A method and related apparatus for cross-document entity alignment that integrates domain ontology constraints and graph reasoning
By integrating domain ontology constraints and graph reasoning, this method solves the problems of high mismatch rate and poor interpretability in cross-document entity alignment in existing technologies, achieving high accuracy and interpretability in entity alignment, and adapting to the inconsistent entity naming and incremental update requirements of multi-source heterogeneous documents in the Industrial Internet.
Patent Information
- Application Number
- CN202611277260.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing cross-document entity alignment methods rely on strings or local semantics, which are difficult to handle term variations, structural sparsity, and insufficient incremental updates, resulting in high mismatch rates, poor interpretability, and a lack of effective entity duplication and relationship recognition capabilities.
By integrating domain ontology constraints and graph reasoning, and combining hierarchical decision-making and local incremental update mechanisms, an entity co-occurrence graph is constructed and multi-feature weighted fusion is introduced. The entity node representation is updated using an attention mechanism, the structural consistency score is calculated, and ontology constraint loss and consistency loss are introduced to achieve high-accuracy and interpretable cross-document entity alignment.
It significantly reduces the risk of misalignment and omission, improves the accuracy and interpretability of entity alignment, has online response and self-optimization capabilities, and adapts to the needs of inconsistent entity naming and incremental updates in multi-source heterogeneous documents of the Industrial Internet.
Smart Images

Figure CN123115379A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial internet knowledge graph construction technology, specifically involving a cross-document entity alignment method and related apparatus that integrates domain ontology constraints and graph reasoning. Background Technology
[0002] As the application of the Industrial Internet in manufacturing, energy, and equipment operation and maintenance deepens, enterprises have gradually accumulated a large amount of multi-source heterogeneous document data, including equipment manuals, inspection records, fault analysis reports, maintenance work orders, process specifications, and expert experience summaries. Inconsistent naming rules across different systems and departments result in the coexistence of aliases, abbreviations, and historical codes for the same entity in different documents. This can easily lead to entity duplication, relationship breaks, conceptual confusion, and knowledge conflicts, thereby affecting the accuracy and maintainability of knowledge graph construction.
[0003] Existing cross-document entity alignment methods mostly rely on string similarity or local semantic similarity, which are easily affected by terminology variations, cross-domain abbreviations, noisy text, and incomplete context. They also struggle to fully utilize the constraints of entities at the relational structure, business rules, and ontology levels. The candidate generation process often lacks effective blocking and semantic filtering, resulting in an excessively large candidate size, increased computational overhead, and a higher false-match rate. On the other hand, relying solely on local textual clues makes it difficult to identify implicit synonyms and hierarchical relationships across documents.
[0004] Furthermore, many existing methods emphasize offline, one-time alignment, lacking an update mechanism for incremental data. When equipment changes, process upgrades, or maintenance strategy adjustments occur, it is difficult to maintain consistency in the old alignment results, which can easily lead to long-term knowledge graph drift, untraceable historical alignments, and conflicts between old and new rules. The interpretability of the alignment results is also weak; engineers find it difficult to understand the basis for merging and the source of confidence, thereby reducing business credibility and actual adoption rate. Summary of the Invention
[0005] The purpose of this invention is to provide an Ontology-Constrained Graph Reasoning for Cross-Document Entity Alignment (OGREA) method and related apparatus that integrates domain ontology constraints and graph reasoning. This addresses the technical problems of existing methods that rely solely on strings or local semantics, making it difficult to handle term variations, structural sparsity, and insufficient incremental updates, resulting in high mismatch rates and poor interpretability. The technical solution disclosed in this invention combines domain ontology constraints with graph structure reasoning and introduces hierarchical decision-making and local incremental update mechanisms to achieve high-accuracy, interpretable, and sustainably evolving cross-document entity alignment.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning, comprising: Step 1: Obtain the multi-source industrial document data to be processed; Step 2: Based on the pre-built domain ontology model and relation constraint rule library, entity extraction is performed on the multi-source industrial document data to generate an entity set and an entity co-occurrence graph. Each entity in the entity set is associated with an entity type, a context vector representation, an attribute vector representation, and an ontology concept representation. In the entity co-occurrence graph, intra-sentence dependency relations and inter-sentence co-occurrence relations are used as document context relation edges, and constraint relations in the relation constraint rule library are used as ontology constraint edges. Step 3: Generate multiple candidate entity pairs based on the entity set, and remove candidate entity pairs that are incompatible in type; for each remaining candidate entity pair, calculate semantic similarity, attribute consistency, ontology semantic distance, initial structure consistency, and string similarity based on context vector representation, attribute vector representation, ontology concept representation, entity co-occurrence graph, and entity name, respectively. After multi-feature weighted fusion and probability normalization, generate a set of candidate entity pairs and the matching probability of each candidate entity pair. Step 4: Using candidate entity pairs in the candidate entity pair set as alignment edges to be verified, and using entities as nodes, integrate document context relationship edges and ontology constraint edges for unified modeling to obtain a cross-document reasoning graph; on the cross-document reasoning graph, calculate the corresponding attention weights for document context relationship edges and ontology constraint edges respectively, and update the entity node representation by weighted summation of each neighborhood information through relation-aware aggregation, and calculate the structural consistency score of each candidate entity pair based on the updated entity node representation; Step 5: Based on entity node representation and structural consistency score, calculate consistency loss using matching probability, introduce ontology constraint loss and consistency loss, combine alignment loss based on marginal ranking and regularization term to construct joint objective function, and output alignment confidence of each candidate entity pair after global constraint optimization and confidence calibration. Step 6: Compare the alignment confidence of each candidate entity pair with the pre-acquired high and low thresholds respectively: if the alignment confidence is greater than or equal to the high threshold, the corresponding candidate entity pair is determined to be aligned; if the alignment confidence is less than the high threshold but greater than or equal to the low threshold, the corresponding candidate entity pair is transferred to manual review; if the alignment confidence is less than the low threshold, the corresponding candidate entity pair is determined to be unaligned and excluded.
[0007] A further improvement to the technical solution of this invention is that, in step 4, the expression for calculating the attention weight is: ; In the formula, For nodes i With adjacent nodes j In relation r Attention weights; For relationship r The attention parameter vector; For relationship r The learnable weight matrix below; , , They are respectively i , j , The current representation vector; For nodes i The neighborhood set under relation r; for Node index in; This represents a vector concatenation operation; This indicates the transpose operation.
[0008] A further improvement to the technical solution of this invention is that, in step 4, the update expression representing the entity node is: ; In the formula, For nodes i In the Layer representation; For activation functions; A set of relation types; For the first Hierarchical Relationship r The learnable weight matrix below; For adjacent node j in the th... Layer representation.
[0009] A further improvement to the technical solution of the present invention is that, in step 5, the joint objective function... for: ; ; ; In the formula, Alignment loss based on marginal ranking; This is the loss due to ontology constraints; This results in a loss of consistency. It is a regularization term; , , These are the weighting coefficients for ontology constraint loss, consistency loss, and regularization term; It is a relational mapping matrix; , These are the representation vectors for the head entity and the tail entity, respectively; It is a relation vector; For labeling or high-confidence pseudo-labels; For candidate entity pairs The matching probability.
[0010] A further improvement of the technical solution of the present invention is that, in step 3, the specific steps of generating multiple candidate entity pairs based on the entity set include: generating candidate entity pairs using a multi-key blocking strategy; wherein, the blocking key includes at least one of the following: encoding prefix, region label, process segment number, and time window.
[0011] A further improvement to the technical solution of the present invention is that, in step 6, the high threshold... and low threshold Determine as follows: ; ; In the formula, Basic threshold; For risk factors, the range of values is: ; This is the risk adjustment coefficient.
[0012] A further improvement to the technical solution of the present invention is that it further includes: Step 7: In response to a new document or a change in the relation constraint rule base, obtain the cross-document reasoning graph constructed in Step 4 and the review result output in Step 6; when the triggering source is a new document, determine the affected entities based on the association between the entities in the new document and the existing entities in the cross-document reasoning graph; when the triggering source is a change in the relation constraint rule base, determine the affected entities in the cross-document reasoning graph based on the updated constraint conditions. Trigger scores are calculated based on the magnitude of attribute changes, relational topology changes, and business risk weighting of the affected entities. When the obtained trigger score exceeds a preset threshold, local re-inference is performed on the k-hop neighborhood of the affected entities on the cross-document inference graph, and the corresponding index and version identifier are updated. If the update result violates ontology constraints or the quality indicators are abnormal, the system is rolled back to the most recent stable version. The verification results are used as new training samples for subsequent use when the score exceeds the threshold to perform local re-inference.
[0013] A second aspect of the present invention provides a cross-document entity alignment system that integrates domain ontology constraints and graph reasoning, comprising: The data acquisition unit is used to acquire multi-source industrial document data to be processed; The entity extraction and representation unit is used to extract entities from the multi-source industrial document data based on a pre-built domain ontology model and a relation constraint rule library, generating an entity set and an entity co-occurrence graph. Each entity in the entity set is associated with an entity type, a context vector representation, an attribute vector representation, and an ontology concept representation. In the entity co-occurrence graph, intra-sentence dependency relations and inter-sentence co-occurrence relations are used as document context relation edges, and constraint relations in the relation constraint rule library are used as ontology constraint edges. The candidate generation and matching unit is used to generate multiple candidate entity pairs based on the entity set and remove candidate entity pairs that are incompatible in type; for each retained candidate entity pair, semantic similarity, attribute consistency, ontology semantic distance, initial structure consistency and string similarity are calculated based on context vector representation, attribute vector representation, ontology concept representation, entity co-occurrence graph and entity name, respectively; after multi-feature weighted fusion and probability normalization, a candidate entity pair set and the matching probability of each candidate entity pair are generated. The reasoning graph construction and structure scoring unit is used to take candidate entity pairs in the candidate entity pair set as alignment edges to be verified, and to use entities as nodes to perform unified modeling by integrating document context relationship edges and ontology constraint edges to obtain a cross-document reasoning graph. On the cross-document reasoning graph, the corresponding attention weights are calculated for document context relationship edges and ontology constraint edges respectively. The neighboring information is weighted and summed through relation-aware aggregation to update the entity node representation. The structure consistency score of each candidate entity pair is calculated based on the updated entity node representation. The constraint optimization and confidence output unit is used to calculate the consistency loss based on entity node representation and structural consistency score, using matching probability, introducing ontology constraint loss and consistency loss, and constructing a joint objective function by combining the alignment loss based on marginal ranking and regularization term. After global constraint optimization and confidence calibration, the alignment confidence of each candidate entity pair is output. The hierarchical decision unit is used to compare the alignment confidence of each candidate entity pair with the pre-acquired high and low thresholds respectively: if the alignment confidence is greater than or equal to the high threshold, the corresponding candidate entity pair is determined to be aligned; if the alignment confidence is less than the high threshold but greater than or equal to the low threshold, the corresponding candidate entity pair is transferred to manual review; if the alignment confidence is less than the low threshold, the corresponding candidate entity pair is determined to be unaligned and excluded.
[0014] In a third aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements a cross-document entity alignment method for fusing domain ontology constraints and graph reasoning as described in any one of the first aspects of the present invention.
[0015] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the cross-document entity alignment method for fusing domain ontology constraints and graph reasoning as described in any one of the first aspects of the present invention.
[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention addresses the problems of high mismatch rates, excessively large candidate size, and poor interpretability caused by inconsistent entity naming, terminology variations, and the coexistence of abbreviations and codes in multi-source heterogeneous documents in the Industrial Internet. It proposes a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning. Its significant advancements lie in: Firstly, by constructing an entity co-occurrence graph in step 2, document contextual relationship edges are established based on intra-sentence dependency relations and inter-sentence co-occurrence relations, while ontology constraint edges are established based on constraint relations in the relation constraint rule base, providing structural evidence support for subsequent reasoning. Secondly, step 3 generates matching probabilities through multi-feature weighted fusion and probability normalization, integrating semantic similarity, Multi-dimensional evidence, including attribute consistency, ontology semantic distance, initial structural consistency, and string similarity, improves the accuracy of candidate entity pair selection. Step 4 updates entity node representations and calculates structural consistency scores through a relation-aware attention mechanism, achieving full mining of global structural evidence and effectively suppressing noise propagation. Step 5 introduces ontology constraint loss and consistency loss to construct a joint objective function, ensuring that the alignment results achieve global optimization while meeting business rule constraints. Step 6 uses a dual-threshold hierarchical decision-making process to achieve hierarchical output of automatic confirmation, manual review, and automatic exclusion, combining high accuracy with engineering usability. The steps of this invention form a closed-loop logic of "candidate generation, structural reasoning, confidence calibration, and hierarchical decision-making," significantly reducing the risk of misalignment and missed alignment.
[0017] This invention further refines and limits the method for calculating attention weights in step 4. Its significant advancement lies in: introducing a relationship-specific attention parameter vector. With learnable weight matrix Attention weights are calculated for document context relationship edges and ontology constraint edges respectively, enabling the model to distinguish the importance of different types of neighborhood information, highlight the contribution of high-value relationship evidence to entity representation, and suppress noise propagation, thereby improving the accuracy of structural consistency scoring and the interpretability of graph reasoning process.
[0018] This invention further refines and limits the entity node representation update method in step 4. Its significant progress lies in: adopting a multi-layer relation-aware aggregation mechanism, by aggregating the weighted information of adjacent nodes in each relation neighborhood layer by layer, the entity node representation can fully integrate ontology constraints and structural evidence, and achieve deep mining of cross-document structural consistency; multi-layer iterative update enables long-distance relation evidence to gradually propagate to the target node, effectively solving the problem of insufficient evidence in structurally sparse scenarios, and improving the ability to identify implicit synonyms and hierarchical relationships.
[0019] This invention further refines and limits the joint objective function in step 5. Its significant progress lies in the organic integration of alignment loss based on marginal ranking, ontology constraint loss, consistency loss and regularization term. Through multi-task joint optimization, it achieves the unity of alignment accuracy, rule compliance and structural consistency. When textual evidence conflicts with ontology constraints, the joint optimization mechanism can automatically weigh the contributions of each party and prioritize the retention of candidate entity pairs that meet ontology constraints and have high time freshness, effectively reducing the risk of misalignment.
[0020] This invention further refines and limits the dual threshold determination method in step 6. Its significant progress lies in: by introducing risk factors to dynamically adjust the high and low thresholds, the adaptive adjustment of the hierarchical decision threshold is achieved; when the risk factor is high, the automatic confirmation interval is narrowed and the manual review interval is widened, effectively controlling the risk of misalignment propagation in high-risk scenarios; when the risk factor is low, the system can appropriately expand the automatic confirmation range to improve processing efficiency, enabling the method to flexibly adapt to the security requirements of different business scenarios.
[0021] This invention further introduces an incremental update mechanism, whose significant advancement lies in: responding to new document additions or changes in the relation constraint rule base, determining the affected entities based on the trigger source type, calculating a weighted trigger score based on the magnitude of attribute changes, the magnitude of relation topology changes, and business risks, and only updating the scores of affected entities that exceed a preset threshold. By skipping neighborhoods to perform local re-inference, the high computational overhead of global recalculation is avoided. Version identification and rollback mechanisms ensure the traceability and consistency of update results, solving the problems of long-term knowledge graph drift and conflicts between old and new rules caused by the lack of incremental update capabilities in existing methods. The review results are fed back as new training samples for subsequent use when the score exceeds the threshold to perform local re-inference, forming a continuous evolution loop of "event triggering, local update, version management, and data feedback", enabling the system to have online response capabilities and self-optimization capabilities. Attached Figure Description
[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the main process of a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning in an embodiment of the present invention; Figure 3 This is a schematic diagram of the principle architecture of a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning in an embodiment of the present invention. Figure 4 This is a schematic diagram of the cross-document reasoning graph and ontology constraint relationship in an embodiment of the present invention; Figure 5 This is a schematic diagram of the confidence threshold decision and verification mechanism in an embodiment of the present invention; Figure 6 This is a schematic diagram of an event-triggered incremental update closed loop in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of protection of the present invention, but rather to illustrate specific selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0025] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other.
[0026] Please see Figure 1 The present invention provides a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning, comprising the following steps: Step 1: Obtain the multi-source industrial document data to be processed.
[0027] Step 2: Based on the pre-built domain ontology model and relation constraint rule library, entity extraction is performed on the multi-source industrial document data to generate an entity set and an entity co-occurrence graph. Each entity in the entity set is associated with an entity type, a context vector representation, an attribute vector representation, and an ontology concept representation. In the entity co-occurrence graph, intra-sentence dependency relations and inter-sentence co-occurrence relations are used as document context relation edges, and constraint relations in the relation constraint rule library are used as ontology constraint edges.
[0028] Step 3: Generate multiple candidate entity pairs based on the entity set, and remove candidate entity pairs that are incompatible in type; for each remaining candidate entity pair, calculate semantic similarity, attribute consistency, ontology semantic distance, initial structure consistency, and string similarity based on context vector representation, attribute vector representation, ontology concept representation, entity co-occurrence graph, and entity name, respectively. After multi-feature weighted fusion and probability normalization, generate a set of candidate entity pairs and the matching probability of each candidate entity pair.
[0029] Specifically, in this step, when generating multiple candidate entity pairs based on the entity set, multiple candidate entity pairs are generated based on a multi-key blocking strategy according to the entity set; wherein, the blocking key used in the multi-key blocking strategy includes at least one of the following: encoding prefix, region label, process segment number, and time window.
[0030] Specifically, for each candidate entity pair, a type consistency check is performed based on the entity types of the two entities. If the entity types of the two entities do not have a compatible relationship within the predefined type hierarchy in the relation constraint rule base, they are eliminated.
[0031] Specifically, this step involves calculating semantic similarity using context vector representation, calculating attribute consistency using attribute vector representation and using it as a gating factor, calculating ontology semantic distance using ontology concept representation, calculating initial structural consistency using entity co-occurrence graph, and calculating string similarity using the names or codes of the two entities in the candidate entity pair. The semantic similarity, attribute consistency, ontology semantic distance, initial structural consistency, and string similarity of each candidate entity pair are then weighted and fused to obtain a comprehensive similarity for each candidate entity pair. Probability normalization is then performed based on the comprehensive similarity to obtain the matching probability of each candidate entity pair. Finally, the set of candidate entity pairs and the matching probability of each candidate entity pair are output.
[0032] Step 4: Using candidate entity pairs in the candidate entity pair set as alignment edges to be verified, and taking entities as nodes, integrate document context relationship edges and ontology constraint edges for unified modeling to obtain a cross-document reasoning graph; on the cross-document reasoning graph, calculate the corresponding attention weights for document context relationship edges and ontology constraint edges respectively, and update the entity node representation by weighted summation of each neighborhood information through relation-aware aggregation, and calculate the structural consistency score of each candidate entity pair based on the updated entity node representation.
[0033] Specifically, in this step of calculating the structural consistency score, based on the updated entity node representation, the vector similarity between the two entity node representations corresponding to each candidate entity pair is calculated, and the vector similarity is used as the structural consistency score of the candidate entity pair.
[0034] Step 5: Based on entity node representation and structural consistency score, calculate consistency loss using matching probability, introduce ontology constraint loss and consistency loss, combine alignment loss based on marginal ranking and regularization term to construct joint objective function, and output alignment confidence of each candidate entity pair after global constraint optimization and confidence calibration.
[0035] Step 6: Compare the alignment confidence of each candidate entity pair with the high threshold and the low threshold respectively: If the alignment confidence is greater than or equal to the high threshold, the corresponding candidate entity pair is automatically confirmed as aligned; if the alignment confidence is less than the high threshold but greater than or equal to the low threshold, the corresponding candidate entity pair is transferred to manual review, and the evidence chain and the reason for review can be recorded; if the alignment confidence is less than the low threshold, the corresponding candidate entity pair is automatically rejected (i.e., judged as unaligned and excluded).
[0036] Specifically, in this step, the base threshold can be adjusted based on the risk factor to calculate the high threshold and the low threshold. The higher the risk factor, the wider the range between the high and low thresholds, and the more stringent the automatic confirmation and rejection decisions.
[0037] Further optional technical solutions include: outputting the manual review results, their evidence chain, and review reasons as review samples to step 7; in step 7, obtaining the cross-document reasoning graph constructed in step 4, the review samples output in step 6, new documents, and relationship constraint rule base change events; calculating the attribute change magnitude of entities, the relationship topology change magnitude, and the weighted trigger score of business risks; when the trigger score exceeds a preset threshold, performing local re-reasoning on the k-hop neighborhood of the affected entities in the cross-document reasoning graph, and updating the corresponding index and version identifier; if the update result violates ontology constraints or the quality indicators are abnormal, rolling back to the most recent stable version; and retraining the review samples to output alignment records.
[0038] In the technical solution disclosed in this invention, a cross-document reasoning graph is constructed, and entity nodes, document context relationship edges, and ontology constraint edges are modeled in a unified manner. Global structural consistency reasoning is performed through a relationship-aware attention mechanism to obtain entity representations that integrate ontology constraints and structural evidence. Ontology constraint loss and consistency loss are introduced to construct a joint objective function for global constraint optimization and confidence calibration, and the alignment confidence of each candidate entity pair is output. Based on dual thresholds and risk factors, hierarchical decision-making is performed, and three types of alignment results are output: automatic confirmation, manual review, or automatic rejection. In response to new documents or change events, local incremental updates are triggered, the review results are fed back into training, and continuously evolving alignment records are output. This invention addresses the issues of inconsistent entity naming, structural sparsity, and insufficient incremental updates in multi-source heterogeneous documents within the Industrial Internet. It proposes a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning. This method efficiently generates candidate entity pairs through a multi-key blocking strategy, combining semantics, attributes, ontology distance, and structural consistency for multi-feature weighted fusion and probability normalization to ensure the accuracy and recall of the candidate set. Based on this, a cross-document reasoning graph is constructed, modeled uniformly with document context relationship edges and ontology constraint edges. A relationship-aware attention mechanism updates entity node representations and calculates structural consistency scores, achieving comprehensive mining of global structural evidence. Furthermore, ontology constraint loss and consistency loss are introduced to construct a joint objective function for global optimization and confidence calibration. Combined with dual-threshold hierarchical decision-making and dynamic adjustment of risk factors, it achieves hierarchical output with automatic confirmation, manual review, and automatic exclusion. Simultaneously, in response to new document additions and changes in the relation constraint rule base, weighted trigger scoring drives local re-reasoning and version rollback, with review samples being fed back for training, forming a continuously evolving closed loop. This invention effectively improves the accuracy and interpretability of entity alignment, significantly reduces the risk of misalignment, and possesses good engineering usability and scalability.
[0039] Please see Figures 2 to 6 This invention provides a cross-document entity alignment method that integrates ontology constraints and graph reasoning, illustrating the implementation process of cross-document entity alignment in a multi-source document scenario of the Industrial Internet, mainly following these steps: Step 1: Construct an ontology model for the industrial domain, define entity type hierarchy, attribute constraints and relation constraints, and form a rule base.
[0040] Step 2: Preprocess the multi-source industrial documents and extract entities to form a document-level entity set and its context vector and attribute vector representation.
[0041] Step 3: Generate candidate entity pairs and calculate the comprehensive similarity and matching probability of candidate entity pairs to obtain the set of candidate entity pairs and their confidence scores.
[0042] Step 4: Construct a cross-document reasoning graph and perform relation-aware graph reasoning to obtain entity representations that fuse structural evidence.
[0043] Step 5: Introduce ontology constraint loss and consistency loss to construct a joint optimization objective, and perform global consistency constraints and calibration on the alignment results.
[0044] Step 6: Based on the dual thresholds and risk factors, make a graded decision and divide the alignment results into automatic confirmation, manual review and automatic rejection.
[0045] Step 7: Calculate the trigger score for new documents and change events and perform local incremental updates to form a closed loop of review feedback and version audit.
[0046] like Figure 3 As shown, the execution system as a whole can be composed of a multi-source document access layer, an ontology construction and rule layer, an entity extraction and encoding layer, a cross-document reasoning layer, and an alignment decision and publishing layer, and further involves a manual review platform and incremental update and reflow training.
[0047] Specifically, in step 1, the process of constructing the industrial domain ontology model and relation constraint rule base is as follows: Step 1.1: Organize the conceptual system of the industrial domain, define entity types such as equipment, components, processes, parameters, faults, and maintenance actions, and their hierarchical relationships. Set attribute field sets and value ranges for different entity types to construct an industrial domain ontology model. Specifically, define required attributes, optional attributes, value ranges, unit specifications, and field integrity constraints for different entity types, forming an attribute constraint set.
[0048] Step 1.2: Establish a relation constraint rule base and encode it into a constraint matrix. This matrix is used to validate the legitimacy of candidate entity alignment edges and relation triples. It includes valid and prohibited combinations of compositional, dependency, causal, and maintenance relations. The rules are then encoded into a constraint matrix for consistency validation. Specifically, this involves generating ontology concept vectors. c e The rule base is encoded into a constraint matrix. R c This is for use in subsequent reasoning and verification.
[0049] Step 1.3: Merge synonyms and abbreviations to create an alias dictionary (Alias) for subsequent candidate generation and verification; and mark the validity period of historical codes to prevent expired codes from participating in the current alignment.
[0050] Specifically, in step 2, the process of multi-source document preprocessing, entity extraction, and representation construction is as follows: Step 2.1: Perform structured preprocessing on the multi-source documents, including terminology normalization, encoding mapping, unit unification and time standardization, output a standardized text set D, and record source system and timestamp metadata.
[0051] Step 2.2: Extract entities using a domain-dictionary-enhanced named entity recognition model to form a document-level entity set. E d And generate a context vector representation for each entity. h e Parse the entity attribute fields to form attribute vectors. a e It also records the source system and timestamp metadata.
[0052] Step 2.3: Construct the entity co-occurrence graph G d Intra-sentence dependency relations and inter-sentence co-occurrence relations are used as structural edges to provide evidence for subsequent structural consistency calculations; the context consistency scores of entities are calculated. q e This is used for subsequent confidence calibration.
[0053] Step 2.4: Match entity types with ontology concept vectors c e Alignment yields entity concept representations. g e This is used for subsequent ontology constraint reasoning.
[0054] Further processing can be performed as follows: standardize the entity context vector to avoid representation bias caused by differences in document length; filter out noise words and stop words to reduce the interference of invalid context on semantic representation; and adaptively adjust the entity context window length, expanding the window for short texts and compressing the window for long texts.
[0055] Specifically, in step 3, the specific implementation process of candidate entity pair generation, comprehensive scoring, and probabilistic approach is as follows: Step 3.1: Generate a set of candidate entity pairs based on multi-key blocking The blocking key includes the encoding prefix, region label, process segment number and time window, which are combined with vector recall in parallel to cover aliases and fuzzy naming.
[0056] Step 3.2: Calculate string similarity for candidate entity pairs. semantic similarity Body distance Consistency with structure And calculate the comprehensive similarity score according to formula (1). Attribute consistency It is used as a gating factor to adjust candidate weights.
[0057] (1) In the formula, to , where are the weighting coefficients corresponding to string similarity, semantic similarity, ontology distance, and structural consistency, respectively. This is a temperature coefficient used to control the smoothness of the distribution; As an attribute consistency gating factor, it participates in candidate weight adjustment and, combined with type consistency checks, reduces the weight or eliminates candidate entity pairs with conflicting attributes or inconsistent types, retaining Top-K candidates to enter the graph reasoning stage.
[0058] Step 3.3: Perform Softmax normalization on the scores of candidate entity pairs and calculate the matching probability according to equation (2). , where τ is a temperature coefficient used to control the smoothness of the distribution.
[0059]
[0060] Step 3.4: Introduce attribute consistency gating and type consistency checks. Candidate entity pairs with conflicting attributes or inconsistent types will have their weights reduced or be eliminated, while the Top-K candidates will be retained to enter the graph reasoning stage.
[0061] Specifically, this step introduces attribute consistency gating, if If the candidate entity pair's score falls below a threshold or there is a conflict in key attributes, its weight is reduced or it is transferred to manual review. Initial screening of candidate entity pairs is performed, retaining the Top-K pairs for the graph reasoning stage. A type consistency check is performed on candidate entity pairs; those with inconsistent types are directly eliminated. A higher candidate retention threshold is applied to high-risk entities to reduce the propagation of misalignment. Semantic consistency verification is performed on candidate entity pairs; if the difference between semantic similarity and structural similarity is too large, its weight is reduced. Temporal consistency checks are performed on candidate entity pairs; entity pairs with obvious time window conflicts are eliminated.
[0062] Please see Figure 4 Specifically, in step 4, the implementation process of cross-document reasoning graph construction and relation-aware graph reasoning is as follows: Step 4.1: Construct a cross-document reasoning graph G=(V,E), where V is the set of entity nodes, E contains context relation edges and ontology constraint edges, and add candidate entity pairs as verifiable alignment edges to the graph structure.
[0063] Step 4.2: Introduce relation attention weights for different relation types and calculate them according to equation (3). This is to highlight high-value relationship evidence and suppress noise propagation.
[0064]
[0065] In the formula, Let be the attention weight of node i to neighbor node j under relation r; Let r be the attention parameter vector of relation r; For relationship r The learnable weight matrix below; , , They are respectively i , j , The current representation vector; This indicates a vector concatenation operation.
[0066] Step 4.3: Node update adopts relation-aware aggregation. The attention weights obtained in step 4.2 are used to weight and aggregate the neighborhood information of each relation to obtain the updated node representation. The calculation process is shown in equation (4), where σ is the activation function and l is the layer index.
[0067]
[0068] In the formula, For nodes i In the Layer representation; For activation functions; A set of relation types; For the first Hierarchical Relationship r The learnable weight matrix below; For adjacent node j in the th... Layer representation.
[0069] Step 4.4: Represent the nodes obtained from equation (4) Perform normalization and residual fusion, calculate the structural consistency score, and feed it back to the comprehensive scoring item. .
[0070] Specifically, in this step, normalization is first performed to obtain a stable representation of the entity structure. ; Calculate structural consistency score It is used for joint decision-making with semantic scores; it prunes cross-document relationship paths to retain evidence that is consistent with ontology constraints and has high path confidence; it truncates relationship attention weights to prevent extreme weights from causing propagation instability; it segments and normalizes structural similarity to make structural scores comparable under different document sizes; and it adds residual connections and intra-layer normalization to node representations to improve inference stability.
[0071] Specifically, in step 5, the implementation process for ontology constraint consistency optimization and training target construction is as follows: Step 5.1: Introduce ontology constraint loss to ensure consistency of relation triples, and calculate according to equation (5). ,in , Let r represent the head entity and the tail entity, respectively, and r be the relation vector. This is the relational mapping matrix.
[0072]
[0073] Step 5.2: Introduce the consistency loss constraint alignment result and calculate according to equation (6). ,in, For annotation or high-confidence pseudo-labels.
[0074]
[0075] Step 5.3: The overall objective function is constructed according to equation (7), where To align task losses, For regularization terms, to These are the weighting coefficients.
[0076]
[0077] In the formula, For alignment loss based on marginal ranking, For regularization terms, and Let r represent the head entity and the tail entity, respectively, and r be the relation vector. It is a relational mapping matrix. For labeling or high-confidence pseudo-labels, For consistency loss, For regularization terms, to , which is the weighting coefficient; when textual evidence conflicts with ontology constraints, candidate entity pairs that satisfy ontology constraints and have higher time freshness are preferred.
[0078] Step 5.4: When textual evidence conflicts with ontology constraints, priority is given to retaining candidate entity pairs that satisfy ontology constraints and have high time freshness, and hard examples are reweighted on boundary samples to improve robustness.
[0079] Specifically, in this step, hard-example reweighting training is performed on low-confidence samples to improve the model's ability to distinguish boundary samples. A regularization term is introduced to suppress overfitting and ensure the alignment results are stable across different batches of data. Moving averages are applied to the model parameters to reduce training fluctuations and improve the stability of alignment confidence. Weight scheduling is applied to the consistency loss to ensure that structural evidence is emphasized in the early stages of training and semantic consistency is strengthened in the later stages.
[0080] See Figure 5 , specifically, in step 6, the specific implementation process of dual-threshold hierarchical decision-making, manual review and evidence chain output is as follows: Step 6.1: Set dual thresholds T1 and T2 to implement hierarchical decision-making, and dynamically adjust them according to the risk factor Risk according to formula (8), where μ is the base threshold when risk factors are not considered, and γ is the risk adjustment coefficient for adjusting the decision threshold according to the risk factor.
[0081]
[0082] Wherein, the value range of Risk is [0,1]; S is the final alignment confidence of the candidate entity pair, T1 is the high decision threshold, T2 is the low decision threshold, and T1>T2; the value range of Risk is [0,1]; when S≥T1, the alignment is automatically confirmed; when T2≤S<T1, manual review is entered and the evidence chain and judgment reason are recorded; when S<T2, it is automatically rejected.
[0083] Step 6.2: When S≥T1, the alignment is automatically confirmed; when T2≤S<T1, manual review is entered; when S<T2, it is automatically rejected, and the decision result and key evidence are written into the audit record.
[0084] Step 6.3: Display string evidence, semantic evidence, attribute consistency evidence and graph structure propagation path summary for the manual review samples, and record the review reasons to flow back to the training set.
[0085] Specifically, in this step, the evidence chain and judgment reasons are recorded for the manual review samples, and the training set is flowed back for model iteration. Stricter thresholds and review priorities are enabled for high-risk entities. Differentiated threshold strategies are set for different business scenarios, and the maintenance scenario prioritizes reducing the risk of missed alignment. Prioritize the ranking of review samples, and prioritize processing high-impact entities and high-frequency entities. Generate audit records for review results, including reasons, evidence chain summaries and operator information. Perform statistical analysis on the review samples, and dynamically adjust the thresholds to control the misalignment rate. A secondary confirmation mechanism is implemented for the review results of high-risk entities to reduce the propagation risk of misalignment.
[0086] See Figure 6 , specifically, in step 7, the specific implementation process of the event-triggered incremental update and version rollback mechanism is as follows: Step 7.1: Calculate the trigger score U(v) for new documents and change events, and measure attribute changes according to formula (9) , relationship changes and the comprehensive impact of Risk(v).
[0087]
[0088] in, Indicates the magnitude of change in entity attributes. The value represents the magnitude of the change in entity relationship topology, Risk(v) represents the entity business risk, and w1 to w3 are the weight coefficients corresponding to changes in entity attributes, changes in entity relationship topology, and entity business risk, respectively. When U(v) exceeds the threshold, local re-inference is performed on the k-hop neighborhood of entity v and the index and version record are updated. If the ontology constraints are violated or the quality indicators are abnormal, the system is rolled back to the most recent stable version and the review samples are fed back for training iteration.
[0089] Step 7.2: When U(v) exceeds the threshold, perform local re-inference on the k-hop neighborhood of entity v and update the index and version record; if the ontology constraint is violated or the quality index is abnormal, roll back to the most recent stable version.
[0090] In an optional implementation of the above steps, the alignment result is output as a structured record of (unified entity ID, original entity ID set, confidence S, evidence chain summary, version number, timestamp) and is synchronously written to the graph database and retrieval index to support subsequent reasoning and tracing.
[0091] Specifically, this step involves: performing consistency checks on the update results; rolling back to the most recent stable version if ontology constraints are violated; writing update logs to an audit table, recording the version number, scope of impact, and rollback flag; triggering a consistency inspection after incremental updates to check for anomalies in the relationship topology after entity merging; performing a consistency snapshot on the partially updated indexes to compare changes in key metrics before and after the update; rolling back entities that failed to update to the previous stable version and marking them as pending observation; performing a partial consistency inspection on the incrementally updated entity set, immediately alerting if anomalies are detected; and periodically summarizing the audit logs to generate an alignment quality monitoring report.
[0092] Existing solutions struggle to handle terminology variations and structural sparsity, easily leading to long-term knowledge graph drift and conflicts between old and new rules. Alignment results are untraceable, and engineers struggle to understand the merging criteria, resulting in low business credibility and insufficient actual adoption rates. The technical solution disclosed in this invention achieves accurate, interpretable, and sustainably evolving entity alignment through ontology constraints and graph structure reasoning, combined with hierarchical decision-making and local incremental updates. Specifically, this invention addresses the problems of inconsistent entity naming, semantic conflicts, and continuous incremental updates in multi-source documents, employing a unified governance process of "candidate generation—structural reasoning—threshold decision—review and feedback." In the candidate generation stage, candidate entity pairs are constructed by integrating string similarity, semantic similarity, attribute consistency, and ontology semantic distance, and a multi-key blocking strategy is used to reduce computational overhead. In the structural reasoning stage, entity nodes, contextual relationship edges, and ontology constraint edges are modeled uniformly, and cross-document consistency evidence is mined through relationship-aware graph reasoning. In the decision-making stage, a dual-threshold grading mechanism and risk adjustment are introduced to achieve hierarchical output of automatic confirmation, manual review, and automatic rejection. In the incremental update stage, local re-reasoning is performed on affected subgraphs based on change triggering functions, and review samples are fed back for training to continuously optimize the model.
[0093] This invention integrates ontology constraints and graph reasoning to achieve global consistency judgment across document entities, significantly reducing misalignment and missed alignment. It employs a threshold-based hierarchical structure and a manual review and feedback mechanism, improving alignment reliability and interpretability in high-risk industrial scenarios. Based on an event-triggered local incremental update mechanism, it reduces the overhead of full graph retraining and enhances online response capabilities. The core improvement lies in the organic integration of ontology constraints, structural reasoning, and incremental closed-loop processing, enabling cross-document entity alignment results to simultaneously possess high accuracy, interpretability, and sustainable evolution capabilities. This meets the requirements of continuous evolution, risk control, and large-scale application in industrial scenarios, and is suitable for equipment ledger management, process knowledge fusion, fault knowledge integration, and operation and maintenance decision support.
[0094] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.
[0095] In this embodiment of the invention, a cross-document entity alignment system that integrates domain ontology constraints and graph reasoning is provided, comprising: The data acquisition unit is used to acquire multi-source industrial document data to be processed; The entity extraction and representation unit is used to extract entities from the multi-source industrial document data based on a pre-built domain ontology model and a relation constraint rule library, generating an entity set and an entity co-occurrence graph. Each entity in the entity set is associated with an entity type, a context vector representation, an attribute vector representation, and an ontology concept representation. In the entity co-occurrence graph, intra-sentence dependency relations and inter-sentence co-occurrence relations are used as document context relation edges, and constraint relations in the relation constraint rule library are used as ontology constraint edges. The candidate generation and matching unit is used to generate multiple candidate entity pairs based on the entity set and remove candidate entity pairs that are incompatible in type; for each retained candidate entity pair, semantic similarity, attribute consistency, ontology semantic distance, initial structure consistency and string similarity are calculated based on context vector representation, attribute vector representation, ontology concept representation, entity co-occurrence graph and entity name, respectively; after multi-feature weighted fusion and probability normalization, a candidate entity pair set and the matching probability of each candidate entity pair are generated. The reasoning graph construction and structure scoring unit is used to take candidate entity pairs in the candidate entity pair set as alignment edges to be verified, and to use entities as nodes to perform unified modeling by integrating document context relationship edges and ontology constraint edges to obtain a cross-document reasoning graph. On the cross-document reasoning graph, the corresponding attention weights are calculated for document context relationship edges and ontology constraint edges respectively. The neighboring information is weighted and summed through relation-aware aggregation to update the entity node representation. The structure consistency score of each candidate entity pair is calculated based on the updated entity node representation. The constraint optimization and confidence output unit is used to calculate the consistency loss based on entity node representation and structural consistency score, using matching probability, introducing ontology constraint loss and consistency loss, and constructing a joint objective function by combining the alignment loss based on marginal ranking and regularization term. After global constraint optimization and confidence calibration, the alignment confidence of each candidate entity pair is output. The hierarchical decision unit is used to compare the alignment confidence of each candidate entity pair with the pre-acquired high and low thresholds respectively: if the alignment confidence is greater than or equal to the high threshold, the corresponding candidate entity pair is determined to be aligned; if the alignment confidence is less than the high threshold but greater than or equal to the low threshold, the corresponding candidate entity pair is transferred to manual review; if the alignment confidence is less than the low threshold, the corresponding candidate entity pair is determined to be unaligned and excluded.
[0096] This invention discloses an industrial internet cross-document entity alignment system that integrates ontology constraints and graph reasoning. The system achieves accurate alignment, interpretable verification, and continuous evolution of entities in multi-source industrial documents through industrial domain ontology modeling, candidate entity pair generation, cross-document reasoning graph construction, relation-aware graph reasoning, dual-threshold hierarchical decision-making, and event-triggered local incremental updates. It reduces the risk of misalignment and improves the consistency and maintainability of industrial knowledge graphs. It is applicable to scenarios such as equipment governance, process knowledge fusion, fault knowledge integration, and operation and maintenance decision support.
[0097] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from a computer storage medium to implement a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to execute operations of a cross-document entity alignment method that integrates domain ontology constraints and graph reasoning.
[0098] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the cross-document entity alignment method related to the fusion of domain ontology constraints and graph reasoning in the above embodiments.
[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.
[0100] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A cross-document entity alignment method integrating domain ontology constraints and graph reasoning, characterized in that, include: Step 1: Obtain the multi-source industrial document data to be processed; Step 2: Based on the pre-built domain ontology model and relation constraint rule library, entity extraction is performed on the multi-source industrial document data to generate an entity set and an entity co-occurrence graph. Each entity in the entity set is associated with an entity type, a context vector representation, an attribute vector representation, and an ontology concept representation. In the entity co-occurrence graph, intra-sentence dependency relations and inter-sentence co-occurrence relations are used as document context relation edges, and constraint relations in the relation constraint rule library are used as ontology constraint edges. Step 3: Generate multiple candidate entity pairs based on the entity set, and remove candidate entity pairs that are incompatible in type; for each remaining candidate entity pair, calculate semantic similarity, attribute consistency, ontology semantic distance, initial structure consistency, and string similarity based on context vector representation, attribute vector representation, ontology concept representation, entity co-occurrence graph, and entity name, respectively. After multi-feature weighted fusion and probability normalization, generate a set of candidate entity pairs and the matching probability of each candidate entity pair. Step 4: Using candidate entity pairs in the candidate entity pair set as alignment edges to be verified, and using entities as nodes, integrate document context relationship edges and ontology constraint edges for unified modeling to obtain a cross-document reasoning graph; on the cross-document reasoning graph, calculate the corresponding attention weights for document context relationship edges and ontology constraint edges respectively, and update the entity node representation by weighted summation of each neighborhood information through relation-aware aggregation, and calculate the structural consistency score of each candidate entity pair based on the updated entity node representation; Step 5: Based on entity node representation and structural consistency score, calculate consistency loss using matching probability, introduce ontology constraint loss and consistency loss, combine alignment loss based on marginal ranking and regularization term to construct joint objective function, and output alignment confidence of each candidate entity pair after global constraint optimization and confidence calibration. Step 6: Compare the alignment confidence of each candidate entity pair with the pre-acquired high and low thresholds respectively: if the alignment confidence is greater than or equal to the high threshold, the corresponding candidate entity pair is determined to be aligned; if the alignment confidence is less than the high threshold but greater than or equal to the low threshold, the corresponding candidate entity pair is transferred to manual review; if the alignment confidence is less than the low threshold, the corresponding candidate entity pair is determined to be unaligned and excluded.
2. The cross-document entity alignment method for fusing domain ontology constraints and graph reasoning according to claim 1, characterized in that, In step 4, the expression for calculating the attention weights is: ; In the formula, For nodes i With adjacent nodes j In relation r Attention weights; For relationship r The attention parameter vector; For relationship r The learnable weight matrix below; , , They are respectively i , j , The current representation vector; For nodes i The neighborhood set under relation r; for Node index in; This represents a vector concatenation operation; This indicates the transpose operation.
3. The cross-document entity alignment method for fusing domain ontology constraints and graph reasoning according to claim 2, characterized in that, In step 4, the update expression represented by the entity node is: ; In the formula, For nodes i In the Layer representation; For activation functions; A set of relation types; For the first Hierarchical Relationship r The learnable weight matrix below; For adjacent node j in the th... Layer representation.
4. The cross-document entity alignment method for fusing domain ontology constraints and graph reasoning according to claim 1, characterized in that, In step 5, the joint objective function for: ; ; ; In the formula, Alignment loss based on marginal ranking; This is the loss due to ontology constraints; This results in a loss of consistency. It is a regularization term; , , These are the weighting coefficients for ontology constraint loss, consistency loss, and regularization term; It is a relational mapping matrix; , These are the representation vectors for the head entity and the tail entity, respectively; It is a relation vector; For labeling or high-confidence pseudo-labels; For candidate entity pairs The matching probability.
5. The cross-document entity alignment method for fusing domain ontology constraints and graph reasoning according to claim 1, characterized in that, Step 3, the specific steps for generating multiple candidate entity pairs based on the entity set include: generating candidate entity pairs using a multi-key blocking strategy; wherein the blocking key includes at least one of the following: encoding prefix, region label, process segment number, and time window.
6. The cross-document entity alignment method for fusing domain ontology constraints and graph reasoning according to claim 1, characterized in that, In step 6, the high threshold and low threshold Determine as follows: ; ; In the formula, Basic threshold; For risk factors, the range of values is: ; This is the risk adjustment coefficient.
7. The cross-document entity alignment method for fusing domain ontology constraints and graph reasoning according to claim 1, characterized in that, Also includes: Step 7: In response to a new document or a change in the relation constraint rule base, obtain the cross-document reasoning graph constructed in Step 4 and the review result output in Step 6; when the triggering source is a new document, determine the affected entities based on the association between the entities in the new document and the existing entities in the cross-document reasoning graph; when the triggering source is a change in the relation constraint rule base, determine the affected entities in the cross-document reasoning graph based on the updated constraint conditions. Trigger scores are calculated based on the magnitude of attribute changes, relational topology changes, and business risk weighting of the affected entities. When the obtained trigger score exceeds a preset threshold, local re-inference is performed on the k-hop neighborhood of the affected entities on the cross-document inference graph, and the corresponding index and version identifier are updated. If the update result violates ontology constraints or the quality indicators are abnormal, the system is rolled back to the most recent stable version. The verification results are used as new training samples for subsequent use when the score exceeds the threshold to perform local re-inference.
8. A cross-document entity alignment system integrating domain ontology constraints and graph reasoning, characterized in that, include: The data acquisition unit is used to acquire multi-source industrial document data to be processed; The entity extraction and representation unit is used to extract entities from the multi-source industrial document data based on a pre-built domain ontology model and a relation constraint rule library, generating an entity set and an entity co-occurrence graph. Each entity in the entity set is associated with an entity type, a context vector representation, an attribute vector representation, and an ontology concept representation. In the entity co-occurrence graph, intra-sentence dependency relations and inter-sentence co-occurrence relations are used as document context relation edges, and constraint relations in the relation constraint rule library are used as ontology constraint edges. The candidate generation and matching unit is used to generate multiple candidate entity pairs based on the entity set and remove candidate entity pairs that are incompatible in type; for each retained candidate entity pair, semantic similarity, attribute consistency, ontology semantic distance, initial structure consistency and string similarity are calculated based on context vector representation, attribute vector representation, ontology concept representation, entity co-occurrence graph and entity name, respectively; after multi-feature weighted fusion and probability normalization, a candidate entity pair set and the matching probability of each candidate entity pair are generated. The reasoning graph construction and structure scoring unit is used to take candidate entity pairs in the candidate entity pair set as alignment edges to be verified, and to use entities as nodes to perform unified modeling by integrating document context relationship edges and ontology constraint edges to obtain a cross-document reasoning graph. On the cross-document reasoning graph, the corresponding attention weights are calculated for document context relationship edges and ontology constraint edges respectively. The neighboring information is weighted and summed through relation-aware aggregation to update the entity node representation. The structure consistency score of each candidate entity pair is calculated based on the updated entity node representation. The constraint optimization and confidence output unit is used to calculate the consistency loss based on entity node representation and structural consistency score, using matching probability, introducing ontology constraint loss and consistency loss, and constructing a joint objective function by combining the alignment loss based on marginal ranking and regularization term. After global constraint optimization and confidence calibration, the alignment confidence of each candidate entity pair is output. The hierarchical decision unit is used to compare the alignment confidence of each candidate entity pair with the pre-acquired high and low thresholds respectively: if the alignment confidence is greater than or equal to the high threshold, the corresponding candidate entity pair is determined to be aligned; if the alignment confidence is less than the high threshold but greater than or equal to the low threshold, the corresponding candidate entity pair is transferred to manual review; if the alignment confidence is less than the low threshold, the corresponding candidate entity pair is determined to be unaligned and excluded.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the cross-document entity alignment method for fusing domain ontology constraints and graph reasoning as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the cross-document entity alignment method for fusing domain ontology constraints and graph reasoning as described in any one of claims 1 to 7.
