A method and system for constructing a scientific research achievement field knowledge graph based on multi-source data fusion

Abstract

The application discloses a scientific research achievement field knowledge graph construction method and system based on multi-source data fusion, and the multi-source data is characterized by vectorization; based on mapping to a unified semantic space by a transformation matrix, a local neighborhood topological index is calculated to check consistency; a vector distance is calculated, evidence theory fusion is performed to obtain a trust divergence, and a change instruction is generated; and topological compression is performed to generate a space-time difference index. The application also constructs a closed-loop feedback mechanism, dynamically updates the transformation matrix based on the trust divergence, and synchronizes the flow and batch states through consistency checking. The application can guarantee semantic mapping topological consistency, realize adaptive optimization of the fusion model, and efficiently reconstruct the evolution history of the graph.

Description

Technical Field

[0001] This invention relates to knowledge graph and data processing technologies, specifically to a method and system for constructing a knowledge graph in the field of scientific research results through multi-source data fusion. Background Technology

[0002] With the increasing frequency of scientific research activities, the amount of data generated from scientific research outcomes (such as academic papers, patent documents, research projects, and science and technology awards) is growing explosively. This data comes from a wide range of sources, is heterogeneous in format (including structured databases, semi-structured metadata, and unstructured text), and exhibits strong temporal evolution characteristics. Effectively integrating this multi-source, heterogeneous data to construct a high-quality knowledge graph of scientific research outcomes is of great significance for scientific research management, technology trend analysis, and talent evaluation.

[0003] Existing knowledge graph construction methods suffer from several technical shortcomings when handling multi-source data fusion, including low semantic mapping fidelity, lack of closed-loop feedback in the fusion process, and difficulty in backtracking graph evolution. Firstly, regarding semantic mapping of multi-source data, existing technologies typically employ static vectorized models or simple linear transformations to map data from different sources to a unified space. However, this approach often ignores the topological characteristics of the data in the original space. For example, two entities semantically close in their original data sources may become distant in the unified semantic space after simple linear mapping, leading to the loss or distortion of semantic relationships and severely impacting the accuracy of subsequent entity alignment. Secondly, most existing fusion models are unidirectional pipeline processes, following a fixed workflow of "data preprocessing - entity alignment - conflict resolution - data storage." This open-loop structure lacks a feedback mechanism. When evidence conflicts arise during the fusion process (such as inconsistent descriptions of the same entity from different sources), the system often can only resolve them using simple strategies like voting or weighted averaging, failing to feed this conflict information back to the upstream mapping model for parameter adjustment. This prevents the system from optimizing its mapping capabilities through errors in the processing, making it difficult for the fusion quality to adaptively improve with increasing data volume. Finally, traditional knowledge graph storage is primarily geared towards static graph queries, typically only saving a snapshot of the graph's latest state. However, research fields exhibit strong evolutionary characteristics (e.g., the rise, development, and decline of a technological topic, or changes in the membership of a research team).

[0004] For example, Chinese patent application CN119783037A discloses a method for multi-source heterogeneous data fusion and processing based on big data. Its core is the fusion and processing of multi-source data, i.e., how to fuse and process data from multiple sources, but it does not explain the purpose of this fusion and processing. Chinese patent application CN120523966A discloses a method and system for constructing knowledge graphs based on large language model technology. Its core is using large language models to construct knowledge graphs, but it neither clarifies the data type of the knowledge graph nor explains the purpose of constructing it. As another example, Chinese patent application CN120067984A discloses a method and system for fusion of multi-source heterogeneous knowledge graph data. Its core is the method for fusion of multi-source heterogeneous knowledge graph data, focusing on fusing multi-source heterogeneous data.

[0005] In summary, existing technologies are insufficient for efficiently reconstructing the spectral structure within a specified historical time window, and cannot meet the needs for refined tracing and analysis of the evolution path of scientific research results.

[0006] Therefore, there is an urgent need for a method to construct a multi-source data fusion knowledge graph that can guarantee the topological consistency of semantic mapping, has closed-loop feedback adaptive capability, and supports spatiotemporal dynamic reconstruction. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies and provide a method and system for constructing a domain knowledge graph of scientific research results based on multi-source data fusion.

[0008] Technical solution: The present invention provides a method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion, comprising the following steps:

[0009] Step 1: Data preprocessing. Obtain multi-source entity data streams in the field of scientific research results, perform standardization and feature vectorization on the multi-source entity data streams, and generate an initial feature vector set. ;

[0010] Step 2, semantic mapping and verification, that is, mapping the initial feature vector set based on the pre-constructed transformation matrix. Mapping to a unified latent semantic space yields aligned semantic vectors, and the topological structure index of the local neighborhood before and after mapping is calculated; if the correlation coefficient of the topological structure index is lower than a preset consistency threshold, the processing flow is blocked.

[0011] Step 3, Entity Fusion and Weight Calculation: Calculate the geometric distance between the multi-source semantic vectors of the same entity, perform evidence theory fusion calculation to obtain the fusion result and trust divergence, and calculate the semantic weight of the relationship between entities to generate a topology change instruction package;

[0012] Step 4: Index storage. Write the obtained topology change instruction package into the streaming buffer, perform topology compression processing to generate spatiotemporal difference index blocks, and write them into the persistent storage engine.

[0013] Step 5, Closed-loop feedback and consistency guarantee steps: Determine whether the trust divergence output in step 3 exceeds the preset conflict threshold; if it does, generate a conflict gradient signal and convert it into an incremental update of the transformation matrix, and re-execute steps 2 to 3; at the same time, compare the topology state of the streaming differential layer and the batch reference layer through the consistency verification mechanism, and trigger re-arbitration if they are inconsistent.

[0014] Step 6: Coordinate output, maintain the execution sequence of steps 1 to 5 of the global state machine management, and respond to evolution path query requests. Dynamically reconstruct the topological structure of the entity association of scientific research results within the specified time window by superimposing spatiotemporal difference index blocks.

[0015] Furthermore, the detailed method of step 1 is as follows:

[0016] Step 1.1: The multi-source entity data stream includes academic paper data stream in JSON format, patent data stream in XML format, and scientific research news stream in plain text format; the academic paper stream includes fields for title, abstract, author, and keywords, and the patent data stream includes fields for patent name, abstract, inventor, and claims; the original multi-source entity data stream is processed by removing HTML tags, filtering stop words, unifying time format, and normalizing personnel name format.

[0017] Step 1.2: For each research result entity, construct its text content as an input text sequence. , , Indicates the first in the input text sequence Each word element;

[0018] Step 1.3: Input the text sequence T into the pre-trained BERT model and extract the vector corresponding to the CLS marker in the last hidden layer of the pre-trained BERT model as the text semantic vector. For highly structured features such as subject classification codes, keyword categories, and patent classification numbers, One-Hot encoding is used to generate additional vectors. ;Transfer text semantic vectors With additional vectors After concatenation, the vectors are fed into a fully connected layer for dimensionality reduction, resulting in an initial feature vector of dimension d. ; ;

[0019] in, This represents a fully connected dimensionality reduction function. This represents a vector concatenation operation;

[0020] Step 1.4: The initial feature vector set is formed by the initial feature vectors corresponding to all scientific research achievement entities.

[0021] ;

[0022] in, This indicates the number of scientific research results entities in the current batch. This represents the m-th initial eigenvector in the initial eigenvector set.

[0023] Furthermore, the detailed steps of semantic mapping and verification in step 2 are as follows:

[0024] Step 2.1: To support dynamic adaptive updates of the semantic space, maintain a dynamic k-nearest neighbor graph with mapped semantic vectors as nodes and nearest neighbor relationships as edges. For each mapped semantic vector In the unified latent semantic space, retrieve its k nearest neighbors (k can take the value 10) and establish edge connections; when a weight adjustment is received, update the semantic mapping relationship using a low-rank update method based on the graph Laplacian matrix L, and the update formula is:

[0025] ;

[0026] in, Represents the graph Laplace matrix. and This represents the low-rank update vector constructed from the weight adjustments. express Transpose of;

[0027] Step 2.2: Construct Vietoris-Rips complexes for both the local neighborhood set before and after mapping; [The text then abruptly shifts to a different topic:] ...filtering parameters... Below, when the distance between two points is less than When three points are connected in pairs, a 2-simple is formed.

[0028] Then, based on the Vietoris-Rips complex, the topological structure indices before mapping are calculated respectively. and mapped topology index ;

[0029] ;

[0030] ;

[0031] in, This represents the number of connected components in the local neighborhood topological complex before mapping. This represents the number of ring structures in the local neighborhood topological complex before mapping. This represents the number of connected components in the local neighborhood topological complex after mapping. This represents the number of ring structures in the local neighborhood topological complex after mapping.

[0032] Step 2.3: Calculate the Pearson correlation coefficient of the topological structure indices before and after mapping. :

[0033] ;

[0034] in, Describing covariance, Indicates standard deviation;

[0035] like Less than the preset consistency threshold If the current mapping causes semantic topology distortion, the current batch of data will be blocked from flowing and the data will be stored in the exception queue; if Not less than Then the mapped semantic vector is input into step 3.

[0036] Furthermore, the detailed process of entity fusion and weight calculation in step 3 is as follows:

[0037] Step 3.1: In the unified latent semantic space, for any two candidate entities... and Corresponding semantic vector and Calculate the Euclidean distance between the two. The formula is as follows:

[0038] ;

[0039] Step 3.2: The Dempster-Shafer evidence theory is used to fuse multiple sources of evidence. For two sources of evidence... and After combination, the proposition support for:

[0040] ;

[0041] in, The coefficient of evidence conflict, also known as the confidence divergence, is calculated using the following formula:

[0042] ;

[0043] Then, based on the fusion confidence Conf(a, b) obtained from the fusion and the entity fusion threshold... The size relationship is used to determine whether candidate entity a and candidate entity b are the same entity;

[0044] Step 3.3: First, use regular expressions to match citation markers in the text and extract the text before and after the citation anchor point of a preset length as the citation context; then, perform word segmentation and TF-IDF vectorization on the citation context and the citation text summary respectively, and calculate their cosine similarity. Next, for the entity pairs identified as related in step 3.2, cosine similarity is used as the basis for further analysis. Calculate the semantic weights of relationships between entities ;

[0045] ;

[0046] Finally, a topology change instruction package is generated based on the entity fusion results and relation weights. ;

[0047] ;

[0048] in, Represents a timestamp. Indicates the operation type. Indicates entity identifier, Corresponding to the fusion confidence level Conf(a, b), The type of relation is indicated by the operation type, which includes LINK, UNLINK, and UPDATE_WEIGHT. LINK indicates adding a new association between two entity nodes, UNLINK indicates deleting an existing association between two entity nodes, and UPDATE_WEIGHT indicates updating the semantic weight or confidence of the association while preserving the existing association.

[0049] Furthermore, the detailed method for index storage in step 4 is as follows:

[0050] Step 4.1: The topology change command packet Pcmd is written to a streaming buffer built on Kafka to decouple the entity fusion computing layer and the persistent storage layer. Let's define the set of topology change commands received within a single batch time window. for:

[0051] ;

[0052] Where L represents the number of topology change instructions received within the time window;

[0053] Step 4.2: Before writing the data to the disk, perform topology compression on the topology change instruction set P: If there are multiple update instructions for the same entity pair within the same batch window, retain the one with the latest timestamp; if there are inverse operations of adding an edge first and then deleting an edge, cancel them out in the instruction set.

[0054] Then the compressed instruction set Calculate the SHA-256 hash value as the fingerprint of the spatiotemporal differential index block. Constructing the spatiotemporal difference index block B:

[0055] ;

[0056] in, Indicates the time bucket identifier;

[0057] Step 4.3: Write the spatiotemporal difference index block B into the RocksDB persistent storage engine based on LSM-Tree, storing the key. Designed as follows:

[0058] ;

[0059] in, Indicates the map partition identifier. Indicates a time bucket. Indicates the entity identifier; Value is the compressed Protocol Buffers format topology instruction block;

[0060] The persistent storage engine here uses a log-structured merged tree storage structure, which supports sequential append writes and range queries through ordered key-value indexes.

[0061] Furthermore, the detailed steps of step 5, closed-loop feedback and consistency guarantee, are as follows:

[0062] Step 5.1: Monitor the trust divergence K output in step S3 in real time. The preset conflict threshold δ is 0.6. If K≤δ is satisfied, the current fusion result is considered credible and closed-loop feedback is not triggered. If K>δ is satisfied, it indicates that there is a serious conflict between different evidence sources and the closed-loop feedback mechanism is triggered.

[0063] Step 5.2: Determine the conflict direction vector based on external evidence. Generate conflict gradient signals ; Calculate the incremental update of the transformation matrix M based on the conflict gradient signal. Perform matrix update After the update is completed, the process is rolled back, and steps 2 and 3 are re-executed using the updated transformation matrix, so that the batch of data can be re-mapped and entity fused in the corrected semantic space; at the same time, conflict gradient signals are... As a back pressure control signal, it is sent to the data preprocessing module or streaming buffer to reduce the data writing frequency;

[0064] When strong external evidence indicates that candidate entities a and b should be the same entity, but the current semantic distance is too large, the conflict gradient direction is set to the attraction direction; when strong external evidence indicates that candidate entities a and b should not be the same entity, but the current semantic distance is too small, the conflict gradient direction is set to the repulsion direction; the conflict gradient signal is used to determine the weight adjustment amount, and the rule is: if the signal indicates that the two entities should be separated in the semantic space, the edge weight on the connection path is reduced; if the signal indicates that the two entities should be aggregated, the edge weight on the connection path is increased.

[0065] ,

[0066] ,

[0067] ,

[0068] in, This represents the conflict gradient direction coefficient, used to determine whether the current feedback is an attractive or repulsive direction. This represents the learning rate or step size coefficient, used to control the updating of the transformation matrix based on the conflict gradient signal each time. The range;

[0069] Step 5.3, Maintain flow cytometry differential stratification and batch baseline layer The batch baseline layer periodically merges spatiotemporal difference index blocks to generate a graph snapshot. At the same time section t, MerkleTree root hashes of the topological states of the two layers of graphs are constructed respectively. If the root hashes are inconsistent, it is determined that the streaming differential layer has a state drift. The graph snapshot of the batch baseline layer is used as the authoritative basis to trigger a re-arbitration process. The stable graph snapshot in the batch baseline layer is used as the authoritative source to cover or correct the erroneous state in the streaming differential layer. If the root hashes are consistent, it is determined that the streaming differential layer is consistent with the batch baseline layer.

[0070] Furthermore, the detailed steps for coordinating the output in step 6 are as follows:

[0071] Global state machine It can be configured with multiple working states:

[0072] When the topology consistency check fails in step 2, the global state machine... Entering ERROR status;

[0073] When trust divergence occurs in step 5 Exceeding the conflict threshold At that time, the global state machine Switch to FEEDBACK state and revert the control process to step 2;

[0074] When the topological state verification of the two-layer graph is inconsistent, the global state machine Transition to CONSISTENCY status and trigger re-arbitration;

[0075] Once steps 1 through 5 have all been executed successfully, the global state machine... Transition to the QUERY_READY state to receive and respond to graph evolution path queries.

[0076] This invention also discloses a knowledge graph construction system for scientific research results based on multi-source data fusion, including a data preprocessing module, a semantic mapping and verification module, an entity fusion and weight calculation module, an index storage module, a closed-loop feedback and consistency guarantee module, and a coordination output module;

[0077] The data preprocessing module acquires multi-source scientific research entity data streams, performs standardization and feature vectorization processing, and generates an initial feature vector set.

[0078] The semantic mapping and verification module maps the initial feature vector set to the unified latent semantic space based on the pre-constructed transformation matrix to obtain the aligned semantic vectors, and calculates the topological structure index of the local neighborhood before and after mapping. If the correlation coefficient of the topological structure index is lower than the preset consistency threshold, the processing flow is blocked.

[0079] The entity fusion and weight calculation module calculates the vector geometric distance between the multi-source semantic vectors of the same entity, performs evidence theory fusion calculation to obtain the fusion result and trust divergence, and calculates the semantic weight of the relationship between entities to generate a topology change instruction package.

[0080] The index storage module writes the topology change instruction package into the streaming buffer, performs topology compression processing to generate spatiotemporal differential index blocks, and writes them into the persistent storage engine.

[0081] The closed-loop feedback and consistency guarantee module determines whether the trust divergence exceeds the preset conflict threshold. If so, it generates a conflict gradient signal and converts it into an incremental update of the transformation matrix, triggering the semantic mapping and verification module and the entity fusion and weight calculation module to re-execute. The consistency verification mechanism compares the topological state of the streaming differential layer and the batch benchmark layer. If they are inconsistent, it triggers re-arbitration.

[0082] The coordination output module maintains the global state machine to manage the execution timing of each module, responds to evolution path query requests, and dynamically reconstructs the topological structure of the entity association of scientific research results within a specified time window by superimposing spatiotemporal difference index blocks.

[0083] Furthermore, the index storage module includes a streaming buffer unit, a topology compression unit, a hash calculation unit, and a persistent storage unit; the streaming buffer unit is built using Kafka, and the persistent storage unit uses RocksDB based on the LSM-Tree architecture, and an ordered key-value index is constructed according to the combination of graph partition identifier, time bucket identifier, and entity identifier.

[0084] The closed-loop feedback and consistency assurance module includes a conflict monitoring unit, a matrix update unit, a state comparison unit, and an arbitration unit. The state comparison unit completes the topological state comparison between the streaming differential layer and the batch baseline layer based on the MerkleTree tree hash structure. When the states are inconsistent, the arbitration unit corrects the streaming differential layer data using the map snapshot of the batch baseline layer as the standard.

[0085] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0086] First, it improves the fidelity and accuracy of semantic mapping. This invention innovatively introduces a topological structure index verification mechanism in the semantic mapping stage. By calculating the topological features of the local neighborhood before and after mapping (such as low-order topological features like the Betti number), it forces the mapping process to maintain the inherent geometric structure of the original data. Once the topological correlation coefficient falls below a threshold, the process is blocked, effectively avoiding the loss of semantic information due to mapping distortion and ensuring high-quality alignment of heterogeneous data in a unified space.

[0087] Second, online adaptive optimization of the fusion system is achieved. Unlike traditional unidirectional pipelines, this invention constructs a closed-loop feedback mechanism of "trust divergence - conflict gradient - transformation matrix update". When high conflict occurs in the fusion result (large trust divergence), the system can backpropagate to generate gradient signals and dynamically adjust the front-end mapping parameters (such as nearest neighbor graph weights) to adaptively correct mapping deviations. This mechanism enables the system to continuously evolve during use, and as the amount of data processed increases, the fusion accuracy continuously improves due to the adaptive adjustment of the model.

[0088] Third, it supports fine-grained spatiotemporal dynamic reconstruction and evolutionary analysis. This invention abandons traditional static snapshot storage and adopts a spatiotemporal differential indexing mechanism, compressing each change in the knowledge graph into a timestamped differential index block. Combined with the overlay logic of the coordination output module, it quickly reconstructs the topological structure of research results within any specified time window, thereby finely displaying the evolution of technical routes, changes in research team collaborations, and other dynamic processes, greatly enhancing the application value of knowledge graphs in the field of research trend analysis.

[0089] Fourth, it ensures consistency between streaming and batch processing. By introducing a consistency verification mechanism based on a tree-structured hash (such as Merkle Tree), the topology of the streaming differential layer and the batch baseline layer are compared in real time, effectively solving the common double-write inconsistency problem in the Lambda architecture. This ensures that the system provides high real-time write performance while ensuring eventual data consistency and reliability. Attached Figure Description

[0090] Figure 1 This is an overall flowchart of the present invention;

[0091] Figure 2 This is a detailed flowchart of the semantic mapping and verification process in this invention;

[0092] Figure 3 This is a schematic diagram illustrating the principle of entity fusion and weight calculation in this invention;

[0093] Figure 4 This is a logical schematic diagram of index storage and spatiotemporal reconstruction in this invention;

[0094] Figure 5 This is a structural block diagram of an embodiment of the present invention. Detailed Implementation

[0095] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0096] like Figure 1 As shown, the method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion of the present invention includes the following steps:

[0097] Step 1: Obtain multi-source entity data streams in the field of scientific research results, and preprocess them to generate an initial feature vector set. ;

[0098] Step 2, semantic mapping and verification, is to map the initial feature vector set to the unified latent semantic space based on the pre-constructed transformation matrix to obtain the aligned semantic vectors, and calculate the topological structure index of the local neighborhood before and after the mapping.

[0099] Step 3, Entity Fusion and Weight Calculation: Calculate the geometric distance between the semantic vectors of the same entity from multiple sources, fuse the evidence theory to obtain the fusion result and trust divergence, and calculate the semantic weight of the relationship between entities to generate a topology change instruction package.

[0100] Step 4: Index storage. Write the topology change instruction package into the streaming buffer, perform topology compression to generate a spatiotemporal difference index block, and write it into the persistent storage engine.

[0101] Step 5: Closed-loop feedback and consistency guarantee. First, determine whether the trust divergence exceeds the preset conflict threshold. If it does, generate a conflict gradient signal and convert it into an incremental update of the transformation matrix, and repeat steps 2 to 3. Then, compare the topology of the streaming differential layer and the batch reference layer through the consistency verification mechanism. If they are inconsistent, trigger re-arbitration.

[0102] Step 6: Coordinate output, maintain the execution sequence of steps 1 to 5 of the global state machine management, and respond to evolution path query requests. Dynamically reconstruct the topological structure of the entity association of scientific research results within the specified time window by superimposing spatiotemporal difference index blocks.

[0103] To facilitate understanding and reproduction of this invention by those skilled in the art, the relevant technical terms are defined as follows:

[0104] Trust divergence: In Dempster-Shafer evidence theory, it refers to the degree of conflict between the support of different sources of evidence for the same proposition. It is calculated by summing the products of m1(X) and m2(Y) for all cases where the intersection of X and Y is an empty set. The value ranges from 0 to 1, and the larger the value, the more serious the evidence conflict.

[0105] Topology change instruction package: A data structure used to describe changes to the knowledge graph structure. It includes fields such as timestamp, operation type (LINK / UNLINK / UPDATE_WEIGHT), entity identifier, and weight value. It serves as the data transfer carrier between streaming and batch processing.

[0106] Spatiotemporal differential index block: A persistent storage unit generated by compressing topology change instructions within the same time window. It contains the hash fingerprint and timestamp information of the instruction set and supports efficient retrieval by time range.

[0107] Topological structure index: A mathematical indicator used to measure the local neighborhood structure characteristics of data points. This invention uses the Betti number as a specific implementation, wherein... Indicates the number of connected components. Indicates the number of ring structures.

[0108] Conflict gradient signal: A feedback signal generated when the trust divergence exceeds a threshold. It is used to guide the incremental update direction of the transformation matrix. A positive direction indicates repulsion and a negative direction indicates attraction.

[0109] The detailed method for step 1 in this embodiment is as follows:

[0110] Step 1.1: The multi-source entity data stream includes academic paper data stream in JSON format, patent data stream in XML format, and scientific research news stream in plain text format; the academic paper stream includes fields for title, abstract, author, and keywords, and the patent data stream includes fields for patent name, abstract, inventor, and claims; the original multi-source entity data stream is processed by removing HTML tags, filtering stop words, unifying time format, and normalizing personnel name format.

[0111] Step 1.2: For each research result entity, construct its text content as an input text sequence. , , Indicates the first in the input text sequence Each word element;

[0112] Step 1.3: Input the text sequence T into the pre-trained BERT model and extract the vector corresponding to the CLS marker in the last hidden layer of the pre-trained BERT model as the text semantic vector. For highly structured features such as subject classification codes, keyword categories, and patent classification numbers, One-Hot encoding is used to generate additional vectors. ;Transfer text semantic vectors With additional vectors After concatenation, the vectors are fed into a fully connected layer for dimensionality reduction, resulting in an initial feature vector of dimension d. ; ;

[0113] in, This represents a fully connected dimensionality reduction function. This represents a vector concatenation operation;

[0114] Step 1.4: The initial feature vector set is formed by the initial feature vectors corresponding to all scientific research achievement entities.

[0115] ;

[0116] in, This indicates the number of scientific research achievement entities in the current batch.

[0117] like Figure 2 As shown, the detailed steps of semantic mapping and verification in step 2 are as follows:

[0118] Step 2.1: Maintain a dynamic k-nearest neighbor graph with mapped semantic vectors as nodes and nearest neighbor relationships as edges. For each mapped semantic vector In the unified latent semantic space, its k nearest neighbors are retrieved and edge connections are established; when a weight adjustment is received, the semantic mapping relationship is updated using a low-rank update method based on the graph Laplacian matrix L, and the update formula is as follows:

[0119] ;

[0120] in, Represents the graph Laplace matrix. and This represents the low-rank update vector constructed from the weight adjustments;

[0121] Step 2.2: Construct Vietoris-Rips complexes for both the local neighborhood set before and after mapping; [The text then abruptly shifts to a different topic:] ...filtering parameters... Below, when the distance between two points is less than When three points are connected in pairs, a 2-simple is formed.

[0122] Then, based on the Vietoris-Rips complex, the topological structure indices before mapping are calculated respectively. and mapped topology index ;

[0123] ;

[0124] ;

[0125] in, This represents the number of connected components in the local neighborhood topological complex before mapping. This represents the number of ring structures in the local neighborhood topological complex before mapping. This represents the number of connected components in the local neighborhood topological complex after mapping. This represents the number of ring structures in the local neighborhood topological complex after mapping;

[0126] Step 2.3: Calculate the Pearson correlation coefficient of the topological structure indices before and after mapping. :

[0127] ;

[0128] in, Describing covariance, Indicates standard deviation;

[0129] like Less than the preset consistency threshold If the current mapping causes semantic topology distortion, the current batch of data will be blocked from flowing and the data will be stored in the exception queue; if Not less than Then the mapped semantic vector is input into step 3.

[0130] like Figure 3As shown, the detailed process of entity fusion and weight calculation in step 3 of the embodiment is as follows:

[0131] Step 3.1: In the unified latent semantic space, for any two candidate entities... and Corresponding semantic vector and Calculate the Euclidean distance between the two. The formula is as follows:

[0132] ;

[0133] Step 3.2: The Dempster-Shafer evidence theory is used to fuse multiple sources of evidence. For two sources of evidence... and After combination, the proposition support for:

[0134] ;

[0135] in, The coefficient of evidence conflict, also known as the confidence divergence, is calculated using the following formula:

[0136] ;

[0137] Then, based on the fusion confidence Conf(a, b) obtained from the fusion and the entity fusion threshold... The size relationship is used to determine whether candidate entity a and candidate entity b are the same entity;

[0138] Step 3.3: First, use regular expressions to match citation markers in the text and extract the text before and after the citation anchor point of a preset length as the citation context; then, perform word segmentation and TF-IDF vectorization on the citation context and the citation text summary respectively, and calculate their cosine similarity. Next, for the entity pairs identified as related in step 3.2, cosine similarity is used as the basis for further analysis. Calculate the semantic weights of relationships between entities ;

[0139] ;

[0140] Finally, a topology change instruction package is generated based on the entity fusion results and relation weights. ;

[0141] ;

[0142] in, Represents a timestamp. Indicates the operation type. Indicates entity identifier, This indicates the relationship type, and the operation types include LINK, UNLINK, and UPDATE_WEIGHT. LINK adds a new association between two entity nodes, UNLINK deletes an existing association between two entity nodes, and UPDATE_WEIGHT updates the semantic weight or confidence of the association while preserving it.

[0143] like Figure 4 As shown, the detailed method for index storage in step 4 of the embodiment is as follows:

[0144] Step 4.1: The topology change command packet Pcmd is written to a streaming buffer built on Kafka to decouple the entity fusion computing layer and the persistent storage layer. Let's define the set of topology change commands received within a single batch time window. for:

[0145] ;

[0146] Where L represents the number of topology change instructions received within the time window;

[0147] Step 4.2: Before writing the data to the disk, perform topology compression on the topology change instruction set P: If there are multiple update instructions for the same entity pair within the same batch window, retain the one with the latest timestamp; if there are inverse operations of adding an edge first and then deleting an edge, cancel them out in the instruction set.

[0148] Then the compressed instruction set Calculate the SHA-256 hash value as the fingerprint of the spatiotemporal differential index block. Constructing the spatiotemporal difference index block B:

[0149] ;

[0150] in, Indicates the time bucket identifier;

[0151] Step 4.3: Write the spatiotemporal difference index block B into the RocksDB persistent storage engine based on LSM-Tree, storing the key. Designed as follows:

[0152] ;

[0153] in, Indicates the map partition identifier. Indicates a time bucket. Represents the entity identifier; Value is the compressed Protocol Buffers format topology instruction block.

[0154] The detailed steps of the closed-loop feedback and consistency guarantee in step 5 above are as follows:

[0155] Step 5.1: Monitor the trust divergence K output in step S3 in real time, and preset the conflict threshold δ; if K≤δ, the current fusion result is considered credible and closed-loop feedback is not triggered; if K>δ, it indicates that there is a serious conflict between different evidence sources and the closed-loop feedback mechanism is triggered.

[0156] Step 5.2: Determine the conflict direction vector based on external evidence. Generate conflict gradient signals ; Calculate the incremental update of the transformation matrix M based on the conflict gradient signal. Perform matrix update After the update is completed, the process is rolled back, and steps 2 and 3 are re-executed using the updated transformation matrix, so that the batch of data can be re-mapped and entity fused in the corrected semantic space; at the same time, conflict gradient signals are... As a back pressure control signal, it is sent to the data preprocessing module or streaming buffer to reduce the data writing frequency;

[0157] ,

[0158] ,

[0159] ,

[0160] in, This represents the conflict gradient direction coefficient, used to determine whether the current feedback is an attractive or repulsive direction. This represents the learning rate or step size coefficient, used to control the updating of the transformation matrix based on the conflict gradient signal each time. The range;

[0161] Step 5.3, Maintain flow cytometry differential stratification and batch baseline layer The batch baseline layer periodically merges spatiotemporal difference index blocks to generate a graph snapshot. At the same time section t, MerkleTree root hashes of the topological states of the two layers of graphs are constructed respectively. If the root hashes are inconsistent, it is determined that the streaming differential layer has a state drift. The graph snapshot of the batch baseline layer is used as the authoritative basis to trigger a re-arbitration process. The stable graph snapshot in the batch baseline layer is used as the authoritative source to cover or correct the erroneous state in the streaming differential layer. If the root hashes are consistent, it is determined that the streaming differential layer is consistent with the batch baseline layer.

[0162] The detailed steps for coordinating the output in step 6 of the embodiment are as follows:

[0163] Global state machine It can be configured with multiple working states:

[0164] When the topology consistency check fails in step 2, the global state machine... Entering ERROR status;

[0165] When trust divergence occurs in step 5 Exceeding the conflict threshold At that time, the global state machine Switch to FEEDBACK state and revert the control process to step 2;

[0166] When the topological state verification of the two-layer graph is inconsistent, the global state machine Transition to CONSISTENCY status and trigger re-arbitration;

[0167] Once steps 1 through 5 have all been executed successfully, the global state machine... Transition to the QUERY_READY state to receive and respond to graph evolution path queries.

[0168] To further verify the technical feasibility of this invention, this embodiment applies the technical solution to a scientific research achievement knowledge graph in a university's scientific research management system. This embodiment aims to integrate multi-source data from library paper databases (structured), the State Intellectual Property Office patent database (semi-structured), and internet scientific research news reports (unstructured) to construct a dynamic knowledge graph containing entities such as researchers, papers, patents, and projects, as well as their relationships.

[0169] The execution entity in this embodiment is a data processing server configured with high-performance computing resources. The specific hardware environment includes: an Intel Xeon Gold 6248R processor (3.0GHz, 24 cores), 256GB of DDR4 ECC memory, and an NVIDIA TeslaT4 computing card (for accelerating vectorization and matrix operations); a storage system consisting of a 2TB NVMe SSD for high-speed caching and a 10TB SAS HDD for persistent storage; and a software environment based on the Ubuntu 20.04 LTS operating system, using Python 3.8 as the primary development language, the PyTorch framework for deep learning model inference, and RocksDB as the underlying storage engine.

[0170] This embodiment includes three types of data streams: a JSON-formatted academic paper data stream, an XML-formatted patent data stream, and a plain text-formatted research news stream. The academic paper stream includes fields for title, abstract, authors, and keywords; the patent data stream includes fields for patent_name, summary, inventors, and claims; and the research news stream is natural language text.

[0171] like Figure 1 and Figure 5 As shown, the detailed steps are as follows:

[0172] S1, Data preprocessing module 10 performs data preprocessing.

[0173] The three types of input data were cleaned and normalized, HTML tags and stop words were removed, the time format was standardized to yyyy-MM-dd, and the personnel name format was standardized. Then, the title, abstract, patent name, claims, or news text corresponding to each research achievement entity were constructed into a text sequence T, which was input into the pre-trained BERT model bert-base-chinese. The 768-dimensional vector corresponding to the CLS tag in the last hidden layer was extracted as the text semantic vector. For highly structured features such as subject classification codes, keyword categories, and patent classification numbers, One-Hot encoding is used to generate additional vectors. Finally, and After concatenation, the input is reduced to a fully connected layer to obtain a 512-dimensional initial feature vector. The initial feature vector set V is formed by the initial feature vectors of all entities in the current batch.

[0174] S2, Semantic Mapping and Verification Module 20 performs semantic mapping and verification.

[0175] Vectors from different sources (paper space, patent space, news space) are mapped to a unified semantic space, ensuring that the topological structure is not distorted.

[0176] First, semantic mapping is performed. A transformation matrix M is predefined, which is... The real matrix is ​​initialized as the identity matrix or a projection matrix obtained through PCA (Principal Component Analysis) pre-training. For any input vector... Calculate the mapped vector Its value is equal to the matrix product of the transformation matrix M and the input vector vi. At this point, This refers to the alignment vectors in the unified latent semantic space. To support subsequent dynamic adaptive updates, this embodiment maintains a dynamic k-nearest neighbor graph. For each newly mapped vector... In the semantic space, the k nearest neighbors (k is 10 in this embodiment) are quickly retrieved using the Faiss library, and edge connections are constructed. Furthermore, when subsequent steps provide feedback on weight adjustments, a low-rank update strategy is adopted: using the Sherman-Morrison formula, the update amount of the graph Laplacian inverse is quickly calculated based on the adjustment increment (which is equal to the outer product of vector u and the transpose of vector v), thus efficiently obtaining the updated transformation matrix and avoiding the huge overhead of full recalculation.

[0177] Next, a topology consistency check is performed. To prevent the mapping process from damaging the inherent structure of the data, the topology structure indicators of the technical data before and after mapping are checked. The specific processing logic is as follows:

[0178] (1) Constructing the topological complex: For the local neighborhood set before mapping and the mapped neighborhood set Construct Vietoris-Rips complexes separately. Set filtering parameters. When the distance between two points is less than ε, a side is connected. When three points are connected in pairs, a triangle (2-simple) is formed.

[0179] (2) Calculate the Betti number: Based on the constructed complex, calculate the 0th order Betti number. (Number of connected components, representing the number of clusters) and the first-order Betti number (The number of holes represents the number of ring-shaped structures.) (The rest of the text appears to be incomplete and requires further context.) , ) as a local topological feature vector .

[0180] (3) Calculate the consistency coefficient: Calculate the topological feature vectors before and after mapping. and Pearson correlation coefficient The calculation method is as follows and The covariance is divided by the product of the standard deviations of the two.

[0181] (4) Blocking decision: The system presets a consistency threshold. Equals 0.85 (this threshold is set based on a large number of experimental observations, when...) When the value is less than 0.85, the downstream entity alignment error rate increases significantly. Less than This indicates that the mapping led to severe semantic distortion (such as splitting what was originally a single cluster into multiple clusters, i.e.) (If an error occurs), the blocking mechanism is triggered, the current mapping result is discarded, the batch of data is placed in the exception queue, and an alarm is triggered.

[0182] S3, Entity Fusion and Weight Calculation Module 30 performs entity fusion and weight calculation.

[0183] First, calculate the geometric distance between the vectors. In the unified semantic space, for any two entity vectors... and Calculate its Euclidean distance The distance is equal to and The L2 norm of the difference vector. The smaller the distance, the more semantically similar the two entities are, and the greater the probability that they belong to the same entity. Subsequently, the module constructs multiple evidence sources based on vector distance, title similarity, author overlap, and keyword similarity, and fuses them using Dempster-Shafer evidence theory.

[0184] Among them, the support mvec(H) based on vector distance is composed of The calculated support mauth(H) based on author overlap is derived from the ratio of the number of elements in the intersection of the author sets to the number of elements in the union. The final fusion confidence is obtained through the DS combination rule. And the trust divergence K. If If the value is greater than 0.9, candidate entity a and candidate entity b are determined to be the same entity; otherwise, entity merging is not performed. The trust divergence K is used to represent the degree of conflict between different evidence sources and serves as the trigger for the closed-loop feedback in step S5.

[0185] For entity pairs that are identified as related, the semantic weights of the relationships are further calculated. Taking paper citation relationships as an example, citation markers in the text are matched using regular expressions, and 100 characters before and after the citation anchor are extracted as citation context. Jieba segmentation is performed on the citation context and the abstract of the cited paper to generate a word sequence. The cosine similarity of their TF-IDF vectors is calculated, and the normalized cosine similarity is used as the relationship weight. Finally, the system generates a topology change instruction package based on the fusion result and relationship weights. The instruction package includes a timestamp, operation type, entity identifier, relationship type, weight value, fusion confidence, and trust divergence.

[0186] S4, Index storage module 40 performs index storage.

[0187] First, the topology change instruction package generated in step S3 is written to a streaming buffer built on Kafka. This streaming buffer acts as a peak-shaving and valley-filling middleware, used to decouple the compute-intensive entity fusion layer from the I / O-intensive persistent storage layer.

[0188] Before writing the data to disk, the module performs topology compression according to a time window. In this embodiment, the batch window can be set to 500ms. If there are multiple update instructions for the same entity pair within a batch window, only the last one is retained; if there are inverse operations of adding an edge and then deleting an edge, they are canceled out in the instruction set. After compression, the SHA-256 hash value of the compressed instruction set is calculated as the fingerprint of the spatiotemporal differential index block, and index nodes are constructed.

[0189] S5, Closed-loop Feedback and Consistency Assurance Module 50 performs closed-loop feedback and consistency assurance.

[0190] The trust divergence K output in step S3 is monitored in real time. In this embodiment, the preset conflict threshold δ is 0.6. If K is not greater than 0.6, closed-loop feedback is not triggered; if K is greater than 0.6, it indicates that there is a significant conflict between different evidence sources, and the module generates a conflict gradient signal.

[0191] Subsequently, the module converts the conflict gradient signal into an incremental update ΔM of the transformation matrix M, and adjusts the transformation matrix in real time through the low-rank update mechanism in step S2. After the matrix is ​​updated, the system triggers a rollback mechanism, re-executing steps S2 and S3 using the updated transformation matrix, so that the current batch of data can be re-mapped and fused in the corrected semantic space.

[0192] S6, Coordination Output Module 60 performs coordination output.

[0193] Maintain a global state machine to manage the execution sequence of steps S1 to S5. When the topology verification in step S2 fails, the state machine blocks the current batch from entering entity fusion; when the closed-loop feedback is triggered in step S5, the state machine controls the system to revert to step S2; when the consistency verification fails, the state machine triggers a re-arbitration process.

[0194] When a query request is received At that time, the module first loads the most recent global map snapshot Gbase(tbase) before tstart, and then retrieves and overlays it in chronological order. The system dynamically reconstructs the graph structure within the target time window by analyzing all spatiotemporal difference index blocks within the range. As a result, the system can output the entity-relationship topology of specified researchers, papers, patents, projects, or technical topics within a given time window, demonstrating the evolution of research results.

[0195] The following specific example data illustrates the collaborative working process of each step in the above embodiments.

[0196] Suppose we receive the following three pieces of raw data from different data sources:

[0197] Data source A (paper repository JSON format): titled "Research on knowledge graph completion method based on deep learning", authors "Zhang San, Li Si", keywords "knowledge graph, deep learning, link prediction", DOI "10.1234 / kg.2025.001".

[0198] Data source B (patent database XML format): The patent title is "A knowledge graph completion device", the inventors are "Zhang San and Wang Wu", the abstract is "This invention relates to a knowledge graph link prediction device based on graph neural network", and the application number is "CN202512345678.9".

[0199] Data source C (scientific research news text format): The content is "Professor Zhang San's team at a certain university has made a significant breakthrough in the field of knowledge graphs. Their research results have been published in papers and have been granted patents."

[0200] The data preprocessing module 10 first cleans the three data sets, removing HTML tags and stop words. Then, the title and summary of data source A are concatenated and input into the BERT model to extract a 768-dimensional feature vector, which is then reduced to a 512-dimensional initial vector. Its partial component values ​​are 0.23, 0.45, etc., and the last component value is 0.67. Similarly, data sources B and C are processed to obtain... (Some component values ​​are 0.21, 0.48, etc., and the last component value is 0.69) and (Some component values ​​are 0.19, 0.42, etc., and the last component value is 0.71).

[0201] The semantic mapping and verification module 20 uses the transformation matrix M to... , , Mapping to a unified semantic space yields , , Then the Betti numbers before and after the mapping were calculated, and the results showed... and Pearson correlation coefficient If the value is greater than the threshold of 0.85, the verification passes. Since step S2 ensures the consistency of the semantic topology, the evidence fusion in step S3 can be computed in a high-quality vector space, avoiding misalignment caused by mapping distortion.

[0202] The entity fusion and weight calculation module 30 calculates the vector distances, where the distance between A and B is 0.15, the distance between A and C is 0.28, and the distance between B and C is 0.31. Based on author overlap (both A and B contain "Zhang San"), a BPA is constructed: the mvec(H) based on the vector distance is equal to the power function value with base e and exponent -0.5 multiplied by 0.15, resulting in 0.927; the mauth(H) based on author overlap is equal to 1 divided by 3, resulting in 0.333. Applying the Dempster combination rule, the fusion confidence Conf(A,B) is 0.951, which is greater than the threshold of 0.9, indicating that data sources A and B correspond to the same entity "Zhang San". The trust divergence K is 0.023, which is less than the threshold of 0.6, so no closed-loop feedback needs to be triggered. Since the topology verification in step S2 ensures the accuracy of the semantic vectors, the fusion result in this step has higher credibility, forming a technical synergy effect of "semantic fidelity - accurate fusion".

[0203] The index storage module 40 encapsulates the fusion result into a topology change instruction package, containing: a timestamp of 2025-06-15T10:30:00Z, an operation type of LINK, and an entity ID of [missing information]. The weight is 0.951. After the instruction packet is written to the Kafka buffer, topology compression is performed, generating a hash value of... Construct a spatiotemporal difference index block and write it to Rocks DB.

[0204] Since K equals 0.023, which is less than the threshold of 0.6, the closed-loop feedback and consistency guarantee module 50 does not trigger parameter adjustments. Simultaneously, comparing the MerkleTree root hash values ​​of the streaming differential layer and the batch baseline layer, they are consistent, eliminating the need for re-arbitration.

[0205] When the user queries At that time, the coordination output module 60 loads the global snapshot of 2024-12-31, overlays the differential index block of the first half of 2025, and reconstructs the association topology of "Zhang San" within this time window, showing its association with paper Entity A and patent Entity B.

[0206] The final output knowledge graph adds a new entity node "Zhang San" (type: researcher), with associated edges including "Zhang San-works-paper A" (weight 0.951) and "Zhang San-invention-patent B" (weight 0.951).

[0207] In addition, this embodiment handles possible abnormal situations as follows:

[0208] Empty data input processing: When the data preprocessing module 10 receives empty data or input with an abnormal format, the system records the exception log and discards the data, without affecting the normal processing of subsequent data. The module maintains an exception counter. When the number of consecutive exceptions exceeds a preset threshold (e.g., 100 times), an alarm is triggered to notify the operation and maintenance personnel.

[0209] Topology verification failure handling: When the Pearson correlation coefficient ρ calculated by the semantic mapping and verification module 20 is lower than the consistency threshold τ, the system blocks the processing flow of the current batch of data, places the batch of data in the exception queue for manual review or reprocessing, and generates an alarm message. The recovery strategy is: after the transformation matrix is ​​updated through the closed-loop feedback mechanism, the data is retrieved from the exception queue and the mapping operation is re-executed.

[0210] Evidence conflict handling: When the trust divergence K calculated by the entity fusion and weight calculation module 30 exceeds the conflict threshold... When this happens, the system does not discard the data directly. Instead, it triggers a closed-loop feedback mechanism to generate conflict gradient signals to update the transformation matrix parameters and re-executes the fusion calculation using the updated parameters. If K still exceeds the threshold after three consecutive retries, the entity pair is marked as "awaiting manual confirmation".

[0211] Storage write timeout handling: When the index storage module 40 times out while writing data to Rocks DB, the system uses an exponential backoff strategy to retry, with a maximum of 5 retries. If it still fails, the data is written to a local backup file and an alarm is triggered. After the storage system recovers, the data is read from the backup file to complete the write.

[0212] Consistency verification failure handling: When the closed-loop feedback and consistency assurance module 50 detects a mismatch between the Merkle Tree root hash value of the streaming differential layer and the batch baseline layer, the system uses the batch baseline layer data as the authoritative source, forcibly overwrites the error state of the streaming layer, and records the time point of inconsistency and the range of data involved for subsequent analysis.

[0213] Resource insufficiency handling: When the system detects that memory usage exceeds 90% or disk space is less than 10%, it automatically reduces the write frequency of the streaming buffer, suspends non-critical background tasks, and triggers a resource alert. Normal processing speed is restored after resources are released.

[0214] In summary, this embodiment effectively solves the three technical problems raised in the background art through the above technical solution:

[0215] To address the issue of low fidelity in existing semantic mapping technologies, this embodiment employs a topological structure index verification mechanism in step S2. This mechanism calculates the Betti number of the local neighborhood before and after mapping and compares it with the Pearson correlation coefficient. This forces the mapping process to maintain the inherent geometric structure of the original data. Once the correlation coefficient falls below a threshold, the process is blocked, fundamentally preventing the loss and distortion of semantic information.

[0216] To address the lack of adaptive optimization capabilities in existing technology fusion processes, this embodiment utilizes a closed-loop feedback mechanism in step S5, using trust divergence as a trigger signal. When evidence conflicts are severe, a conflict gradient is automatically generated and backpropagated to update the transformation matrix parameters, thereby achieving online adaptive optimization of the fusion system. This allows the fusion accuracy to continuously improve as the amount of data processed increases.

[0217] To address the problem that existing technologies cannot efficiently trace the evolution history of a map, this embodiment uses the spatiotemporal differential indexing mechanism in step S4 and the superposition reconstruction logic in step S6 to compress the map changes into differential index blocks with timestamps. Combined with the dynamic superposition capability of the coordination output module, it achieves rapid reconstruction of the map structure within any time window, meeting the need for refined tracing of the evolution path of scientific research results.

[0218] This embodiment provides an alternative implementation based on graph neural networks and graph databases. Compared with the above embodiments, this embodiment optimizes and replaces the feature extraction and storage engines, making it suitable for scenarios with higher requirements for graph structure mining depth.

[0219] In the data preprocessing stage, this embodiment uses the Graph SAGE algorithm instead of BERT for feature vectorization. Specifically, it constructs a heterogeneous information network by using research entities as graph nodes and citation relationships as edges. Through Graph SAGE's inductive learning capability, it aggregates the feature information of neighboring nodes to generate node embedding vectors. This approach can better capture the structured features of entities and is particularly suitable for academic networks with dense citation relationships.

[0220] In the index storage phase, this embodiment uses the native graph database Neo4j instead of RocksDB as the persistent storage engine. Although RocksDB has advantages in write performance, Neo4j has the advantage of native graph traversal when performing complex multi-hop path queries. To address the issue that Neo4j's write throughput is lower than that of LSM-Tree storage, this embodiment introduces a "write merging" mechanism in the index storage module, merging topology change commands within the same time window into a single Cypher batch execution statement, thereby balancing write performance and query capabilities.

[0221] Regarding parameter selection, this embodiment has performed sensitivity analysis and adjustments on key parameters:

[0222] (1) Consistency threshold In Example 1, the threshold was set to 0.85. In the graph neural network scenario of this example, since the vectors generated by Graph SAGE already contain structural information and the topology is more stable, this threshold can be appropriately reduced to 0.75 to reduce unnecessary blocking and improve system throughput. If this value is set too low (e.g., less than 0.5), semantic distortions cannot be effectively intercepted; if it is set too high (e.g., greater than 0.95), the system will become overly sensitive, frequently triggering retries and affecting processing efficiency.

[0223] (2) Conflict threshold δ: In the evidence theory fusion, this embodiment sets it to 0.7. This is because GraphSAGE has a high tolerance for noisy data and allows for a higher degree of evidence conflict.

[0224] By implementing the above alternatives, the present invention can not only utilize textual semantics, but also make full use of graph structure information, thus verifying the universality and effectiveness of the proposed "closed-loop feedback" and "spatiotemporal reconstruction" architecture under different specific algorithms.

[0225] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for constructing a domain knowledge graph of scientific research results through multi-source data fusion, characterized in that: Includes the following steps: Step 1: Obtain multi-source entity data streams in the field of scientific research results, and preprocess them to generate an initial feature vector set. ; Step 2, semantic mapping and verification, is to map the initial feature vector set to the unified latent semantic space based on the pre-constructed transformation matrix to obtain the aligned semantic vectors, and calculate the topological structure index of the local neighborhood before and after the mapping. Step 3, Entity Fusion and Weight Calculation: Calculate the geometric distance between the semantic vectors of the same entity from multiple sources, fuse the evidence theory to obtain the fusion result and the trust divergence K, and calculate the semantic weight of the relationship between entities to generate a topology change instruction package; the trust divergence K refers to the evidence conflict coefficient, which is used to represent the degree of conflict between different evidence sources; Step 4: Index storage. Write the topology change instruction package into the streaming buffer, perform topology compression to generate a spatiotemporal difference index block, and write it into the persistent storage engine. Step 5, closed-loop feedback and consistency assurance, detailed method is as follows: Step 5.1: Monitor the trust divergence K output in step S3 in real time, and preset the conflict threshold δ; if K≤δ, the current fusion result is considered credible and closed-loop feedback is not triggered; if K>δ, it indicates that there is a serious conflict between different evidence sources and the closed-loop feedback mechanism is triggered. Step 5.2: Determine the conflict direction vector based on external evidence. Generate conflict gradient signals ; Calculate the incremental update of the transformation matrix M based on the conflict gradient signal. Perform matrix update ; After the update is completed, the process is rolled back, and steps 2 and 3 are re-executed using the updated transformation matrix, so that the current batch of data can be re-mapped and entity fused in the corrected semantic space; at the same time, conflict gradient signals are... As a back pressure control signal, it is sent to the data preprocessing module or streaming buffer to reduce the data writing frequency; Step 5.3, Maintain flow cytometry differential stratification and batch baseline layer The batch baseline layer periodically merges the spatiotemporal difference index blocks to generate a map snapshot; at the same time section t, MerkleTree root hashes of the topological states of the two layers of map are constructed respectively; if the root hashes are inconsistent, it is determined that the streaming differential layer has a state drift, and the batch baseline layer map snapshot is used as the authoritative basis to trigger re-arbitration, and the stable map snapshot in the batch baseline layer is used as the authoritative source to cover or correct the erroneous state in the streaming differential layer. If the root hash is consistent, then the streaming differential layer is determined to be consistent with the batch baseline layer; Step 6: Coordinate output, maintain the execution sequence of steps 1 to 5 of the global state machine management, and respond to evolution path query requests. Dynamically reconstruct the topological structure of the entity association of scientific research results within the specified time window by superimposing spatiotemporal difference index blocks.

2. The method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion according to claim 1, characterized in that, The detailed method for step 1 is as follows: Step 1.1: The multi-source entity data stream includes academic paper data stream in JSON format, patent data stream in XML format, and scientific research news stream in plain text format; the academic paper stream includes fields for title, abstract, author, and keywords, and the patent data stream includes fields for patent name, abstract, inventor, and claims; the original multi-source entity data stream is processed by removing HTML tags, filtering stop words, unifying time format, and normalizing personnel name format. Step 1.2: For each research result entity, construct its text content as an input text sequence. , , Indicates the first in the input text sequence Each word element; Step 1.3: Input the text sequence T into the pre-trained BERT model and extract the vector corresponding to the CLS marker in the last hidden layer of the pre-trained BERT model as the text semantic vector. For strong features such as subject classification codes, keyword categories, and patent classification numbers, One-Hot encoding is used to generate additional vectors. ; Text semantic vectors With additional vectors After concatenation, the vectors are fed into a fully connected layer for dimensionality reduction, resulting in an initial feature vector of dimension d. ; ; in, This represents a fully connected dimensionality reduction function. This represents a vector concatenation operation; Step 1.4: The initial feature vector set is formed by the initial feature vectors corresponding to all scientific research achievement entities. ； in, This indicates the number of scientific research results entities in the current batch. This represents the m-th initial eigenvector in the initial eigenvector set.

3. The method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion according to claim 1, characterized in that, The detailed steps of semantic mapping and verification in step 2 are as follows: Step 2.1: Maintain a dynamic k-nearest neighbor graph with mapped semantic vectors as nodes and nearest neighbor relationships as edges. ; For each mapped semantic vector In the unified latent semantic space, its k nearest neighbors are retrieved and edge connections are established; when a weight adjustment is received, the semantic mapping relationship is updated using a low-rank update method based on the graph Laplacian matrix L, and the update formula is as follows: ； in, Represents the graph Laplace matrix. and This represents the low-rank update vector constructed from the weight adjustments. express Transpose of; Step 2.2: Construct Vietoris-Rips complexes for both the local neighborhood set before and after mapping; [The text then abruptly shifts to a different topic:] ...filtering parameters... Below, when the distance between two points is less than When three points are connected in pairs, a 2-simple is formed. Then, based on the Vietoris-Rips complex, the topological structure indices before mapping are calculated respectively. and mapped topology index ; ； ； in, This represents the number of connected components in the local neighborhood topological complex before mapping. This represents the number of ring structures in the local neighborhood topological complex before mapping. This represents the number of connected components in the local neighborhood topological complex after mapping. This represents the number of ring structures in the local neighborhood topological complex after mapping; Step 2.3: Calculate the Pearson correlation coefficient of the topological structure indices before and after mapping. : ； in, Describing covariance, Indicates standard deviation; like Less than the preset consistency threshold If the current mapping causes semantic topology distortion, the current batch of data will be blocked from flowing and the data will be stored in the exception queue; if Not less than Then the mapped semantic vector is input into step 3.

4. The method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion according to claim 1, characterized in that, The detailed process of step 3, entity fusion and weight calculation, is as follows: Step 3.1: In the unified latent semantic space, for any two candidate entities... and Corresponding semantic vector and Calculate the Euclidean distance between the two. The formula is as follows: ； Step 3.2: The Dempster-Shafer evidence theory is used to fuse multiple sources of evidence. For two sources of evidence... and After combination, the proposition support for: ； in, The coefficient of evidence conflict, also known as the confidence divergence, is calculated using the following formula: ； Then, based on the fusion confidence Conf(a, b) obtained from the fusion and the entity fusion threshold... The size relationship is used to determine whether candidate entity a and candidate entity b are the same entity; Step 3.3: First, use regular expressions to match citation markers in the text and extract the text before and after the citation anchor point of a preset length as the citation context; then, perform word segmentation and TF-IDF vectorization on the citation context and the citation text summary respectively, and calculate their cosine similarity. Next, for the entity pairs identified as related in step 3.2, cosine similarity is used as the basis for further analysis. Calculate the semantic weights of relationships between entities ; ； Finally, a topology change instruction package is generated based on the entity fusion results and relation weights. ; ； in, Represents a timestamp. Indicates the operation type. Indicates entity identifier, Corresponding to the fusion confidence level Conf(a, b), The type of relation is indicated by the operation type, which includes LINK, UNLINK, and UPDATE_WEIGHT. LINK indicates that a new association is added between two entity nodes, UNLINK indicates that an existing association is deleted between two entity nodes, and UPDATE_WEIGHT indicates that the semantic weight or confidence of the association is updated while preserving the entity association.

5. The method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion according to claim 1, characterized in that, The detailed method for storing the index in step 4 is as follows: Step 4.1: The topology change command packet Pcmd is written to a streaming buffer built on Kafka to decouple the entity fusion computing layer and the persistent storage layer. Let's define the set of topology change commands received within a single batch time window. for: ； Where L represents the number of topology change instructions received within the time window; Step 4.2: Before writing the data to the disk, perform topology compression on the topology change instruction set P: If there are multiple update instructions for the same entity pair within the same batch window, retain the one with the latest timestamp; if there are inverse operations of adding an edge first and then deleting an edge, cancel them out in the instruction set. Then the compressed instruction set Calculate the SHA-256 hash value as the fingerprint of the spatiotemporal differential index block. Constructing the spatiotemporal difference index block B: ； in, Indicates the time bucket identifier; Step 4.3: Write the spatiotemporal difference index block B into the RocksDB persistent storage engine based on LSM-Tree, storing the key. Designed as follows: ； in, Indicates the map partition identifier. Indicates a time bucket. Represents the entity identifier; Value is the compressed Protocol Buffers format topology instruction block.

6. The method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion according to claim 1, characterized in that, Step 5: Closed-loop feedback and consistency assurance process: Conflict gradient signal Incremental update volume The expression for matrix update is as follows: ， ， ， in, This represents the conflict gradient direction coefficient, used to determine whether the current feedback is an attractive or repulsive direction. This represents the learning rate or step size coefficient, used to control the updating of the transformation matrix based on the conflict gradient signal each time. The range.

7. The method for constructing a domain knowledge graph of scientific research results based on multi-source data fusion according to claim 1, characterized in that, The detailed steps for coordinating the output in step 6 are as follows: Global state machine It can be configured with multiple working states: When the topology consistency check fails in step 2, the global state machine... Entering ERROR status; When trust divergence occurs in step 5 Exceeding the conflict threshold At that time, the global state machine Switch to FEEDBACK state and revert the control process to step 2; When the topological state verification of the two-layer graph is inconsistent, the global state machine Transition to CONSISTENCY status and trigger re-arbitration; Once steps 1 through 5 have all been executed successfully, the global state machine... Transition to the QUERY_READY state to receive and respond to graph evolution path queries.

8. A knowledge graph construction system for scientific research results based on multi-source data fusion, characterized in that, It includes a data preprocessing module, a semantic mapping and verification module, an entity fusion and weight calculation module, an index storage module, a closed-loop feedback and consistency guarantee module, and a coordination output module; The data preprocessing module acquires multi-source scientific research entity data streams, performs standardization and feature vectorization processing, and generates an initial feature vector set. The semantic mapping and verification module maps the initial feature vector set to the unified latent semantic space based on the pre-constructed transformation matrix to obtain the aligned semantic vectors, and calculates the topological structure index of the local neighborhood before and after mapping. If the correlation coefficient of the topological structure index is lower than the preset consistency threshold, the processing flow is blocked. The entity fusion and weight calculation module calculates the vector geometric distance between the multi-source semantic vectors of the same entity, performs evidence theory fusion calculation to obtain the fusion result and trust divergence, and calculates the semantic weight of the relationship between entities to generate a topology change instruction package. The index storage module writes the topology change instruction package into the streaming buffer, performs topology compression processing to generate spatiotemporal differential index blocks, and writes them into the persistent storage engine. The closed-loop feedback and consistency assurance module monitors the trust divergence K output in step S3 in real time, with a preset conflict threshold δ. If K ≤ δ, the current fusion result is considered credible, and closed-loop feedback is not triggered. If K > δ, it indicates a serious conflict between different evidence sources, triggering the closed-loop feedback mechanism. The conflict direction vector is determined based on external evidence. Generate conflict gradient signals ; Calculate the incremental update of the transformation matrix M based on the conflict gradient signal. Perform matrix update ; After the update is completed, the process is rolled back, and steps 2 and 3 are re-executed using the updated transformation matrix, so that the current batch of data can be re-mapped and entity fused in the corrected semantic space; at the same time, conflict gradient signals are... As a back pressure control signal, it is sent to the data preprocessing module or streaming buffer to reduce the data writing frequency; Maintaining flow cytometry differential stratification and batch baseline layer The batch baseline layer periodically merges the spatiotemporal difference index blocks to generate a map snapshot; at the same time section t, MerkleTree root hashes of the topological states of the two layers of map are constructed respectively; if the root hashes are inconsistent, it is determined that the streaming differential layer has a state drift, and the batch baseline layer map snapshot is used as the authoritative basis to trigger re-arbitration, and the stable map snapshot in the batch baseline layer is used as the authoritative source to cover or correct the erroneous state in the streaming differential layer. If the root hash is consistent, then the streaming differential layer is determined to be consistent with the batch baseline layer; The coordination output module maintains the global state machine to manage the execution timing of each module, responds to evolution path query requests, and dynamically reconstructs the topological structure of the entity association of scientific research results within a specified time window by superimposing spatiotemporal difference index blocks.

9. The multi-source data fusion-based knowledge graph construction system for scientific research results according to claim 8, characterized in that, The index storage module includes a streaming buffer unit, a topology compression unit, a hash calculation unit, and a persistent storage unit. The streaming buffer unit is built using Kafka, and the persistent storage unit uses RocksDB based on the LSM-Tree architecture. An ordered key-value index is constructed according to the combination of graph partition identifier, time bucket identifier, and entity identifier. The closed-loop feedback and consistency assurance module includes a conflict monitoring unit, a matrix update unit, a state comparison unit, and an arbitration unit. The state comparison unit completes the topological state comparison between the streaming differential layer and the batch baseline layer based on the MerkleTree tree hash structure. When the states are inconsistent, the arbitration unit corrects the streaming differential layer data using the map snapshot of the batch baseline layer as the standard.

Images