A Deep Learning-Based Knowledge Extraction and Fusion Method

By extracting and fusing multi-source knowledge data through deep learning models, a heterogeneous knowledge evidence graph is constructed, which solves the problems of entity alignment and structure verification in multi-source knowledge extraction and fusion, and realizes the generation of knowledge graphs with accurate entity recognition, reliable knowledge fusion and strong dynamic update capability.

CN123115437APending Publication Date: 2026-11-17BEIJING DATA100 INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611550500.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies suffer from several drawbacks in the process of multi-source knowledge extraction and fusion. These include a single entity alignment basis and a lack of structural verification during the fusion process. This results in duplicate entities not being merged or different entities being merged incorrectly. Furthermore, the fused knowledge graph contains missing nodes, incorrect relation endpoints, and abnormal relation connections, making it difficult to update effectively.

Method used

We employ a deep learning-based knowledge extraction and fusion method. By using a confidence-enhanced PFN model to extract entities, attributes, and relationships, we generate entity semantic features and knowledge extraction confidence scores, construct a knowledge evidence heterogeneous graph, perform virtual merging and independent retention, identify synonymous entities, aliased entities, and homonymous entities, and generate a dynamic fused knowledge graph.

Benefits of technology

It improves the accuracy of multi-source entity alignment results, ensures the reliability and structural consistency of knowledge fusion, enables dynamic updates of the knowledge graph, reduces duplicate retention and erroneous merging, and enhances the integrity and update reliability of the knowledge graph.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115437A_ABST
    Figure CN123115437A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based knowledge extraction and fusion method, relating to the field of knowledge graph construction technology, comprising the following steps: collecting and preprocessing multi-source knowledge data to generate a multi-source knowledge dataset; extracting entities, attributes, and relationships using a confidence-enhanced Power Flow Network (PFN) model to generate entity semantic features, knowledge extraction confidence scores, and multi-source knowledge evidence units; screening entity combinations to be judged to generate candidate entity pairs; constructing a heterogeneous knowledge evidence graph and extracting local subgraphs of candidate entities; performing virtual merging and independent retention on candidate entity pairs to generate entity merging inference results; combining entity semantic features, attribute information, relational neighborhoods, and entity merging inference results to complete multi-source entity alignment; processing duplicate and conflicting knowledge, correcting multi-source knowledge evidence units based on structural anomalies, and generating a dynamically fused knowledge graph. This invention improves the accuracy and reliability of entity alignment, knowledge fusion, and graph structure correction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction technology, and in particular to a knowledge extraction and fusion method based on deep learning. Background Technology

[0002] In recent years, with the development of natural language processing, knowledge graphs, deep learning, and multi-source data fusion technologies, knowledge extraction and fusion for text data, tabular data, semi-structured data, database record data, and knowledge graph data has gradually become an important direction in intelligent information processing. Existing technologies typically first perform format conversion, field mapping, text segmentation, and terminology normalization on multi-source knowledge data. Then, deep learning models are used to identify entities, attributes, and relationships. Entity alignment is achieved through entity name matching, semantic similarity calculation, or attribute comparison. Subsequently, duplicate entities, attributes, and relationships are merged to form a unified knowledge graph. Some solutions also incorporate data source, extraction confidence, and time information to filter or update knowledge records to support knowledge retrieval, semantic association, and intelligent analysis.

[0003] However, existing technologies still suffer from problems in multi-source knowledge extraction and fusion processes, such as relying on a single entity alignment criterion and lacking structural verification during the fusion process. On the one hand, different data sources may present the same knowledge object with full names, abbreviations, aliases, and homonyms, and the extraction confidence levels of entities, attributes, and relationships vary. Matching based solely on entity names or semantic similarity can easily lead to duplicate entities not being merged or different entities being incorrectly merged. On the other hand, existing solutions typically fuse directly based on attribute content and relationship records, lacking the ability to infer the complete state of relationships, attribute conflict states, and neighborhood structure changes after the virtual merging of candidate entities. This makes it difficult to use source support and relationship connections in the knowledge graph to determine the retention status of conflicting knowledge. Furthermore, the resulting missing nodes, incorrect relationship endpoints, and abnormal relationship connections cannot be used to correct the original knowledge evidence, resulting in the retention of incorrect entity identifiers, attribute attributions, and relationship connections during the knowledge graph update process. Summary of the Invention

[0004] One objective of this invention is to propose a knowledge extraction and fusion method based on deep learning. This invention fully utilizes deep learning, heterogeneous knowledge evidence graphs, and entity merging inference methods to jointly extract entities, attributes, and relationships from multi-source knowledge, perform entity alignment, conflict resolution, and structural correction, and generate a dynamically fused knowledge graph. It has the advantages of accurate entity recognition, reliable knowledge fusion, high structural consistency, and strong dynamic update capability.

[0005] A knowledge extraction and fusion method based on deep learning according to an embodiment of the present invention includes the following steps: Collect multi-source knowledge data and preprocess it to generate a multi-source knowledge dataset; The multi-source knowledge dataset is input into the confidence-enhanced PFN model to extract entities, attributes and relationships, and generate entity semantic features, knowledge extraction confidence and multi-source knowledge evidence units. Based on entity names and semantic features, filter the combinations of entities to be judged, associate the attribute information, relational neighborhood and data source of each entity, and generate candidate entity pairs; Using multi-source knowledge evidence units as nodes, association edges are established based on entity co-occurrence, attribute correspondence, relational connection, and source support to construct a heterogeneous graph of knowledge evidence, and candidate entity local subgraphs are extracted; Virtual merging and independent retention are performed on candidate entity pairs respectively. The relationship integrity state, attribute conflict state and neighborhood structure change are compared to generate entity merging inference results. By combining entity semantic features, attribute information, relational neighborhood, knowledge extraction confidence and entity merging inference results, synonymous entities, aliased entities and homonymous entities are identified, and multi-source entity alignment results are generated. Based on the multi-source entity alignment results, entities, attributes, and relationships are fused. Duplicate and conflicting knowledge is processed according to the data source credibility status, knowledge extraction confidence, time consistency, and relationship support status. Multi-source knowledge evidence units are corrected based on structural anomalies to generate a dynamic fused knowledge graph.

[0006] Optionally, the multi-source knowledge data includes text data, tabular data, semi-structured data, database record data, and knowledge graph data. The preprocessing includes character encoding unification, format conversion, field mapping, text segmentation, abnormal data removal, missing data completion, terminology standardization, source tagging, and time alignment.

[0007] Optionally, the generation of the entity semantic features, knowledge extraction confidence, and multi-source knowledge evidence units includes: The multi-source knowledge dataset is input into the confidence-enhanced PFN model, which includes a semantic encoding layer, a partitioning filtering layer, a joint extraction layer, a confidence calibration layer, and a knowledge evidence generation layer. In the semantic encoding layer, the multi-source knowledge dataset is encoded according to the content units and word order to generate a contextual semantic representation; In the partitioning filtering layer, entity gating weights, relation gating weights, and shared gating weights are calculated based on the context semantic representation, and the context semantic representation is divided into entity semantic partitions, relation semantic partitions, and shared semantic partitions. In the joint extraction layer, entity semantic partitioning and shared semantic partitioning are fused to identify entity boundaries and entity types, and relation semantic partitioning and shared semantic partitioning are fused to match attribute correspondence states and relation connection states between entities, generating entity, attribute, relation and entity semantic features. In the confidence calibration layer, the confidence of knowledge extraction is determined by combining the entity boundary recognition results, entity type recognition results, attribute matching results, and relationship matching results. In the knowledge evidence generation layer, entities, attributes, relationships, entity semantic features, and knowledge extraction confidence are bound together according to data sources and content units to generate multi-source knowledge evidence units.

[0008] Optionally, the generation of the candidate entity pairs includes: The entities in the multi-source knowledge evidence unit are grouped according to entity type, and entities with different data sources are selected from the same entity type and combined in pairs. For each entity combination, the entity name is normalized and segmented. The character overlap status, word inclusion status and abbreviation correspondence status between entity names are calculated. The entity combination that satisfies any name correspondence status is determined as the first entity combination to be judged. Calculate the semantic similarity between the semantic features of entities in the same entity combination, and determine the entity combination with a semantic similarity not lower than the preset semantic screening threshold as the second entity combination to be judged. Merge the first and second entity combinations to be determined and remove duplicate combinations to generate the entity combination to be determined. Retrieve attribute information, associated entities, relationship type and relationship direction of two entities in each combination of entities to be judged from the multi-source knowledge evidence unit, and generate the relationship neighborhood of each of the two entities; Two entities to be judged are bound together with their entity names, semantic features, attribute information, relational neighborhoods, and data sources. Entity pairs with the same data source and content unit are eliminated to generate candidate entity pairs.

[0009] Optionally, the construction of the knowledge evidence heterogeneous graph and the candidate entity local subgraph includes: Configure node identifiers for each multi-source knowledge evidence unit and bind entities, attributes, relationships, entity semantic features, knowledge extraction confidence, data sources, and content units; Within the same content unit, establish entity co-occurrence edges for nodes corresponding to entities that appear together; Establish attribute-corresponding edges for nodes of entities with the same attribute type, identical attribute content, or corresponding relationships; Based on the starting entity, ending entity, relationship type, and relationship direction, establish relationship connection edges between corresponding nodes; For nodes with different data sources but supporting the same entity, attribute, or relationship, establish source support edges, and determine the association strength by combining knowledge extraction confidence and the number of source supports; Nodes and various associated edges are combined according to node identifiers to generate a heterogeneous graph of knowledge evidence; In the heterogeneous graph of knowledge evidence, locate the nodes corresponding to the candidate entity pairs, expand outward along two connection levels: entity co-occurrence edge, attribute correspondence edge, relation connection edge, and source support edge, retain the nodes and associated edges within the expansion range, and generate a local subgraph of candidate entities.

[0010] Optionally, the generation of the entity merging deduction result includes: In the local subgraph of the candidate entity, locate the two nodes corresponding to the candidate entity pair, and copy the local subgraph of the candidate entity to form a virtual merged deduction branch and an independently retained deduction branch; In the virtual merge simulation branch, the two nodes are replaced with virtual merge nodes, attribute information and data sources are merged, the original entity co-occurrence edges, attribute corresponding edges, relationship connection edges and source support edges are reconnected to the virtual merge nodes, duplicate association edges are merged, and a virtual merge state is generated. In the independent retention derivation branch, maintain two nodes and their associated edges to generate an independent retention state; Count the number of retained relational connections, disconnected relational connections, and self-joined relational connections under the two states respectively, and generate the complete relational state; Compare the standard attribute content corresponding to the same attribute type, mark the attribute records with inconsistent content and no attribute correspondence as attribute conflicts, and generate an attribute conflict status. By comparing the associated entities, relationship types, relationship directions, and number of associated edges before and after the two nodes are merged, the retention of the common neighborhood, the increase or decrease of associated edges, and the isolation of nodes are determined, and the changes in the neighborhood structure are generated. The results of entity merging are generated by considering the complete state of the combination relationship, the conflicting state of attributes, and the changes in the neighborhood structure.

[0011] Optionally, the generation of the multi-source entity alignment result includes: Retrieve the entity name, semantic features, attribute information, relational neighborhood, and knowledge extraction confidence of the candidate entity pair, and extract the relational integrity state, attribute conflict state, and neighborhood structure change from the entity merging inference results; By combining entity semantic features, attribute correspondence status and relational neighborhood features, it is determined whether two entities point to the same knowledge object. The determination result is verified by knowledge extraction confidence and entity merging inference results. Candidate entity pairs that point to the same knowledge object but have different entity names are identified as synonymous entities; Synonymous entities whose names have corresponding full names, abbreviations, or aliases are identified as alias entities; Candidate entity pairs with the same entity name but pointing to different knowledge objects are identified as homonymous entities; Establish entity identifier correspondence between synonymous entities and aliased entities, retain independent entity identifiers for entities with the same name but different meanings, bind entity type, data source and corresponding entity identifier, and generate multi-source entity alignment results.

[0012] Optionally, the generation of the dynamically fused knowledge graph includes: Based on the unified entity identifier, synonymous entities and alias entities are merged, while the independent entity identifiers of entities with the same name but different meanings are retained, and attributes and relationships are mapped to the corresponding entities; Organize the knowledge to be merged according to entity identifier, attribute type and relational connection, and identify duplicate and conflicting knowledge. The knowledge credibility status is determined based on the credibility status of the data source and the confidence level of knowledge extraction, and the knowledge retention status is determined by combining time consistency and relation support status. Merge the data sources and knowledge extraction confidence levels corresponding to duplicate knowledge, filter conflicting knowledge according to knowledge retention status, and record the conflict resolution results; The retained entities, attributes, and relationships are written into the knowledge graph to identify structural anomalies caused by missing nodes, incorrect relationship endpoints, and abnormal relationship connections. Based on the multi-source knowledge evidence units corresponding to the structural anomaly location, entity identification, attribute attribution and relational connection are corrected, and the nodes, associated edges and time information in the knowledge graph are updated to generate a dynamic fusion knowledge graph.

[0013] The beneficial effects of this invention are: This invention inputs multi-source knowledge data into a confidence-enhanced PFN model to jointly extract entities, attributes, and relationships, simultaneously generating entity semantic features, knowledge extraction confidence scores, and multi-source knowledge evidence units. This allows knowledge content from different data sources to be organized according to a unified structure. Entity semantic partitioning, relationship semantic partitioning, and shared semantic partitioning are used to retain entity identification information and relationship connection information, reducing the loss of associations caused by separate extraction of entities, attributes, and relationships. Knowledge extraction confidence scores record the credibility of each extraction result, providing comparable knowledge evidence for subsequent entity alignment and knowledge fusion.

[0014] This invention filters entity pairs based on entity names and semantic features, binding attribute information, relational neighborhoods, and data sources to candidate entity pairs to construct a heterogeneous knowledge evidence graph using multi-source knowledge evidence units. By performing virtual merging and independent retention on candidate entity pairs, and comparing relational integrity, attribute conflict, and neighborhood structure changes, the impact of merging two entities on the original knowledge structure can be determined before actual merging. This ensures that the identification of synonyms, aliases, and homonymous entities no longer relies solely on name similarity or single semantic similarity, reducing the duplication of the same knowledge object and the erroneous merging of different knowledge objects, thus improving the accuracy of multi-source entity alignment results.

[0015] This invention fuses entities, attributes, and relationships based on multi-source entity alignment results. It combines data source credibility, knowledge extraction confidence, temporal consistency, and relationship support status to handle duplicate and conflicting knowledge. This ensures that the data source and knowledge extraction confidence of duplicate records are uniformly preserved, while determining the corresponding knowledge retention status for conflicting records. After the knowledge graph is generated, it continues to identify missing nodes, incorrect relation endpoints, and abnormal relation connections. Based on structural anomalies, it reverse-locates and corrects multi-source knowledge evidence units, updating nodes, associated edges, and temporal information in the knowledge graph. This makes knowledge extraction, entity alignment, knowledge fusion, and structural correction a continuous processing procedure, improving the completeness, consistency, and update reliability of the dynamically fused knowledge graph. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a knowledge extraction and fusion method based on deep learning proposed in this invention; Figure 2 This is a schematic diagram illustrating the generation of multi-source knowledge evidence units based on a deep learning-based knowledge extraction and fusion method proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2 A knowledge extraction and fusion method based on deep learning includes the following steps: Collect multi-source knowledge data and preprocess it to generate a multi-source knowledge dataset; The multi-source knowledge dataset is input into the confidence-enhanced PFN model to extract entities, attributes and relationships, and generate entity semantic features, knowledge extraction confidence and multi-source knowledge evidence units. Based on entity names and semantic features, filter the combinations of entities to be judged, associate the attribute information, relational neighborhood and data source of each entity, and generate candidate entity pairs; Using multi-source knowledge evidence units as nodes, association edges are established based on entity co-occurrence, attribute correspondence, relational connection, and source support to construct a heterogeneous graph of knowledge evidence, and candidate entity local subgraphs are extracted; Virtual merging and independent retention are performed on candidate entity pairs respectively. The relationship integrity state, attribute conflict state and neighborhood structure change are compared to generate entity merging inference results. By combining entity semantic features, attribute information, relational neighborhood, knowledge extraction confidence and entity merging inference results, synonymous entities, aliased entities and homonymous entities are identified, and multi-source entity alignment results are generated. Based on the multi-source entity alignment results, entities, attributes, and relationships are fused. Duplicate and conflicting knowledge is processed according to the data source credibility status, knowledge extraction confidence, time consistency, and relationship support status. Multi-source knowledge evidence units are corrected based on structural anomalies to generate a dynamic fused knowledge graph.

[0019] In this embodiment, the multi-source knowledge data includes text data, tabular data, semi-structured data, database record data, and knowledge graph data; preprocessing includes character encoding unification, format conversion, field mapping, text segmentation, abnormal data removal, missing data completion, terminology standardization, source marking, and time alignment.

[0020] In this embodiment, the generation of entity semantic features, knowledge extraction confidence, and multi-source knowledge evidence units includes: The multi-source knowledge dataset is input into the confidence-enhanced PFN model, which includes a semantic encoding layer, a partitioning filtering layer, a joint extraction layer, a confidence calibration layer, and a knowledge evidence generation layer. The improvements to the confidence-enhanced PFN model are specifically reflected in the following aspects: First, data source identifiers and content unit identifiers are added to the semantic encoding layer of the original PFN model, allowing the generated contextual semantic representation to retain the source differences of multi-source knowledge data. Second, in the partitioning filtering layer, entity gating channels, relation gating channels, and shared gating channels are set up respectively. Entity gating weights, relation gating weights, and shared gating weights are calculated based on the contextual semantic representation, and the entity semantic partitions, relation semantic partitions, and shared semantic partitions are dynamically adjusted. Third, the original entity and relation joint extraction structure is extended to an entity, attribute, and relation joint extraction structure. Entity boundary recognition results, entity type recognition results, attribute matching results, and relation matching results are generated simultaneously in the joint extraction layer. Fourth, a confidence calibration layer is set up to jointly calibrate the entity boundary probability, entity type probability, attribute matching probability, and relation matching probability to generate knowledge extraction confidence. The training process of the confidence-enhanced PFN model is as follows: Samples containing entity, attribute, and relation annotations are selected from text data, tabular data, semi-structured data, database record data, and knowledge graph data. Entity boundaries, entity types, attribute correspondences, entity relationships, data sources, and trusted labels are added to generate the confidence-enhanced PFN model training dataset, which is then divided into training, validation, and test sets in an 8:1:1 ratio. Each sample is segmented according to content units, with the maximum input sequence length set to 256 words. Samples exceeding the maximum input sequence length are slide-segmented with an overlap length of 32 words. The semantic encoding layer is initialized using pre-trained language model parameters, while the partitioning filter layer, joint extraction layer, confidence calibration layer, and knowledge evidence generation layer are initialized with random parameters. During training, the AdamW optimizer is used, with the learning rate of the semantic encoding layer set to 0.00002. The learning rate for the remaining layers was set to 0.0001, the batch size to 32, the number of training epochs to 30, the weight decay coefficient to 0.01, the dropout rate to 0.1, the warm-up steps for the learning rate to 10% of the total training steps, and the gradient clipping threshold to 1.0. In each training epoch, entity boundary loss, entity type loss, attribute matching loss, relation matching loss, and confidence calibration loss were calculated sequentially, with weights for each loss function set to 0.25, 0.20, 0.20, 0.20, and 0.15 respectively. The confidence-enhanced PFN model parameters were updated based on the weighted total loss. After each training epoch, validation metrics for entity recognition, attribute matching, relation matching, and knowledge extraction confidence were calculated using the validation set. Training was stopped if no improvement was observed for five consecutive epochs. The model parameters corresponding to the highest validation metric in each epoch were saved, and the performance of the confidence-enhanced PFN model was verified using the test set. In the semantic encoding layer, the multi-source knowledge dataset is encoded according to the content units and word order to generate a contextual semantic representation; The semantic coding layer process specifically involves: grouping the multi-source knowledge dataset according to the data source identifier and content unit identifier; dividing each content unit into tokens based on their original character positions; and generating an input token sequence according to the order in which tokens appear in the content unit; adding a sequence start identifier and a sequence end identifier to the start and end positions of the input token sequence, respectively; setting the maximum length of a single input token sequence to 256 tokens; performing sliding segmentation on content units exceeding the maximum length according to an overlap length of 32 tokens; padding input token sequences shorter than the maximum length and generating a valid token mask; and obtaining the token embedding representation, position embedding representation, and data embedding representation corresponding to each token. Source embedding representation and content unit embedding representation are combined according to the corresponding word positions to generate an encoded input representation. Based on an effective word mask, the padding positions in the input word sequence are masked. The encoded input representation is then passed sequentially through multiple self-attention sublayers and feedforward network sublayers. Attention calculation, residual connections, layer normalization, and nonlinear mapping are performed in each layer to update the semantic association between each word and other words within the same content unit. The encoded vectors corresponding to each word are output according to the arrangement of the input word sequence. The encoded vectors within the same content unit are combined according to word order to generate a contextual semantic representation containing word context information, data source information, and content unit information. In the partitioning filtering layer, entity gating weights, relation gating weights, and shared gating weights are calculated based on the context semantic representation, and the context semantic representation is divided into entity semantic partitions, relation semantic partitions, and shared semantic partitions. The specific processing of the partitioning filtering layer is as follows: The encoding vector corresponding to each word is read according to the content unit and word position. This encoding vector is then input into the entity gating channel, relation gating channel, and shared gating channel, respectively. In the three gating channels, independent linear transformation matrices are used to map the encoding vectors, obtaining entity gating scores, relation gating scores, and shared gating scores. These three gating scores are then normalized at the same word position to generate entity gating weights, relation gating weights, and shared gating weights with values ​​ranging from 0 to 1 and a weight sum of 1. The entity gating weights are then multiplied element-wise with their corresponding encoding vectors, and dimensional mapping is performed using the entity feature projection matrix to generate entity partitioning vectors. Finally, the relation gating weights are multiplied with their corresponding encoding vectors... The elements are multiplied element-wise, and the dimensions are mapped through the relation feature projection matrix to generate relation partition vectors. The shared gate weights are multiplied element-wise with the corresponding encoding vectors, and the dimensions are mapped through the shared feature projection matrix to generate shared partition vectors. In this embodiment, the vector dimension of the context semantic representation is set to 768 dimensions, the dimension of the three partition vectors is set to 256 dimensions, and the dropout rate of the gate channel is set to 0.1. The entity partition vector, relation partition vector, and shared partition vector of each word are arranged according to the word order. The corresponding encoding vectors are converted into 256-dimensional vectors through linear projection. Residual connections and layer normalization are performed with the three partition vectors respectively to generate entity semantic partitions, relation semantic partitions, and shared semantic partitions. In the joint extraction layer, entity semantic partitioning and shared semantic partitioning are fused to identify entity boundaries and entity types, and relation semantic partitioning and shared semantic partitioning are fused to match attribute correspondence states and relation connection states between entities, generating entity, attribute, relation and entity semantic features. The joint extraction layer processes the following steps: First, it concatenates the entity partition vectors corresponding to the same lexical position with the shared partition vectors. Then, it converts the concatenated vectors into a 256-dimensional entity joint representation using an entity feature mapping matrix. Next, it inputs the entity joint representations into the entity boundary recognition channel and the entity type recognition channel, respectively, to calculate the entity start position, entity interior position, entity end position, and non-entity position corresponding to each lexical, thus determining the entity type for each entity. Then, it connects the entity start position, entity interior position, and entity end position according to the lexical order, combining consecutive lexicals with a recognition probability of not less than 0.60 into entities. Finally, it performs weighted aggregation on the entity joint representations within the entity coverage area to generate the entity semantic features corresponding to each entity. Finally, it concatenates the relation partition vectors corresponding to the same lexical position with the shared partition vectors. Next, a 256-dimensional relational joint representation is generated through the relational feature mapping matrix. Based on the start and end positions of the identified entities, the relational representations corresponding to each entity are extracted from the relational joint representation, and entity combinations are formed according to the order of entities within the same content unit. For each entity combination, the attribute correspondence state and relational connection state are calculated. The attribute correspondence state includes the correspondence between the attribute type and attribute content and the entity. The relational connection state includes the relation type, relational direction, and entities at both ends of the relation. The attribute correspondence state and relational connection state with a matching probability of not less than 0.55 are retained. Entities, attributes, and relations are associated according to content units and word positions. Entities are bound to their corresponding semantic features. The entities, attributes, relations, and semantic features are output for further processing by the confidence calibration layer. In the confidence calibration layer, the confidence of knowledge extraction is determined by combining the entity boundary recognition results, entity type recognition results, attribute matching results, and relationship matching results. The confidence calibration layer specifically processes the following: Based on the lexical range covered by the entity, the predicted probabilities of the entity's start position, internal position, and end position are averaged to generate entity boundary confidence. The entity boundary confidence is then combined with the predicted probabilities corresponding to the entity type using weights of 0.6 and 0.4 to generate entity extraction confidence. Similarly, the predicted probabilities corresponding to attribute matching results are combined with the entity extraction confidence of associated entities using weights of 0.6 and 0.4 to generate attribute extraction confidence. Finally, the predicted probabilities corresponding to relation matching results are combined with the entity extraction confidence of the entities at both ends of the relation using weights of 0.5, 0.25, and 0.25 to generate relation extraction confidence. The system calls the confidence calibration parameters saved after the confidence-enhanced PFN model training is completed, and performs temperature scaling and normalization on the entity extraction confidence, attribute extraction confidence, and relation extraction confidence. The temperature parameters for entities, attributes, and relations are 1.05, 1.10, and 1.15, respectively. The calibrated entity extraction confidence, attribute extraction confidence, and relation extraction confidence are then bound to the corresponding entities, attributes, and relations. Extraction results with a knowledge extraction confidence below 0.50 are removed, extraction results with a knowledge extraction confidence between 0.50 and 0.70 are marked as results to be verified, and extraction results with a knowledge extraction confidence of not less than 0.70 are marked as valid results. In the knowledge evidence generation layer, entities, attributes, relationships, entity semantic features, and knowledge extraction confidence are bound together according to data sources and content units to generate multi-source knowledge evidence units.

[0021] In this embodiment, the generation of candidate entity pairs includes: The entities in the multi-source knowledge evidence unit are grouped according to entity type, and entities with different data sources are selected from the same entity type and combined in pairs. For each entity combination, the entity name is normalized and segmented. The character overlap status, word inclusion status and abbreviation correspondence status between entity names are calculated. The entity combination that satisfies any name correspondence status is determined as the first entity combination to be judged. The determination of the first entity combination to be judged is specifically as follows: For each entity combination, the two entity names are subjected to character normalization processing. Full-width characters are converted to half-width characters, English characters are uniformly converted to lowercase, spaces, punctuation marks, parentheses, and connecting characters not involved in entity recognition are deleted, while retaining Chinese characters, letters, and numbers, generating two normalized entity names. The normalized entity names are then segmented according to Chinese word boundaries, English word boundaries, and consecutive position of numbers. The segmentation results are combined into two entity name word sequences according to their original order. Character sets are established for the characters in the two normalized entity names, the number of identical characters is counted, and the ratio of the number of identical characters to the number of characters in the shorter of the two normalized entity names is calculated to obtain the character overlap ratio. When the character overlap ratio is not less than 0.80, the corresponding entity combination is marked as having character overlap. The two entity name word sequences are compared sequentially. When the shorter entity name... When all lexical elements in the lexical element sequence are consecutively contained in the longer entity name lexical element sequence in their original order, or when the longer entity name contains the complete shorter entity name, the corresponding entity combination is marked as a lexical element inclusion state. The first character of each lexical element is extracted from the longer entity name lexical element sequence, and the last lexical elements representing organizational form, geographical level, and business category are deleted. The extracted results are concatenated in sequence into an abbreviation character sequence. The abbreviation character sequence is compared with the shorter entity name for consistency. At the same time, the full name, abbreviation, and alias corresponding records generated in the preprocessing of the multi-source knowledge dataset are retrieved. When the abbreviation character sequence is consistent with the shorter entity name, or when the two entity names belong to the same full name and abbreviation mapping relationship in the corresponding records, the corresponding entity combination is marked as an abbreviation corresponding state. Entity combinations with any of the following states are retained and bound to the corresponding name corresponding state to generate the first entity combination to be judged. Calculate the semantic similarity between the semantic features of entities in the same entity combination, and determine the entity combination with a semantic similarity not lower than the preset semantic screening threshold as the second entity combination to be judged. The determination of the second entity combination to be judged is as follows: The semantic features of the two entities in the same entity combination are checked for dimensionality, and the two entity semantic features are unified into a 256-dimensional vector. The feature values ​​of each dimension are normalized to generate a first entity semantic vector and a second entity semantic vector. The directional consistency between the first entity semantic vector and the second entity semantic vector is calculated according to the corresponding dimension to generate semantic similarity, which is limited to a value range of 0 to 1. Semantic filtering thresholds are set according to entity type, where the semantic filtering thresholds for person entities, organization entities, location entities, and product entities are set to 0.82, 0.80, 0.84, and 0.78, respectively. The entity type corresponding to the entity combination is read, and the semantic filtering threshold corresponding to the entity type is retrieved. The semantic similarity is compared with the semantic filtering threshold, and entity combinations with semantic similarity lower than the semantic filtering threshold are eliminated. The two entities, entity semantic features, semantic similarity, and entity type in the retained entity combination are bound together to generate the second entity combination to be judged. Merge the first and second entity combinations to be determined and remove duplicate combinations to generate the entity combination to be determined. Retrieve attribute information, associated entities, relationship type and relationship direction of two entities in each combination of entities to be judged from the multi-source knowledge evidence unit, and generate the relationship neighborhood of each of the two entities; The generation of the relation neighborhood is specifically as follows: Each entity to be judged is retrieved from the multi-source knowledge evidence units according to its entity identifier; the corresponding attribute type, attribute content, and attribute extraction confidence are extracted; duplicate attribute records are merged to generate an attribute set for the entity to be judged; relation records with the entity to be judged as the starting or ending entity of the relation are searched; the associated entity, relation type, relation direction, and relation extraction confidence of the other end of the relation are extracted; relation records with a relation extraction confidence of less than 0.50 are removed; duplicate records with the same associated entity and consistent relation type and relation direction are merged; the entity to be judged, the attribute set of the entity to be judged, the associated entity, the relation type, the relation direction, and the corresponding knowledge extraction confidence are bound together to generate the relation neighborhood. Two entities to be judged are bound together with their entity names, semantic features, attribute information, relational neighborhoods, and data sources. Entity pairs with the same data source and content unit are eliminated to generate candidate entity pairs.

[0022] In this embodiment, the construction of the knowledge evidence heterogeneous graph and the candidate entity local subgraph includes: Configure node identifiers for each multi-source knowledge evidence unit and bind entities, attributes, relationships, entity semantic features, knowledge extraction confidence, data sources, and content units; Within the same content unit, establish entity co-occurrence edges for nodes corresponding to entities that appear together; Establish attribute-corresponding edges for nodes of entities with the same attribute type, identical attribute content, or corresponding relationships; The establishment of attribute-corresponding edges is specifically as follows: Extract the attribute type and content corresponding to each entity, and group attribute records according to attribute type; standardize the format of attribute content under the same attribute type, unifying the unit of measurement and numerical precision for numeric attributes, the date format for time attributes, and performing character and terminology normalization on text attributes to generate standard attribute content; compare the standard attribute content corresponding to different entities item by item, and mark attribute content as consistent when the standard attribute content is completely identical; mark attribute records where the numerical values ​​are consistent after unit conversion and numerical precision unification, the effective time ranges of time attributes overlap, and text attributes are mapped to the same standardized content in the full name, abbreviation, and alias records as having a corresponding relationship in attribute content; establish attribute-corresponding edges between corresponding nodes of entities with the same attribute type and consistent attribute content or a corresponding relationship, and bind the attribute type, attribute correspondence status, attribute extraction confidence, and corresponding data source to the attribute-corresponding edges. Based on the starting entity, ending entity, relationship type, and relationship direction, establish relationship connection edges between corresponding nodes; The establishment of relational connection edges is specifically as follows: extract the starting entity, ending entity, relation type, relation direction, and relation extraction confidence level corresponding to each relation; locate the corresponding nodes of the starting entity and ending entity in the heterogeneous knowledge evidence graph according to the entity identifier; establish directed relational connection edges with the starting node of the starting entity and the ending node of the ending entity as the endpoint, and bind the relation type, relation direction, relation extraction confidence level, data source, and content unit to the relational connection edges; for relations without directionality, establish undirected relational connection edges between the corresponding nodes of the two entities; merge duplicate relation records with the same starting entity, ending entity, relation type, and relation direction, retain the data source and relation extraction confidence level corresponding to each relation record, and complete the establishment of relational connection edges; For nodes with different data sources but supporting the same entity, attribute, or relationship, establish source support edges, and determine the association strength by combining knowledge extraction confidence and the number of source supports; The establishment of source support edges and the determination of association strength are as follows: The multi-source knowledge evidence units are grouped according to data source, and nodes with different data sources are cross-compared; nodes whose entity names are consistent after character normalization and whose entity types are the same are marked as supporting the same entity; nodes whose attribute types and standard attribute content are consistent are marked as supporting the same attribute; nodes whose starting entity, ending entity, relation type, and relation direction are consistent are marked as supporting the same relation; source support edges are established between nodes in each marked group, and the supporting object type, data source, and knowledge extraction confidence are bound to the source support edges; the number of different data sources corresponding to the same entity, attribute, or relation is counted, the average knowledge extraction confidence of the corresponding nodes is calculated, and the average knowledge extraction confidence is combined with the data source number normalization result with a weight of 30% and weighted by 70% to generate an association strength ranging from 0 to 1, and the association strength is written into the corresponding source support edge; Nodes and various associated edges are combined according to node identifiers to generate a heterogeneous graph of knowledge evidence; In the heterogeneous graph of knowledge evidence, locate the nodes corresponding to the candidate entity pairs, expand outward along two connection levels: entity co-occurrence edge, attribute correspondence edge, relation connection edge, and source support edge, retain the nodes and associated edges within the expansion range, and generate a local subgraph of candidate entities.

[0023] In this embodiment, the generation of entity merging deduction results includes: In the candidate entity local subgraph, locate the two nodes corresponding to the candidate entity pair, and copy the candidate entity local subgraph to form a virtual merge inference branch and an independent retention inference branch. Specifically, in the candidate entity local subgraph, locate the two candidate entity nodes according to the entity identifiers bound to the candidate entity pair, verify the entity type, data source, and node identifiers corresponding to the two candidate entity nodes, and record the entity co-occurrence edges, attribute correspondence edges, relationship connection edges, and source support edges connecting the two candidate entity nodes; copy all nodes, associated edges, and their binding information in the candidate entity local subgraph, configure different inference branch identifiers for the copied results, and make the two inference branches independent of the original candidate entity local subgraph; mark one copied result as a virtual merge inference branch and mark the two candidate entity nodes as nodes to be merged; mark the other copied result as an independent retention inference branch, keeping the two candidate entity nodes and their associated edges unchanged; record the initial number of nodes, number of associated edges, attribute information, relationship connection status, and relationship neighborhood of the two inference branches respectively, and record the independent retention inference branch as the comparison benchmark for the virtual merge status; In the virtual merge simulation branch, the two nodes are replaced with virtual merge nodes, attribute information and data sources are merged, the original entity co-occurrence edges, attribute corresponding edges, relationship connection edges and source support edges are reconnected to the virtual merge nodes, duplicate association edges are merged, and a virtual merge state is generated. The generation of the virtual merge state is as follows: Retrieve the node identifiers, entity types, attribute information, data sources, and associated edges corresponding to the two nodes to be merged; delete the nodes to be merged and create a virtual merge node; record the original node identifier of the nodes to be merged as the source node identifier of the virtual merge node; write the common entity type of the two nodes to be merged into the virtual merge node; merge the attribute information of the nodes to be merged according to attribute type; deduplicate attribute records with the same attribute type and standard attribute content and retain all data sources; retain attribute records with the same attribute type but different standard attribute content separately; merge the data sources and knowledge extraction confidence scores corresponding to the nodes to be merged, and delete duplicate data sources; replace the corresponding endpoints of the entity co-occurrence edges, attribute corresponding edges, relationship connection edges, and source support edges originally connected to the nodes to be merged with the virtual merge node, keeping the type, direction, and binding information of the associated edges unchanged; merge associated edges with the same starting node, ending node, associated edge type, and relationship direction, and summarize the corresponding data sources and knowledge extraction confidence scores; convert the associated edges between the nodes to be merged into self-connection records of the virtual merge node and mark them separately to generate the virtual merge state. In the independent retention derivation branch, maintain two nodes and their associated edges to generate an independent retention state; Count the number of retained relational connections, disconnected relational connections, and self-joined relational connections under the two states respectively, and generate the complete relational state; The generation of a complete relational state is as follows: Using the relational connections in the independently retained state as a comparison benchmark, each relational connection is numbered according to its starting entity, ending entity, relation type, and relation direction. In the virtual merged state, each corresponding relational connection is searched item by item. If both ends of the relation are still connected to valid nodes and the relation type and direction remain consistent, it is marked as a retained relational connection. If the relation endpoints cannot be located after node replacement, the relation direction changes, or the relational connection is deleted, it is marked as a disconnected relational connection. If both ends of the relation are replaced with virtual merged nodes, it is marked as a self-connected relational connection. The number of retained relational connections, disconnected relational connections, and self-connected relational connections are counted separately, and the proportion of retained relational connections to the total number of relational connections in the independently retained state is calculated. The number of retained relational connections and their proportion, the number of disconnected relational connections, and the number of self-connected relational connections are combined to generate the complete relational state. Compare the standard attribute content corresponding to the same attribute type, mark the attribute records with inconsistent content and no attribute correspondence as attribute conflicts, and generate an attribute conflict status. The generation of attribute conflict states is as follows: The attribute types, standard attribute contents, attribute extraction confidence levels, and data sources corresponding to two candidate entity nodes are extracted and grouped according to attribute types. Attribute records with the same attribute type in the candidate entity nodes are compared item by item. Records with identical standard attribute contents or corresponding attribute relationships are marked as non-conflicting attribute records, while those with inconsistent standard attribute contents and no corresponding attribute relationships are marked as conflicting attribute records. Attribute records existing only in one candidate entity node are marked as unilateral attribute records and are not included in the attribute conflict count. The number of conflicting attribute records and the number of identical attribute types participating in the comparison are counted to determine the attribute conflict ratio. According to the attribute types corresponding to the conflicting attribute records, the conflicting standard attribute contents, attribute extraction confidence levels, and data sources are recorded. The number of conflicting attribute records, the attribute conflict ratio, and each conflicting attribute record are combined to generate the attribute conflict state. By comparing the associated entities, relationship types, relationship directions, and number of associated edges before and after the two nodes are merged, the retention of the common neighborhood, the increase or decrease of associated edges, and the isolation of nodes are determined, and the changes in the neighborhood structure are generated. The generation of neighborhood structure changes is as follows: using the associated entities, relationship types, relationship directions, and number of associated edges connected to two candidate entity nodes in the independent retention state as the neighborhood benchmark before merging, extracting the same information corresponding to the virtual merge nodes in the virtual merge state; marking associated entities with consistent entity identifiers and consistent relationship types and directions before and after merging as common neighborhood retentions; counting the newly added associated entities and associated edges after merging, the associated edges reduced due to repeated merging, and the associated edges lost due to endpoint replacement, generating the associated edge increase / decrease state; checking whether other nodes in the local subgraph of the candidate entity are still connected to at least one valid associated edge after the virtual merge, marking nodes whose number of associated edges changes from non-zero to zero as isolated nodes; combining the number of common neighborhood retentions and their proportion to the total number of associated entities before merging, the number of newly added associated edges, the number of reduced associated edges, the number of lost associated edges, and the number of isolated nodes to generate neighborhood structure changes; The results of entity merging are generated by considering the complete state of the combination relationship, the conflicting state of attributes, and the changes in the neighborhood structure.

[0024] In this embodiment, the generation of multi-source entity alignment results includes: Retrieve the entity name, semantic features, attribute information, relational neighborhood, and knowledge extraction confidence of the candidate entity pair, and extract the relational integrity state, attribute conflict state, and neighborhood structure change from the entity merging inference results; Combining entity semantic features, attribute correspondence states, and relational neighborhood features, this method determines whether two entities point to the same knowledge object. The determination result is verified using knowledge extraction confidence and entity merging inference results. Specifically, the semantic features of the two entities in a candidate entity pair are normalized, and the semantic similarity between the two entity semantic features is calculated. Attribute records with consistent attribute types, consistent attribute content, and corresponding attribute content are extracted from the attribute correspondence states. The proportion of corresponding attributes to the total number of attributes participating in the comparison is counted to determine attribute consistency. The associated entities, relation types, and relation directions connecting the two entities are extracted from the relational neighborhood features. Relationship records with consistent associated entities pointing to the same relation type and relation direction are matched to determine relational neighborhood consistency. Semantic similarity, attribute consistency, and relational neighborhood consistency are combined with weights of 0.40, 0.30, and 0.30 respectively to generate an initial determination score for the same knowledge object. Candidate entity pairs with an initial judgment score of not less than 0.70 are marked as pointing to the same knowledge object; the knowledge extraction confidence scores of the two entities, their attributes, and relationships are read, and the comprehensive knowledge extraction confidence score of the candidate entity pairs is calculated; the relationship integrity status, attribute conflict status, and neighborhood structure changes are extracted from the entity merging deduction results. When the relationship connection retention ratio is not less than 0.80, the attribute conflict ratio is not higher than 0.10, and no isolated nodes are formed, the entity merging deduction results are marked as supporting merging; when the relationship connection retention ratio is lower than 0.60, the attribute conflict ratio is higher than 0.30, or isolated nodes are formed, the entity merging deduction results are marked as not supporting merging; when the comprehensive knowledge extraction confidence score is not less than 0.70 and the entity merging deduction results support merging, the judgment of pointing to the same knowledge object is retained; when the entity merging deduction results do not support merging, the corresponding candidate entity pairs are adjusted to point to different knowledge objects, and the remaining candidate entity pairs are marked as pending review, generating the verified knowledge object judgment results; Candidate entity pairs that point to the same knowledge object but have different entity names are identified as synonymous entities; Synonymous entities whose names have corresponding full names, abbreviations, or aliases are identified as alias entities; Candidate entity pairs with the same entity name but pointing to different knowledge objects are identified as homonymous entities; Establish entity identifier correspondence between synonymous entities and aliased entities, retain independent entity identifiers for entities with the same name but different meanings, bind entity type, data source and corresponding entity identifier, and generate multi-source entity alignment results.

[0025] In this embodiment, the generation of dynamically fused knowledge graphs includes: Based on the unified entity identifier, synonymous entities and alias entities are merged, while the independent entity identifiers of entities with the same name but different meanings are retained, and attributes and relationships are mapped to the corresponding entities; Organize the knowledge to be merged according to entity identifier, attribute type and relational connection, and identify duplicate and conflicting knowledge. The knowledge credibility status is determined based on the credibility status of the data source and the confidence level of knowledge extraction, and the knowledge retention status is determined by combining time consistency and relation support status. The determination of knowledge credibility and knowledge retention status is as follows: The total number of multi-source knowledge evidence units corresponding to each data source, and the number of multi-source knowledge evidence units connected to at least one source support edge, are counted. The number of multi-source knowledge evidence units connected to source support edges is divided by the total number of multi-source knowledge evidence units corresponding to the data source to generate the data source credibility status of that data source. The data source credibility status and knowledge extraction confidence level corresponding to each entity, attribute, and relation to be merged are read. The data source credibility status and knowledge extraction confidence level are normalized to a numerical range of 0 to 1, and weighted summed according to weights of 0.45 and 0.55 respectively to generate the knowledge credibility status. When the same knowledge to be merged corresponds to multiple data sources, the knowledge credibility status corresponding to each data source is calculated separately, and then the average value of each knowledge credibility status is calculated as the knowledge credibility status of the knowledge to be merged. Attributes and relations are grouped according to a unified entity identifier. The occurrence time and effective time range corresponding to the knowledge in the same group are compared. If the content is the same and the effective time range overlaps, or the order of content changes is consistent with the order of occurrence time, it is determined to be time-consistent. If the content is different and the effective time range overlaps, it is determined to be time-unconsistent. For inconsistent time settings, source support edges corresponding to the knowledge to be fused are extracted from the heterogeneous knowledge evidence graph, and the average association strength of each source support edge is calculated. For relation records, the entities at both ends of the relation, the relation type, and the relation direction are checked. If all are consistent, the relation connection consistency value is set to 1; otherwise, it is set to 0. For entity records and attribute records, the relation connection consistency value is set to 1. The average association strength and the relation connection consistency value are combined with weights of 0.70 and 0.30, respectively, to generate the relation support state. If there are no source support edges, the average association strength is set to 0. The knowledge credibility state, time consistency, and relation support state are judged. If the knowledge credibility state is not lower than 0.70, the time is consistent, and the relation support state is not lower than 0.60, the knowledge to be fused is marked as retained knowledge. If the time is inconsistent, the knowledge credibility states corresponding to conflicting knowledge are compared. The knowledge with the higher knowledge credibility state is retained. If the difference in knowledge credibility state is less than 0.05, the knowledge with the higher relation support state is retained. If the knowledge credibility state is lower than 0.50 and the relation support state is lower than 0.40, the knowledge to be fused is marked as not retained knowledge, and a knowledge retention state is generated. The process involves merging the data sources and knowledge extraction confidence levels corresponding to duplicate knowledge, filtering conflicting knowledge based on its retention status, and recording the conflict resolution results. Specifically, attribute records are grouped according to unified entity identifier, attribute type, and standard attribute content; relation records are grouped according to the entities at both ends of the relation, relation type, and relation direction; records with identical content are identified as duplicate knowledge; for duplicate knowledge, one record is retained, the corresponding data sources are merged, the average of the knowledge extraction confidence levels is calculated, and a merged knowledge extraction confidence level is generated; records with different attribute content or inconsistent relation connections are identified as conflicting knowledge, and corresponding records are retained and non-retained records are deleted based on their knowledge retention status; when multiple retained records exist, they are distinguished according to the occurrence time and effective time range, and when the effective time ranges overlap, the record with the higher knowledge confidence level is retained; the data source, knowledge extraction confidence level, retained record, and deleted record corresponding to conflicting knowledge are bound together to generate the conflict resolution results. The retained entities, attributes, and relationships are written into the knowledge graph, and structural anomalies caused by missing nodes, incorrect relationship endpoints, and abnormal relationship connections are identified. Specifically, entity nodes are created for retained entities based on unified entity identifiers, and the corresponding attribute type, attribute content, data source, and knowledge extraction confidence are written into the entity nodes. The corresponding entity nodes are located according to the unified entity identifiers of the entities at both ends of the relationship, and the relationship type, relationship direction, and time information are written into the relationship connection between the two entity nodes to form the knowledge graph. The unified entity identifiers referenced by attributes and relationships are checked item by item. If the referenced unified entity identifier does not exist in the knowledge graph, it is marked as a missing node. If the starting entity or ending entity of the relationship is inconsistent with the corresponding entity in the multi-source entity alignment result, it is marked as an incorrect relationship endpoint. If the relationship type, relationship direction is inconsistent with the corresponding relationship record, or there are contradictory relationship connections between the same entity nodes, it is marked as an abnormal relationship connection. The anomaly type, corresponding entity identifier, attribute record or relationship record, and data source are bound together to generate structural anomalies. Based on the multi-source knowledge evidence units corresponding to the structural anomaly location, entity identification, attribute attribution and relational connection are corrected, and the nodes, associated edges and time information in the knowledge graph are updated to generate a dynamic fusion knowledge graph.

[0026] The generation of a dynamically fused knowledge graph involves: reading entity identifiers, attribute records, relation records, and data sources for structurally abnormal bindings; locating the corresponding entities, attributes, and relations in multi-source knowledge evidence units according to data sources and content units; for entities with missing nodes, retrieving the unified entity identifier from the multi-source entity alignment results, replacing the original entity identifier with the unified entity identifier, and supplementing the corresponding entity node; for attributes with incorrect attribution, rewriting the attribute into the corresponding entity node based on the correspondence between attributes and entities in the multi-source knowledge evidence unit; for relations with incorrect endpoints, correcting the unified entity identifiers at both ends of the relation based on the starting entity, ending entity, and multi-source entity alignment results in the relation record; for relations with abnormal connections, comparing relation type, relation direction, and time information, deleting relation connections inconsistent with the multi-source knowledge evidence unit, and restoring the corresponding relation connections based on the knowledge retention status; writing the corrected entity identifiers, attribute attributions, and relation connections into the knowledge graph, updating the corresponding nodes, associated edges, and time information, and generating a dynamically fused knowledge graph.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-source policy knowledge integration scenario in a government knowledge service center. This center needs to uniformly organize policy documents published on government websites, item forms maintained by business departments, service guide pages, internal database records, and existing knowledge graphs for policy inquiries, item consultations, and document retrieval. Different data sources have full names, abbreviations, aliases, and historical names for the same institution and item; some items with the same name belong to different regions or different business departments. The original processing method mainly relied on matching entity names and fixed dictionaries, which easily resulted in "Provincial Market Supervision Administration," "Provincial Market Supervision Bureau," and "Provincial Market Supervision Bureau" being retained as multiple entities. It also easily led to the incorrect merging of "Enterprise Registration Window" with the same name in different regions, causing duplicate institutions, item confusion, incorrect attribute attribution, and abnormal relationship connections.

[0028] This embodiment collects approximately 180,000 policy texts, 36,000 table records, 21,000 service guide field combinations, 74,000 database records, and approximately 420,000 existing knowledge graph triples. Policy texts include policy name, issuing department, applicable subjects, processing conditions, and application materials. Table records include item code, material name, and processing level. Database records include institution code, item code, and business status. Implementers perform preprocessing on all types of data to generate a multi-source knowledge dataset, retaining the data source, original name, and time information corresponding to each content unit.

[0029] In this scenario, a multi-source knowledge dataset is input into the confidence-enhanced PFN model. Through a semantic encoding layer, partitioning filtering layer, joint extraction layer, confidence calibration layer, and knowledge evidence generation layer, institutional entities, policy entities, matter entities, and material entities, along with their attributes and relationships, are extracted, generating entity semantic features, knowledge extraction confidence, and multi-source knowledge evidence units. The system filters entity combinations to be judged based on entity names and semantic features, retrieves the corresponding attribute information, associated entities, relationship types, and relationship directions for each entity, forms a relationship neighborhood, and generates candidate entity pairs. Using multi-source knowledge evidence units as nodes, it establishes entity co-occurrence edges, attribute correspondence edges, relationship connection edges, and source support edges to construct a heterogeneous knowledge evidence graph. It then extracts local subgraphs of candidate entities within two connection levels around the candidate entity pairs. For each candidate entity pair, virtual merging and independent retention are performed, comparing relationship integrity, attribute conflict, and neighborhood structure changes to identify synonymous entities, aliased entities, and entities with the same name but different meanings. Based on the multi-source entity alignment results, entities, attributes, and relationships are fused. Duplicate and conflicting knowledge is processed according to the data source credibility status, knowledge extraction confidence, time consistency, and relationship support status. Then, multi-source knowledge evidence units are corrected according to structural anomalies to generate a dynamic fused knowledge graph.

[0030] In practical knowledge processing, the names "Provincial Market Supervision Administration," "Provincial Market Supervision Bureau," and "Provincial Market Supervision Bureau" originate from policy documents, service guides, and database records, respectively. These three names correspond to the same agency code and are linked to the same policy entities and service items. The system identifies these three names as the same knowledge object, establishes entity identifier correspondences, and retains the data source and knowledge extraction confidence level for each name. The knowledge data also contains three entities all named "Enterprise Registration Window," corresponding to provincial, municipal, and district-level government service halls, respectively. These three entities share the same name but differ in their affiliated agencies, administrative divisions, service items, and relationship directions. After virtual merging, service items and addresses from different regions will be centrally connected to the same node, leading to attribute conflicts; when retained independently, each entity maintains a complete connection with its corresponding agency and service item. Based on this, the system identifies the three entities as homonymous entities and retains independent entity identifiers, preventing the mixing of service materials, service addresses, and consultation information from different regions.

[0031] To address duplicate knowledge, a certain business establishment matter uses three different names in policy documents, service guides, and database records: "legal person's identity certificate," "legal representative's identity document," and "legal representative's identity certificate materials," respectively. The system identifies these three records as sharing the same material knowledge based on attribute type, standard attribute content, and relationship support status, merging the corresponding data sources and knowledge extraction confidence levels. Regarding conflicting knowledge, a certain business license matter has historical policy records containing six materials, while current service guides and database records contain five. The system distinguishes between historical records and currently valid records based on time consistency and knowledge retention status, retaining the corresponding time information. During the knowledge graph writing process, the system also identifies some policy relationships still connected to historical institution identifiers, and some application materials incorrectly categorized as adjacent matters due to field hierarchy errors. After locating the corresponding multi-source knowledge evidence units through structural anomalies, the system corrects entity identifiers, attribute attribution, and relationship connections, and updates the nodes, edges, and time information in the knowledge graph, ensuring a consistent knowledge structure among policies, institutions, matters, and materials.

[0032] Table 1 Performance Comparison of Multi-Source Knowledge Extraction and Fusion Methods

[0033] As shown in Table 1, the method of this invention exhibits a relatively stable performance improvement in entity, attribute, and relation extraction. The F1 score for entity recognition increased from 89.4% in the traditional method to 92.1%, and the F1 scores for attribute and relation extraction increased from 86.8% to 89.7%. This improvement mainly stems from the confidence-enhanced PFN model's joint extraction of entities, attributes, and relations, and the processing of entity recognition information, relation connection information, and shared context information through entity semantic partitioning, relation semantic partitioning, and shared semantic partitioning, respectively. Traditional methods typically identify entities first and then match attributes and relations based on the entity recognition results. Errors in entity boundary or entity type identification can easily propagate to subsequent processing. The method of this invention simultaneously determines entity boundaries, entity types, attribute correspondence states, and relation connection states through a joint extraction layer, and uses a confidence calibration layer to determine the confidence level of knowledge extraction, reducing the continuous impact of entity recognition errors on the attribute and relation extraction process.

[0034] The multi-source entity alignment accuracy of the method in this invention reaches 92.6%, higher than the 88.3% of traditional methods. Traditional methods mainly rely on entity names, character overlap, or semantic similarity for entity alignment. When full names, abbreviations, aliases, or historical names are used in different data sources, the same knowledge object is easily retained as multiple entities; when multiple knowledge objects use the same name, misalignment is prone to occur. The method in this invention not only utilizes entity names and semantic features, but also binds attribute information, relational neighborhood, and data source to candidate entity pairs, performing virtual merging and independent retention through local subgraphs of candidate entities. By comparing relational integrity, attribute conflict, and neighborhood structure changes, it can be determined whether merging candidate entities causes relational breakage, attribute conflict, or node isolation, ensuring that the multi-source entity alignment process considers both semantic consistency and knowledge structure consistency.

[0035] The correct retention rate of conflicting knowledge increased from 84.7% to 89.9%. Traditional methods typically filter conflicting records based on fixed data source priority, update time, or single extraction confidence level, making it difficult to comprehensively assess the source support and time validity of knowledge content. This invention determines the knowledge credibility status based on the data source credibility and knowledge extraction confidence level, and further determines the knowledge retention status by combining time consistency and relationship support. For duplicate knowledge, the corresponding data sources and knowledge extraction confidence levels are merged; for conflicting knowledge, filtering is performed based on knowledge credibility status, time information, and relationship support, reducing the overlap of low-credibility records with high-credibility records and the mixing of historical knowledge with currently valid knowledge.

[0036] Regarding the quality of knowledge graph structure, the method of this invention achieves a structural anomaly detection rate of 88.4%, a 9.8 percentage point improvement compared to the 78.6% of traditional methods; the accuracy of structural anomaly correction increases from 80.9% to 87.2%. This improvement is directly related to the reverse correction mechanism for structural anomalies. After writing entities, attributes, and relationships into the knowledge graph, the method of this invention continues to check for missing nodes, incorrect relation endpoints, and abnormal relation connections. Based on the structural anomalies, it locates the corresponding multi-source knowledge evidence units and corrects entity identification, attribute attribution, and relation connections. Traditional methods typically end processing after knowledge is written into the graph, and errors formed during the extraction or entity alignment stages tend to persist. The method of this invention applies the graph structure check results in reverse to the multi-source knowledge evidence units, creating a closed-loop processing flow for knowledge extraction, entity alignment, knowledge fusion, and structural correction.

[0037] After merging duplicate knowledge, filtering conflicting knowledge, and correcting structural anomalies, the graph relation consistency rate of the method of this invention is improved from 91.2% to 95.0%. This result demonstrates that the starting and ending entities, relation types, and relation directions in the knowledge graph can maintain a higher degree of consistency. The performance improvement stems not only from the increased accuracy of basic extraction but also from the constraints imposed by entity merging inference on erroneous entity merging, the filtering of conflicting relations by knowledge retention status, and the correction of erroneous relation endpoints by structural anomaly correction. Therefore, the method of this invention can improve the overall reliability of multi-source knowledge extraction, entity alignment, conflict handling, and graph structure correction while achieving a reasonable performance improvement.

[0038] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge extraction and fusion method based on deep learning, characterized in that, Includes the following steps: Collect multi-source knowledge data and preprocess it to generate a multi-source knowledge dataset; The multi-source knowledge dataset is input into the confidence-enhanced PFN model to extract entities, attributes and relationships, and generate entity semantic features, knowledge extraction confidence and multi-source knowledge evidence units. Based on entity names and semantic features, filter the combinations of entities to be judged, associate the attribute information, relational neighborhood and data source of each entity, and generate candidate entity pairs; Using multi-source knowledge evidence units as nodes, association edges are established based on entity co-occurrence, attribute correspondence, relational connection, and source support to construct a heterogeneous graph of knowledge evidence, and candidate entity local subgraphs are extracted; Virtual merging and independent retention are performed on candidate entity pairs respectively. The relationship integrity state, attribute conflict state and neighborhood structure change are compared to generate entity merging inference results. By combining entity semantic features, attribute information, relational neighborhood, knowledge extraction confidence and entity merging inference results, synonymous entities, aliased entities and homonymous entities are identified, and multi-source entity alignment results are generated. Based on the multi-source entity alignment results, entities, attributes, and relationships are fused. Duplicate and conflicting knowledge is processed according to the data source credibility status, knowledge extraction confidence, time consistency, and relationship support status. Multi-source knowledge evidence units are corrected based on structural anomalies to generate a dynamic fused knowledge graph.

2. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The multi-source knowledge data includes text data, tabular data, semi-structured data, database record data, and knowledge graph data. The preprocessing includes character encoding unification, format conversion, field mapping, text segmentation, abnormal data removal, missing data completion, terminology standardization, source marking, and time alignment.

3. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The generation of entity semantic features, knowledge extraction confidence, and multi-source knowledge evidence units includes: The multi-source knowledge dataset is input into the confidence-enhanced PFN model, which includes a semantic encoding layer, a partitioning filtering layer, a joint extraction layer, a confidence calibration layer, and a knowledge evidence generation layer. In the semantic encoding layer, the multi-source knowledge dataset is encoded according to the content units and word order to generate a contextual semantic representation; In the partitioning filtering layer, entity gating weights, relation gating weights, and shared gating weights are calculated based on the context semantic representation, and the context semantic representation is divided into entity semantic partitions, relation semantic partitions, and shared semantic partitions. In the joint extraction layer, entity semantic partitioning and shared semantic partitioning are fused to identify entity boundaries and entity types, and relation semantic partitioning and shared semantic partitioning are fused to match attribute correspondence states and relation connection states between entities, generating entity, attribute, relation and entity semantic features. In the confidence calibration layer, the confidence of knowledge extraction is determined by combining the entity boundary recognition results, entity type recognition results, attribute matching results, and relationship matching results. In the knowledge evidence generation layer, entities, attributes, relationships, entity semantic features, and knowledge extraction confidence are bound together according to data sources and content units to generate multi-source knowledge evidence units.

4. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The generation of the candidate entity pairs includes: The entities in the multi-source knowledge evidence unit are grouped according to entity type, and entities with different data sources are selected from the same entity type and combined in pairs. For each entity combination, the entity name is normalized and segmented. The character overlap status, word inclusion status and abbreviation correspondence status between entity names are calculated. The entity combination that satisfies any name correspondence status is determined as the first entity combination to be judged. Calculate the semantic similarity between the semantic features of entities in the same entity combination, and determine the entity combination with a semantic similarity not lower than the preset semantic screening threshold as the second entity combination to be judged. Merge the first and second entity combinations to be determined and remove duplicate combinations to generate the entity combination to be determined. Retrieve attribute information, associated entities, relationship type and relationship direction of two entities in each combination of entities to be judged from the multi-source knowledge evidence unit, and generate the relationship neighborhood of each of the two entities; Two entities to be judged are bound together with their entity names, semantic features, attribute information, relational neighborhoods, and data sources. Entity pairs with the same data source and content unit are eliminated to generate candidate entity pairs.

5. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The construction of the knowledge evidence heterogeneous graph and the candidate entity local subgraph includes: Configure node identifiers for each multi-source knowledge evidence unit and bind entities, attributes, relationships, entity semantic features, knowledge extraction confidence, data sources, and content units; Within the same content unit, establish entity co-occurrence edges for nodes corresponding to entities that appear together; Establish attribute-corresponding edges for nodes of entities with the same attribute type, identical attribute content, or corresponding relationships; Based on the starting entity, ending entity, relationship type, and relationship direction, establish relationship connection edges between corresponding nodes; For nodes with different data sources but supporting the same entity, attribute, or relationship, establish source support edges, and determine the association strength by combining knowledge extraction confidence and the number of source supports; Nodes and various associated edges are combined according to node identifiers to generate a heterogeneous graph of knowledge evidence; In the heterogeneous graph of knowledge evidence, locate the nodes corresponding to the candidate entity pairs, expand outward along two connection levels: entity co-occurrence edge, attribute correspondence edge, relation connection edge, and source support edge, retain the nodes and associated edges within the expansion range, and generate a local subgraph of candidate entities.

6. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The generation of the entity merging deduction results includes: In the local subgraph of the candidate entity, locate the two nodes corresponding to the candidate entity pair, and copy the local subgraph of the candidate entity to form a virtual merged deduction branch and an independently retained deduction branch; In the virtual merge simulation branch, the two nodes are replaced with virtual merge nodes, attribute information and data sources are merged, the original entity co-occurrence edges, attribute corresponding edges, relationship connection edges and source support edges are reconnected to the virtual merge nodes, duplicate association edges are merged, and a virtual merge state is generated. In the independent retention derivation branch, maintain two nodes and their associated edges to generate an independent retention state; Count the number of retained relational connections, disconnected relational connections, and self-joined relational connections under the two states respectively, and generate the complete relational state; Compare the standard attribute content corresponding to the same attribute type, mark the attribute records with inconsistent content and no attribute correspondence as attribute conflicts, and generate an attribute conflict status. By comparing the associated entities, relationship types, relationship directions, and number of associated edges before and after the two nodes are merged, the retention of the common neighborhood, the increase or decrease of associated edges, and the isolation of nodes are determined, and the changes in the neighborhood structure are generated. The results of entity merging are generated by considering the complete state of the combination relationship, the conflicting state of attributes, and the changes in the neighborhood structure.

7. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The generation of the multi-source entity alignment result includes: Retrieve the entity name, semantic features, attribute information, relational neighborhood, and knowledge extraction confidence of the candidate entity pair, and extract the relational integrity state, attribute conflict state, and neighborhood structure change from the entity merging inference results; By combining entity semantic features, attribute correspondence status and relational neighborhood features, it is determined whether two entities point to the same knowledge object. The determination result is verified by knowledge extraction confidence and entity merging inference results. Candidate entity pairs that point to the same knowledge object but have different entity names are identified as synonymous entities; Synonymous entities whose names have corresponding full names, abbreviations, or aliases are identified as alias entities; Candidate entity pairs with the same entity name but pointing to different knowledge objects are identified as homonymous entities; Establish entity identifier correspondence between synonymous entities and aliased entities, retain independent entity identifiers for entities with the same name but different meanings, bind entity type, data source and corresponding entity identifier, and generate multi-source entity alignment results.

8. The knowledge extraction and fusion method based on deep learning according to claim 1, characterized in that, The generation of the dynamic fusion knowledge graph includes: Based on the unified entity identifier, synonymous entities and alias entities are merged, while the independent entity identifiers of entities with the same name but different meanings are retained, and attributes and relationships are mapped to the corresponding entities; Organize the knowledge to be merged according to entity identifier, attribute type and relational connection, and identify duplicate and conflicting knowledge. The knowledge credibility status is determined based on the credibility status of the data source and the confidence level of knowledge extraction, and the knowledge retention status is determined by combining time consistency and relation support status. Merge the data sources and knowledge extraction confidence levels corresponding to duplicate knowledge, filter conflicting knowledge according to knowledge retention status, and record the conflict resolution results; The retained entities, attributes, and relationships are written into the knowledge graph to identify structural anomalies caused by missing nodes, incorrect relationship endpoints, and abnormal relationship connections. Based on the multi-source knowledge evidence units corresponding to the structural anomaly location, entity identification, attribute attribution and relational connection are corrected, and the nodes, associated edges and time information in the knowledge graph are updated to generate a dynamic fusion knowledge graph.