships.
[0012] In some embodiments, the multi-source heterogeneous science and technology innovation data includes structured project data, semi-structured annual report data, and unstructured text data.
[0013] In some embodiments, the initial framework for constructing a knowledge graph in the field of science and technology innovation specifically includes:
[0014] Determine the entity types and relationship types of the knowledge graph in the field of science and technology innovation;
[0015] Determine the temporal dimension constraints for entities and relationships in the knowledge graph of the science and technology innovation field;
[0016] An initial framework for a knowledge graph in the field of science and technology innovation is constructed based on the entity type, the relationship type, and the time dimension constraint.
[0017] In some embodiments, performing entity association identification on the multi-source heterogeneous science and technology innovation data to generate multiple fact triples of the multi-source heterogeneous science and technology innovation data specifically includes:
[0018] An initial set of fact triples was identified from the multi-source heterogeneous science and technology innovation data;
[0019] The initial set of fact triples is associated and merged to obtain multiple fact triples of the multi-source heterogeneous science and technology innovation data.
[0020] In some embodiments, the temporal fusion of various fact triples based on the initial framework of the science and technology innovation domain knowledge graph to form the temporal knowledge graph of the multi-source heterogeneous science and technology innovation data specifically includes:
[0021] Based on the initial framework of the knowledge graph in the science and technology innovation field, the types of each fact triple are verified to obtain each normalized triple;
[0022] Temporal correlation is performed on each normalized triple to obtain the temporal network structure;
[0023] The temporal network structure and the initial framework of the science and technology innovation domain knowledge graph are indexed and fused to obtain the temporal knowledge graph of the multi-source heterogeneous science and technology innovation data.
[0024] In some embodiments, performing time-constrained graph neural network inference on the temporal knowledge graph based on the time tag to generate feature representations of each entity node in the temporal knowledge graph at the user query time point specifically includes:
[0025] Based on the time tag, the temporal knowledge graph is extracted to obtain the temporal subgraph of each entity node in the temporal knowledge graph at the user query time point;
[0026] Temporal neighborhood features are aggregated for each temporal subgraph to obtain the local temporal aggregated features of each entity node;
[0027] Multi-level message passing is performed on each local temporal aggregate feature to generate the feature representation of each entity node in the temporal knowledge graph at the user query time point.
[0028] Secondly, this application provides a fusion acquisition system for multi-source heterogeneous scientific and technological innovation data, including:
[0029] The building module is used to construct the initial framework for a knowledge graph in the field of science and technology innovation;
[0030] The processing module is used to acquire multi-source heterogeneous science and technology innovation data, perform entity association recognition on the multi-source heterogeneous science and technology innovation data, and generate multiple fact triples of the multi-source heterogeneous science and technology innovation data.
[0031] The processing module is also used to perform temporal fusion of each fact triple based on the initial framework of the knowledge graph in the science and technology innovation field to form a temporal knowledge graph of the multi-source heterogeneous science and technology innovation data.
[0032] The processing module is also used to obtain the time tag when the user interacts and queries the multi-source heterogeneous science and technology innovation data, and to perform graph neural network reasoning under time constraints on the time-series knowledge graph based on the time tag, so as to generate the feature representation of each entity node in the time-series knowledge graph at the time point of the user query.
[0033] The execution module is used to output the fusion acquisition results of the multi-source heterogeneous science and technology innovation data based on the feature representation. The fusion acquisition results include explicit and implicit associations between entities.
[0034] Thirdly, this application provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described method for fusion acquisition of multi-source heterogeneous scientific and technological innovation data.
[0035] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for fusion acquisition of multi-source heterogeneous scientific and technological innovation data.
[0036] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:
[0037] The method and system for fusion acquisition of multi-source heterogeneous science and technology innovation data provided in this application establish a unified, domain-specific, and structured benchmark for the fusion acquisition of multi-source heterogeneous science and technology innovation data by constructing an initial framework for a knowledge graph in the science and technology innovation field. It performs entity association identification and generates fact triples from the multi-source heterogeneous science and technology innovation data, breaking down the heterogeneous barriers and data silos, transforming scattered data into standardized representation units, and achieving standardized data organization. Based on the initial framework, it performs temporal fusion of fact triples to form a temporal knowledge graph, which can adapt to the dynamic evolution of science and technology innovation data, endowing static knowledge with a time dimension feature, and fully preserving... This study focuses on the temporal evolution of science and technology innovation data. By combining user interaction queries with time tags to perform temporal-constrained graph neural network inference and generate entity node feature representations, it can accurately match users' time-dimensional query needs, uncover entity association features at different time points in dynamic science and technology innovation data, and overcome the limitations of static graphs in adapting to dynamic data. Finally, based on the feature representations, it outputs fused collection results containing explicit and implicit associations, directly achieving the core objective of automatically mining and outputting explicit and implicit associations between science and technology innovation entities from dynamically evolving science and technology innovation data. This completes the deep fusion collection and association mining of multi-source heterogeneous science and technology innovation data. Using this approach, it is possible to automatically mine and output explicit and implicit associations in science and technology innovation from dynamically evolving science and technology innovation data. Attached Figure Description
[0038] Figure 1 This is an exemplary flowchart of a method for fusing and collecting multi-source heterogeneous scientific and technological innovation data, as shown in some embodiments of this application.
[0039] Figure 2 This is an exemplary flowchart illustrating the construction of an initial framework according to some embodiments of this application;
[0040] Figure 3 This is a schematic diagram of the workflow of a fusion acquisition system for multi-source heterogeneous scientific and technological innovation data, as shown in some embodiments of this application.
[0041] Figure 4This is a schematic diagram of the structure of a fusion acquisition system for multi-source heterogeneous scientific and technological innovation data, as shown in some embodiments of this application;
[0042] Figure 5 This is a schematic diagram of the structure of a computer device that implements a method for fusion acquisition of multi-source heterogeneous scientific and technological innovation data, according to some embodiments of this application. Detailed Implementation
[0043] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0044] refer to Figure 1 The figure is an exemplary flowchart of a method for fusing and acquiring multi-source heterogeneous scientific and technological innovation data according to some embodiments of this application. The method for fusing and acquiring multi-source heterogeneous scientific and technological innovation data mainly includes the following steps:
[0045] In step 101, an initial framework for a knowledge graph in the field of science and technology innovation is constructed.
[0046] In some embodiments, reference Figure 2 The diagram is an exemplary flowchart of constructing an initial framework in some embodiments of this application. In this embodiment, the initial framework for constructing a knowledge graph in the field of science and technology innovation can be implemented by the following steps:
[0047] In step 1011, the entity types and relation types of the knowledge graph in the field of science and technology innovation are determined;
[0048] In step 1012, the temporal dimension constraints of entities and relationships in the knowledge graph of the science and technology innovation field are determined;
[0049] In step 1013, the initial framework of the knowledge graph in the field of science and technology innovation is constructed by the entity type, the relationship type and the time dimension constraint.
[0050] In specific implementation, determining the entity types and relation types of the knowledge graph in the science and technology innovation field involves: selecting a certain number of project application forms, completion reports, corporate annual reports, policy documents, and structured project databases as analysis samples; for unstructured text data, a combination of pattern matching and dictionary lookup is used for concept extraction: a seed dictionary covering commonly used terms in the science and technology innovation field is pre-constructed, and candidate concepts in the text are located through string matching; simultaneously, the syntactic structure of the text is analyzed to identify noun phrases and their modifiers, and core vocabulary with business meaning is extracted; for semi-structured annual report data, the table is parsed... The table header fields and row / column structure are used to identify column labels representing entity categories. For structured project databases, the table names, field names, and primary / foreign key relationships between tables are directly read. Table names are considered candidate entity types, and relationships between tables are considered candidate relationship types. After the initial extraction, the extraction results are reviewed and categorized, and synonyms are merged. For example, the undertaking unit and the supporting unit are uniformly classified as institutional entities. Concepts unrelated to scientific and technological innovation are eliminated, ultimately forming entity types including projects, institutions, personnel, achievements, and policies, as well as relationship types including collaboration, funding, output, and association.
[0051] In addition, in specific implementation, the temporal constraints of entities and relationships in the knowledge graph of the science and technology innovation field are determined as follows: Valid start and end times are assigned to entities and relationships respectively to support temporal evolution expression; for structured project data, the project initiation year and completion year are directly extracted from the database as the temporal constraints of entities; for tables in semi-structured annual report data, statistical periods are extracted as the time intervals of relationships through rule parsing; for unstructured text, a temporal information extraction tool based on regular expressions and heuristic rules is used to identify the explicit time expressions in the document and standardize them into a unified format; for entities or relationships whose temporal constraints cannot be directly obtained, inference is made or they are marked as default lifecycles based on the temporal consistency of upstream and downstream related data; finally, the temporal constraints of entities and relationships in the knowledge graph of the science and technology innovation field are obtained.
[0052] Furthermore, in specific implementation, the initial framework for constructing a knowledge graph in the field of science and technology innovation, based on the entity types, relation types, and time dimension constraints, is as follows: The framework is defined using a graph database modeling approach. A set of data attributes is defined for each entity type; for example, a project entity includes attributes such as project number, project name, and funding amount, while a personnel entity includes attributes such as name, title, and affiliated institution. Corresponding object attributes are defined for each relation type, clarifying the entity type constraints at both ends of the relation. For example, the undertaking relation connects the personnel entity and the project entity, and the output relation connects the project entity and the result entity. During the definition process, the time dimension constraint is used as the basis for each entity type. The common attributes of the relationship types are embedded, meaning that each entity instance carries its own start time and end time, and each relationship instance also carries its corresponding valid start time and end time. After the definition is completed, the framework is reviewed by a combination of manual inspection and logical verification: checking one by one whether the entity types cover all business scenarios, whether there are any missing or redundant relationship types, and whether the assignment logic of the time dimension constraints is consistent; at the same time, checking whether there are circular dependencies in the inheritance and association between types to ensure that the framework structure is reasonable, and finally obtaining the initial framework of the knowledge graph in the field of science and technology innovation. Other methods can be used to construct it in other embodiments, which are not limited here.
[0053] It should be noted that the entity type in this application represents the core object category in the field of science and technology innovation, reflecting the classification system of nodes in the knowledge graph; the relation type represents the semantic connection method between different entities, reflecting the business logic and interaction mode between various subjects in science and technology innovation activities; the time dimension constraint represents the time constraint of entities or relations on the time axis, reflecting the dynamic evolution characteristics of science and technology innovation data; the initial framework represents the framework of the structured standard of knowledge graph in the field of science and technology innovation, reflecting the unified standard of multi-source heterogeneous data fusion, which can be used to guide the standardized implementation of data extraction, mapping, storage and reasoning applications.
[0054] In step 102, multi-source heterogeneous science and technology innovation data is acquired, and entity association identification is performed on the multi-source heterogeneous science and technology innovation data to generate multiple fact triples of the multi-source heterogeneous science and technology innovation data.
[0055] It should be noted that the multi-source heterogeneous science and technology innovation data in this application refers to a collection of original data in the field of science and technology innovation obtained from different sources and with various structural forms. It reflects the overall picture of information generated by various entities at different stages in science and technology innovation activities. Specifically, it includes structured data, semi-structured data, and unstructured data. Among them, structured data includes project establishment information tables, funding allocation records, and personnel affiliation tables in project databases, which are stored in a regular row and column format. Semi-structured data includes enterprise annual reports and scientific research institution statistical annual reports, which are presented in the form of tables, key-value pairs, or hierarchical tags, but the structure is not uniform. Unstructured data includes project application texts, original policy documents, academic papers, and news reports, which exist in the form of free text.
[0056] In some embodiments, entity association identification of the multi-source heterogeneous science and technology innovation data to generate multiple fact triples of the multi-source heterogeneous science and technology innovation data can be achieved by the following steps:
[0057] An initial set of fact triples was identified from the multi-source heterogeneous science and technology innovation data;
[0058] The initial set of fact triples is associated and merged to obtain multiple fact triples of the multi-source heterogeneous science and technology innovation data.
[0059] In specific implementation, identifying the initial fact triplet set from the multi-source heterogeneous science and technology innovation data involves the following steps: For structured project data, structured data parsing technology is used to directly read the field relationships in the data table. For example, fields such as project name, funding institution, and funding amount are extracted from the project information table. The project name is used as the head entity, funding as the relationship, and the funding institution as the tail entity. At the same time, the funding time of the corresponding project is associated to generate the initial fact triplet. For semi-structured annual report data, regular expression matching is used to identify the entities and related information contained in the annual report. For example, from the R&D cooperation section of the annual report of a research institution, text fragments of institution A, cooperating with, and institution B are matched to extract the head entity institution A, the relationship cooperation, and the tail entity institution B, and combined with the annual report publication time. First, a time dimension constraint is determined to form initial triples. For unstructured text data, an entity recognition method based on a domain dictionary is adopted. First, a science and technology innovation domain entity dictionary is constructed, which includes commonly used words and synonyms corresponding to categories such as projects, institutions, personnel, achievements, and policies. Through string matching and keyword retrieval, various entities are identified from the text. Then, a relation extraction method based on dependency parsing is adopted to analyze the syntactic structure between entities and determine the relation type between entities. For example, from the case that researcher Zhang led the research and development of XX patent, the head entity Zhang, the relation output, and the tail entity XX patent are identified to generate initial fact triples. The initial triples generated by all types of data are summarized to obtain the initial fact triple set. Other methods can also be used in other embodiments, which are not limited here.
[0060] Furthermore, in specific implementation, the initial set of fact triples is associated and merged to obtain multiple fact triples of the multi-source heterogeneous science and technology innovation data. Specifically, the first step involves using a string similarity matching algorithm for entity disambiguation, calculating the similarity of entity names in different triples. For example, the similarity matching algorithm could be an edit distance algorithm. A reasonable similarity threshold is set. This threshold can be initially determined by statistically analyzing the similarity distribution of entity names in the science and technology innovation field to identify the critical value for distinguishing synonymous and heteronymous entities. Then, a portion of labeled science and technology innovation entity samples are selected for verification and fine-tuning. Finally, a fixed threshold suitable for this field is determined, and entities with similarity higher than the threshold are classified as the same entity. The first step involves identifying entities, such as Tsinghua University and Tsinghua University, as the same institutional entity, and unifying their entity representations. The second step is to deduplicate triples by comparing the head entity, relation, tail entity, and time dimension constraints of all initial triples, and deleting completely duplicate triples to avoid data redundancy. The third step is to complete the triples by combining well-known association rules in the field of science and technology innovation. For example, if a personnel entity belongs to an institutional entity, then the project entity that the personnel entity participates in has an association relationship with that institutional entity. Triples with missing relations or entities are supplemented and improved, ultimately resulting in multiple fact triples of multi-source heterogeneous science and technology innovation data after deduplication, disambiguation, and completion. Other methods can also be used in other embodiments, which are not limited here.
[0061] It should be noted that the initial fact triple set in this application represents a scattered combination of entities, relationships, and corresponding time information initially identified in the science and technology innovation data, reflecting the basic association information that is intuitively present within each science and technology innovation data segment; the fact triple represents the effective association and corresponding time constraints between entities in the science and technology innovation field that have been sorted out and confirmed, reflecting the real and unified entity association logic in multi-source science and technology innovation data, and can be used to analyze the relationship between data in multi-source heterogeneous science and technology innovation data.
[0062] In step 103, based on the initial framework of the knowledge graph in the science and technology innovation field, the fact triples are fused in time to form the time-series knowledge graph of the multi-source heterogeneous science and technology innovation data.
[0063] In some embodiments, the temporal fusion of various fact triples based on the initial framework of the science and technology innovation domain knowledge graph to form the temporal knowledge graph of the multi-source heterogeneous science and technology innovation data can be achieved by the following steps:
[0064] Based on the initial framework of the knowledge graph in the science and technology innovation field, the types of each fact triple are verified to obtain each normalized triple;
[0065] Temporal correlation is performed on each normalized triple to obtain the temporal network structure;
[0066] The temporal network structure and the initial framework of the science and technology innovation domain knowledge graph are indexed and fused to obtain the temporal knowledge graph of the multi-source heterogeneous science and technology innovation data.
[0067] In specific implementation, the type verification of each fact triple is performed according to the initial framework of the science and technology innovation knowledge graph to obtain each normalized triple. Specifically, based on the entity types, relation types, and time dimension constraints specified in the initial framework of the science and technology innovation knowledge graph, a refined type verification is carried out on each fact triple. Breaking through the traditional single verification mode, a three-level processing method of matching, verification, and correction is adopted. First, by combining string exact matching and fuzzy matching, it is verified whether the head entity and tail entity in the triple belong to the five entity types of projects, institutions, personnel, achievements, and policies preset by the framework. At the same time, the relationship is compared to see if it conforms to the protocol. The system employs four pre-defined relationship types: creation, funding, output, and association. It then addresses time-related constraints by validating the time intervals for reasonableness. For example, the relationship's effective time must not be later than the entity's expiration time, and the time format must be standardized as start time - end time. Invalid triples with mismatched entity types, incorrect relationship types, or abnormal time constraints are eliminated. For triples with inconsistent entity descriptions but identical essence, a domain synonym mapping table is used for standardization correction, unifying the descriptions of entities and relationships. The final result is standardized triples with consistent format, compliant types, and valid information. Other implementation methods can also be used in other embodiments, which are not limited here.
[0068] In addition, in specific implementation, the time-series association of each normalized triple is performed to obtain the time-series network structure as follows: For all normalized triples, firstly, according to the characteristics of science and technology innovation data, such as project data by quarter, policy data by year, and achievement data by specific date, different time granularities are adapted, and the effective time interval bound to each triple is decomposed into time nodes of corresponding granularities. Then, with the entity as the core, an entity association time-series ledger is established, and triples belonging to the same entity but different time nodes are linked together in chronological order. At the same time, information on the status changes of the entity in different time intervals is supplemented, such as institutional mergers, project initiation and acceptance, personnel job changes, etc. The complete effective time life cycle is marked for each entity node, and the effective and termination periods are bound to each relationship edge. A time-series network structure that combines entity association logic, time dynamic characteristics, and status change information is constructed. Other methods can also be used in other embodiments, which are not limited here.
[0069] In addition, in specific implementation, the temporal network structure and the initial framework of the science and technology innovation domain knowledge graph are indexed and fused to obtain the temporal knowledge graph of the multi-source heterogeneous science and technology innovation data. Specifically, on the one hand, through hash indexing technology, a one-to-one mapping is established between entity nodes and relation edges in the temporal network structure and entity types and relation types in the initial framework, realizing the framework's normative constraints on the temporal network; on the other hand, through dictionary mapping technology, the time dynamic information and entity state change information in the temporal network are embedded into the time dimension constraints of the initial framework, supplementing the dynamic characteristics of the initial framework. At the same time, an index update mechanism is established to ensure that newly added normalized triples can be quickly integrated into the existing structure. Finally, the association information, time information, and entity state information of all normalized triples are integrated to form a temporal knowledge graph that combines static framework norms and dynamic time characteristics and is adapted to multi-source heterogeneous science and technology innovation data. Other methods can also be used in other embodiments, which are not limited here.
[0070] It should be noted that the standardized triplet in this application represents a standardized association unit of multi-source science and technology innovation data that conforms to the entity and relationship type norms in the field of science and technology innovation and has accurate time constraints. It reflects the objective basic association facts in multi-source science and technology innovation data after removing ambiguity and errors, and can be used as a standard data unit for temporal association and network construction. The temporal network structure represents the association links and duration intervals of various entities and their interrelationships in the field of science and technology innovation as time changes. It reflects the association evolution and state changes of science and technology innovation elements such as projects, institutions, personnel, and achievements at different time nodes, and can be used to build a basic network skeleton that integrates spatial association and temporal attributes. The temporal knowledge graph represents a science and technology innovation full-domain association structure that combines static domain norms and dynamic temporal characteristics. It reflects the complete association logic, evolution law, and hierarchical relationship of multi-source heterogeneous science and technology innovation data in the time dimension, and can directly support reasoning calculation under temporal constraints, explicit and implicit association mining, and the generation and output of fused collection results.
[0071] In step 104, the time tag of the user's interactive query on the multi-source heterogeneous science and technology innovation data is obtained, and the graph neural network reasoning under time constraints is performed on the time-series knowledge graph based on the time tag to generate the feature representation of each entity node in the time-series knowledge graph at the user's query time point.
[0072] It should be noted that the time stamp in this application represents the time constraints carried in the query request sent by the user to the system, reflecting the time range or point in time corresponding to the data that the user wants to retrieve or analyze. Specifically, it includes the absolute time value directly entered by the user, the relative time expression, the time interval, and the fuzzy time description, such as the fuzzy time description as recent or earlier. This time stamp is used to parse the user's query intent.
[0073] In some embodiments, performing time-constrained graph neural network inference on the time-series knowledge graph based on the time tag to generate feature representations of each entity node in the time-series knowledge graph at the user query time point can be achieved through the following steps:
[0074] Based on the time tag, the temporal knowledge graph is extracted to obtain the temporal subgraph of each entity node in the temporal knowledge graph at the user query time point;
[0075] Temporal neighborhood features are aggregated for each temporal subgraph to obtain the local temporal aggregated features of each entity node;
[0076] Multi-level message passing is performed on each local temporal aggregate feature to generate the feature representation of each entity node in the temporal knowledge graph at the user query time point.
[0077] In specific implementation, the temporal knowledge graph is extracted based on the time tag to obtain the temporal subgraph of each entity node in the temporal knowledge graph at the user query time point. Specifically: First, based on the time tag of the user interaction query, an adaptive extraction method of time window is adopted. Combined with the time characteristics of different types of science and technology innovation data, the extraction range is accurately set, breaking through the limitations of the traditional fixed window. For project-type data, since project implementation is mostly in quarterly phases, the effective extraction window of 3 months before and after the query time point is defined as the core to ensure coverage of the related information of the current stage of the project. For achievement-type data, since the release of achievements has a clear and specific date, an extraction window of 1 month before and after the query time point is set to accurately capture the real-time associations related to the achievements. For policy-type data, since policy implementation is mostly in annual cycles, the extraction window is set to the year in which the query time point is located to fully cover the related entities and relationships during the policy's effective period. Then, the neighbor... Using a table traversal technique, the temporal knowledge graph is filtered node-by-node and edge-by-edge. First, the effective time lifecycle ledger of each entity node is retrieved, and the start and end times in the ledger are compared with the query time point to filter out entity nodes whose lifecycles include the query time point. Then, the effective and termination periods of each relation edge are checked, and relation edges whose periods cover the query time point are retained. At the same time, the time attribute information corresponding to these entity nodes and relation edges is extracted, and the nodes are grouped into independent units according to individual entity nodes. A temporal subgraph containing only the associated nodes, relations, and time information within the effective range of the query time point is constructed for each entity node. For entities or relations with unclear time intervals, the time completion rules known in the field of science and technology innovation are adopted to supplement the most recent effective time node associated with the entity or relation, ensuring the completeness and accuracy of each temporal subgraph and completing the temporal extraction process. Other methods can also be used in other embodiments, which are not limited here.
[0078] In addition, in the specific implementation, temporal neighborhood feature aggregation is performed on each temporal subgraph to obtain the local temporal aggregated features of each entity node. Specifically, the association level between the neighboring entities and the target entity is first defined. Neighboring entities with direct relationships such as collaboration, funding, and output are classified as first-level neighbors with a weight of 0.7. Neighboring entities indirectly associated through an intermediate entity are classified as second-level neighbors with a weight of 0.3, ensuring that neighboring entities with higher degree of association have a greater impact on the features of the target entity. At the same time, a linear decay rule is used to allocate time weights. Based on the time granularity of the corresponding data, the closer the relationship effective time of the neighboring entity is to the query time, the higher the time weight coefficient. Add a time granularity, such as 1 quarter for projects, 1 month for results, and 1 year for policies, with the time weight coefficient decaying by 0.1, down to a minimum of 0.1. Then, extract the basic features of each neighboring entity, including entity type, core attributes, and relationship type. Core attributes include the domain of the organization entity, the professional title of the personnel entity, and the project approval level of the project entity. Weight these basic features and the two-dimensional weights are summed, and then the basic attribute features of the target entity at the query time point are integrated with the current status information. Through feature integration and deduplication, duplicate attribute information is removed, and finally the local temporal aggregation features of each entity node are obtained. Other methods can be used in other embodiments, which are not limited here.
[0079] Furthermore, in specific implementation, multi-level message passing is performed on each local temporal aggregation feature to generate the feature representation of each entity node in the temporal knowledge graph at the user query time point. Specifically, a layered and progressively deepening transmission mode is adopted, divided into three interconnected transmission levels: shallow, middle, and deep, breaking the limitations of traditional single-level transmission. The shallow transmission focuses on the basic association information in the local temporal aggregation features, concatenating the target entity's type, core attributes, and the basic features and direct relationships of first-level neighboring entities. After concatenation, redundant information removal technology is used to delete duplicate entity types, attributes, and other invalid data, retaining the core basic features. The middle-level transmission focuses on mining temporal evolution information, extracting the state changes of the target entity before and after the query time point, such as the changes in the project entity's initiation, implementation, and acceptance stages, the merger and split of the organization entity, and the existence of relationships corresponding to neighboring entities. For example, the duration of funding relationships and the points of change in collaborative relationships are analyzed using a feature-weighted fusion method. Features with significant state changes and key aspects of relationship continuity are given higher weights, highlighting the dynamic evolution characteristics over time. Deep transmission focuses on mining implicit association information. Through multi-step relationship deduction, it sorts out the indirect association links between the target entity and second-level and higher neighboring entities. It uses feature cross-comparison technology to cross-analyze the features of different neighboring entities to mine potential associations between features. For example, the matching association between the institution to which the researchers of a certain achievement entity belong and the support field of a certain policy entity. At the same time, invalid association features generated during the cross-comparison process are eliminated. After each layer of transmission is completed, feature verification is performed to ensure the accuracy and effectiveness of the features. Finally, the features after the three layers of transmission are integrated to form the feature representation of each entity node at the time of user query. Other methods can also be used in other embodiments, which are not limited here.
[0080] It should be noted that the temporal subgraph in this application represents the effective association range of each entity at the target time, reflecting the direct neighborhood association and time constraint state of a single entity at a specified query time, and can be used to delineate the effective boundary of feature calculation; the local temporal aggregation feature represents the core temporal features of the target entity at the query time point, reflecting the strength and dynamic changes of the association between the entity and its surrounding related objects at the corresponding time, and can achieve preliminary condensation and noise reduction of local association information; the feature representation represents the complete information of the entity at the user query time point, integrating its own attributes, neighborhood associations, temporal evolution and potential associations, reflecting the complete association logic and intrinsic attribute features of the scientific and technological innovation entity in the time dimension, and can be directly used to mine the explicit and implicit associations between entities.
[0081] In step 105, the fusion acquisition result of the multi-source heterogeneous scientific and technological innovation data is output based on the feature representation. The fusion acquisition result includes explicit and implicit associations between entities.
[0082] In some embodiments, the fusion acquisition result of the multi-source heterogeneous science and technology innovation data based on the feature representation can be achieved by the following steps:
[0083] Based on the aforementioned feature representation, the explicit and implicit associations between the entities corresponding to the multi-source heterogeneous science and technology innovation data are determined;
[0084] The fusion collection results of the multi-source heterogeneous science and technology innovation data are generated based on the explicit and implicit associations between the entities.
[0085] In specific implementation, determining the explicit and implicit associations between entities corresponding to the multi-source heterogeneous science and technology innovation data based on the feature representation is as follows: For the determination of explicit associations, a cosine similarity comparison algorithm is used. First, direct association identifiers are extracted from the feature representation of each entity, such as relational tags like "collaboration," "funding," and "output" contained in the features, as well as attribute information indicating direct association between entities. The feature representations of two entities are converted into feature vectors. By calculating the cosine similarity between the vectors, a reasonable threshold is set. Based on the entity association characteristics in the science and technology innovation field, the similarity threshold is set to 0.75, that is, the direct association identifier matching degree of two entities reaches 75% or above, and there are clear relational tags in the features, that is, it is determined that there is an explicit association between the two. For weak association scenarios such as achievements and policies, institutions and institutions, the judgment threshold is lowered to 0.65 to balance the comprehensiveness and accuracy of association identification. When the similarity of the feature vectors of two entities is higher than this threshold, and there are clear direct association identifiers in the features, it is determined that there is an explicit association between the two. At the same time, information such as time constraints and association types bound in the features are extracted to clarify the specific attributes of the explicit association, such as projects. If the feature vector similarity between an entity and an institution is higher than 0.75 and the feature contains the "funding" identifier, then an explicit funding association is determined, and the funding time interval is recorded. For determining implicit associations, a frequent itemset mining algorithm is used. First, the feature representations of all entities are decomposed to extract core attributes, such as the research field of institutional entities, the research direction of personnel entities, the technical field of achievement entities, and the support scope of policy entities. These core attributes are used as itemsets, and a support threshold is set. Based on the distribution characteristics of science and technology innovation data, a support threshold of 5% is set. Frequently occurring itemset combinations are mined. Then, combined with association rules in the science and technology innovation field (e.g., potential compatibility associations between institutional entities in the same research field and policy entities in the corresponding field, and potential output associations between personnel entities in the same research direction and related achievement entities), frequent itemset combinations are screened and verified. Combinations without actual association are eliminated, and entity pairs with potential associations are determined to be implicitly associated. Simultaneously, feature cross-validation is used to further confirm the rationality of implicit associations and avoid misjudgments. Other methods can be used in other embodiments, which are not limited here.
[0086] Furthermore, in specific implementation, the generation of the fusion collection results of the multi-source heterogeneous science and technology innovation data based on the explicit and implicit associations between entities is as follows: First, a classification and organization method is used to classify explicit associations according to relationship type. Each explicit association entry clearly marks the head entity, tail entity, association type, time constraint, and core attribute information. Implicit associations are graded according to association strength. Each implicit association entry marks the head entity, tail entity, association basis, and association strength. Then, a structured integration method is used to group by entity type. Each entity corresponds to all its explicit and implicit associations, and the core attributes and time information of the entity itself are supplemented to form a clear and complete fusion collection result. At the same time, redundant information removal technology is used to delete duplicate association entries to ensure the simplicity and accuracy of the results. Finally, the output is a collection result that can be directly used for user query and meets the needs of science and technology innovation data fusion, clearly presenting all the association relationships between entities. Other methods can also be used in other embodiments, which are not limited here.
[0087] It should be noted that the explicit associations in this application represent the directly observable entity correspondences in science and technology innovation data, reflecting the established connections of collaboration, funding, and output that objectively exist and can be directly verified among projects, institutions, personnel, achievements, and policies. They can be used to intuitively present the basic association structure of science and technology innovation data, providing users with clear and verifiable direct association evidence. The implicit associations represent the hidden attribute matching and logical linkage relationships between entities, reflecting the inherent connections and developmental relationships of different science and technology innovation entities in terms of research fields, technological directions, and policy adaptation. They can be used to mine association clues that are not directly recorded in the data, improving the comprehensiveness of science and technology innovation data association mining. The fusion collection results represent the complete association set after time-series fusion and inference calculation of multi-source heterogeneous science and technology innovation data, reflecting all association logic, dynamic characteristics, and overall association patterns of various science and technology innovation entities at the user's query time point. They can be used to meet user interactive query needs, support science and technology innovation resource analysis and decision-making, and provide a unified data foundation for science and technology innovation data management, association expansion, and in-depth application.
[0088] In some embodiments, reference Figure 3This diagram illustrates the workflow of a multi-source heterogeneous science and technology innovation data fusion acquisition system in some embodiments of this application. 110 corresponds to a multi-source heterogeneous science and technology innovation data acquisition scenario, encompassing various entity elements and multi-source data sources within science and technology innovation-related business scenarios such as port operations. 120 is a data acquisition device responsible for acquiring multi-source heterogeneous science and technology innovation business data from scenario 110. 130 is a data fusion processing device, performing entity association recognition, temporal fusion, and graph neural network inference on the acquired multi-source data to achieve structured integration and implicit association mining of the data. 140 is a data visualization terminal, used to present the processed fusion acquisition results in an intuitive form. Data interaction between modules is achieved through bidirectional arrows, while data feedback and iterative optimization are realized through a closed-loop link at the bottom, fully covering the entire process from data acquisition and fusion processing to visualization display, providing system support for the efficient acquisition, integration, and application of science and technology innovation data.
[0089] Furthermore, in another aspect of this application, in some embodiments, this application provides a fusion acquisition system for multi-source heterogeneous scientific and technological innovation data, with reference to... Figure 4 The figure is a schematic diagram of the structure of a fusion acquisition system for multi-source heterogeneous scientific and technological innovation data according to some embodiments of this application. The fusion acquisition system 400 for multi-source heterogeneous scientific and technological innovation data includes: a construction module 401, a processing module 402, and an execution module 403, which are described below:
[0090] Construction module 401, in this application, is mainly used to construct the initial framework of the knowledge graph in the field of science and technology innovation;
[0091] Processing module 402, in this application, is used to acquire multi-source heterogeneous science and technology innovation data, perform entity association identification on the multi-source heterogeneous science and technology innovation data, and generate multiple fact triples of the multi-source heterogeneous science and technology innovation data.
[0092] It should be noted that the processing module 402 in this application is also used to perform temporal fusion of each fact triple based on the initial framework of the knowledge graph in the science and technology innovation field to form a temporal knowledge graph of the multi-source heterogeneous science and technology innovation data.
[0093] Additionally, it should be noted that the processing module 402 in this application is also used to obtain the time tag when the user interacts with the multi-source heterogeneous science and technology innovation data, and to perform graph neural network reasoning under time constraints on the time-series knowledge graph based on the time tag, so as to generate the feature representation of each entity node in the time-series knowledge graph at the time point of the user query.
[0094] The execution module 403 in this application is mainly used to output the fusion acquisition result of the multi-source heterogeneous scientific and technological innovation data based on the feature representation. The fusion acquisition result includes explicit and implicit associations between entities.
[0095] In addition, this application also provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described method for fusion acquisition of multi-source heterogeneous scientific and technological innovation data.
[0096] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing a method for fusing and acquiring multi-source heterogeneous scientific and technological innovation data, according to some embodiments of this application. The method for fusing and acquiring multi-source heterogeneous scientific and technological innovation data in the above embodiments can be achieved through... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.
[0097] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).
[0098] The communication bus 502 can be used to transmit information between the aforementioned components.
[0099] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.
[0100] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The method used in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.
[0101] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.
[0102] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).
[0103] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.
[0104] In addition, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for fusion acquisition of multi-source heterogeneous scientific and technological innovation data.
[0105] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0106] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for fusing and acquiring multi-source heterogeneous scientific and technological innovation data, characterized in that, Includes the following steps: An initial framework for constructing a knowledge graph in the field of science and technology innovation; Acquire multi-source heterogeneous science and technology innovation data, perform entity association identification on the multi-source heterogeneous science and technology innovation data, and generate multiple fact triples of the multi-source heterogeneous science and technology innovation data; Based on the initial framework of the knowledge graph in the field of science and technology innovation, the fact triples are fused in time to form the time-series knowledge graph of the multi-source heterogeneous science and technology innovation data. The time tag of the user’s interactive query on the multi-source heterogeneous science and technology innovation data is obtained, and the graph neural network reasoning under time constraints is performed on the time-series knowledge graph based on the time tag to generate the feature representation of each entity node in the time-series knowledge graph at the user’s query time point. Based on the aforementioned feature representation, the fusion collection results of the multi-source heterogeneous science and technology innovation data are output, and the fusion collection results include explicit and implicit associations between entities.
2. The method of claim 1, wherein, The initial framework includes entity types, relationship types, and time dimension constraints. The entity types include at least project entities, institutional entities, personnel entities, outcome entities, and policy entities. The relationship types include at least collaborative relationships, funding relationships, output relationships, and association relationships. The time dimension constraints are used to record the effective time interval of an entity or relationship.
3. The method of claim 1, wherein, The multi-source heterogeneous science and technology innovation data includes structured project data, semi-structured annual report data, and unstructured text data.
4. The method of claim 1, wherein, The initial framework for constructing a knowledge graph in the field of science and technology innovation specifically includes: Determine the entity types and relationship types of the knowledge graph in the field of science and technology innovation; Determine the temporal dimension constraints for entities and relationships in the knowledge graph of the science and technology innovation field; An initial framework for a knowledge graph in the field of science and technology innovation is constructed based on the entity type, the relationship type, and the time dimension constraint.
5. The method of claim 1, wherein, Entity association identification is performed on the multi-source heterogeneous science and technology innovation data to generate multiple fact triples of the multi-source heterogeneous science and technology innovation data, specifically including: An initial set of fact triples was identified from the multi-source heterogeneous science and technology innovation data; The initial set of fact triples is associated and merged to obtain multiple fact triples of the multi-source heterogeneous science and technology innovation data.
6. The method of claim 1, wherein, Based on the initial framework of the knowledge graph in the science and technology innovation field, the various fact triples are fused in a time series to form the time series knowledge graph of the multi-source heterogeneous science and technology innovation data, specifically including: Based on the initial framework of the knowledge graph in the science and technology innovation field, the types of each fact triple are verified to obtain each normalized triple; Temporal correlation is performed on each normalized triple to obtain the temporal network structure; The temporal network structure and the initial framework of the science and technology innovation domain knowledge graph are indexed and fused to obtain the temporal knowledge graph of the multi-source heterogeneous science and technology innovation data.
7. The method of claim 1, wherein, Based on the time tag, performing time-constrained graph neural network inference on the temporal knowledge graph to generate feature representations of each entity node in the temporal knowledge graph at the user query time point specifically includes: Based on the time tag, the temporal knowledge graph is extracted to obtain the temporal subgraph of each entity node in the temporal knowledge graph at the user query time point; Temporal neighborhood features are aggregated for each temporal subgraph to obtain the local temporal aggregated features of each entity node; Multi-level message passing is performed on each local temporal aggregate feature to generate the feature representation of each entity node in the temporal knowledge graph at the user query time point.
8. A multi-source heterogeneous scientific and technological data-oriented fusion collection system, characterized in that, include: The building module is used to construct the initial framework for a knowledge graph in the field of science and technology innovation; The processing module is used to acquire multi-source heterogeneous science and technology innovation data, perform entity association recognition on the multi-source heterogeneous science and technology innovation data, and generate multiple fact triples of the multi-source heterogeneous science and technology innovation data. The processing module is also used to perform temporal fusion of each fact triple based on the initial framework of the knowledge graph in the science and technology innovation field to form a temporal knowledge graph of the multi-source heterogeneous science and technology innovation data. The processing module is also used to obtain the time tag when the user interacts with the multi-source heterogeneous science and technology innovation data, and to perform graph neural network reasoning under time constraints on the time-series knowledge graph based on the time tag, so as to generate the feature representation of each entity node in the time-series knowledge graph at the time point of the user query. An execution module is used to output the fusion acquisition results of the multi-source heterogeneous science and technology innovation data based on the feature representation. The fusion acquisition results include explicit and implicit associations between entities.
9. A computer device, comprising: The computer device includes a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the fusion acquisition method for multi-source heterogeneous scientific and technological innovation data as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the fusion acquisition method for multi-source heterogeneous scientific and technological innovation data as described in any one of claims 1 to 7.