Controllable graph attention and candidate semantic collaborative driven cross-graph entity alignment method

Abstract

This invention discloses a cross-graph entity alignment method driven by controllable graph attention and candidate semantics, comprising the following steps: Step 1, construction of multi-view controllable entity representation; Step 2, generation of candidate entities with high recall; Step 3, selection of candidate set semantics; Step 4, confidence calibration and rejection. In the candidate generation stage, this invention effectively suppresses the propagation and accumulation of heterogeneous structural noise through multi-view controllable entity representation and a dual-path controllable graph attention mechanism, significantly improving the discriminativeness and scene adaptability of entity representation. By introducing a Soft-InfoNCE soft contrast learning objective in the candidate recall optimization process, the recall rate of truly matching entities is significantly improved. In the fine-ranking stage, a candidate set joint semantic discrimination mechanism overcomes the limitations of independent matching for each candidate, reducing the risk of mismatch. A dual-threshold rejection mechanism effectively identifies suspended entities, avoiding misalignment problems in open scenes.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a cross-graph entity alignment method driven by controllable graph attention and candidate semantics. Background Technology

[0002] Knowledge graphs, as a core technology for structurally representing real-world entities and their relationships, have wide-ranging application value in many fields such as knowledge fusion, intelligent question answering, semantic retrieval, and decision support. However, due to the inherent differences in modeling patterns, naming rules, relational systems, and knowledge coverage among different data sources, there are a large number of synonymous entities pointing to the same real-world object but with different forms of expression in different knowledge graphs. Cross-knowledge graph entity alignment, as a core link in identifying such synonymous entities and realizing the unified organization and fusion of multi-source heterogeneous knowledge, is an important foundation for the practical application of knowledge graph technology.

[0003] However, in practical applications, the knowledge graphs to be aligned often exhibit significant heterogeneity. Different graphs differ greatly in entity size, relation type, structural distribution, and temporal representation. Existing entity alignment technologies have the following shortcomings when dealing with such highly heterogeneous scenarios: 1. For entity alignment tasks, entities typically contain three key clues: name semantic information, local relational structure information, and temporal context information. However, in highly heterogeneous scenarios, existing methods struggle to achieve stable collaborative modeling of these three key clues for entity alignment. In particular, entity alignment methods based on graph structure propagation generally assume that neighborhood structure information can directly enhance entity representation. In highly heterogeneous scenarios, when faced with issues such as missing cross-graph relations, topological mismatch, and structural noise, inconsistent neighborhood relations and invalid structural information are synchronously introduced and continuously accumulated during graph propagation, severely weakening the discriminative ability of entity representation and directly affecting the quality of candidate recall and the reliability of subsequent matching results. Furthermore, these methods also struggle to address issues such as name aliases, abbreviations, multilingual transliterations, differences in description granularity, and inconsistent temporal information annotation granularity, incomplete scope, and inconsistent expression forms. 1. To achieve complementary and stable fusion of multi-dimensional information; 2. Existing fine ranking methods based on pre-trained language models and large language models mostly adopt a candidate-by-candidate independent matching approach, performing semantic matching and relevance judgment on each candidate entity separately. They lack the ability to jointly model the relative differences between entities within the candidate set, which not only leads to a large amount of repetitive computation and unnecessary consumption of model resources, but also makes it impossible for the model to use the comparative information between candidates to optimize the ranking results. This easily leads to problems such as unstable ranking and mismatch of highly similar interference items, and makes it difficult to accurately identify aliases and abbreviations that are similar to entities without surface word overlap; 3. Existing technologies generally lack effective rejection mechanisms for open knowledge fusion scenarios. In real applications, not all source entities can find corresponding entities that can be aligned in the target knowledge graph. These dangling entities without real matching objects are often incorrectly linked to semantically similar but not real corresponding candidate entities in the processing flow of existing methods due to the logic of the system forcibly outputting matching results, when there is insufficient matching evidence. This introduces a large amount of incorrectly aligned data, seriously polluting the quality of the fused knowledge base and affecting the effect of subsequent downstream applications. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-graph entity alignment method driven by controllable graph attention and candidate semantics, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a cross-graph entity alignment method driven by controllable graph attention and candidate semantics, comprising the following steps: Step 1, construction of multi-view controllable entity representation; Step 2, generation of high-recall candidate entities; Step 3, selection of candidate set semantics; Step 4, confidence calibration and rejection.

[0006] In step one above, multi-dimensional features are extracted from entities in the source knowledge graph and the target knowledge graph respectively, and adaptive fusion is performed to generate a unified entity vector representation across the graph.

[0007] In step two above, similarity calculation based on unified entity vector representation generates a controlled set of candidate entities for source entity retrieval in the target knowledge graph.

[0008] In step three above, the source entity and candidate entity sets are organized into a unified input sequence, and the joint matching, scoring and ranking of candidate entities are completed through a large language model to output the optimal candidate entity.

[0009] In step four above, the matching reliability is determined based on the selection confidence of the best candidate and the marginal difference between the second-best candidate. If the evidence is insufficient, a no-match result is output.

[0010] In step one, the multi-dimensional features include name semantic representation extracted from the perspective of name semantics, structural representation extracted from the perspective of relational structure, and temporal representation extracted from the perspective of temporal context; specifically, semantic information is extracted from the entity's name, alias, title, and text description, and name semantic representation is generated by mapping through a pre-trained text encoding model. Based on the local neighborhood structure of entities in the knowledge graph, a controllable graph attention network is used to model structural features and generate structural representations. For knowledge graph facts with time tags associated with entities, complete the encoding and aggregation of time information to generate a time representation. .

[0011] The controllable graph attention network explicitly distinguishes between two independent paths during message passing: the preservation of a node's own semantics and the propagation of neighborhood relationships. Specifically:

[0012] Let entity The set of neighboring nodes is For any neighboring node ,entity With neighboring nodes The structural propagation weights between them are denoted as:

[0013]

[0014] in, Representing entities With neighboring nodes Relationship types Represents a relation-aware attention scoring function. , Entities With neighboring nodes The initial feature vector;

[0015] Define independent control coefficients for the semantic reservation terms of the node itself. :

[0016]

[0017] in, Indicates the strength of the retention of the entity's own information. This represents the self-loop information control function; based on this, the entity The update from a structural perspective is represented as:

[0018]

[0019] in, and These represent the transformation matrices of the semantic path itself and the neighborhood propagation path, respectively. This represents a non-linear activation function.

[0020] The time representation extracted from the perspective of time context specifically refers to: assuming the entity... The relevant time set is The time information is encoded and aggregated to obtain a time representation:

[0021]

[0022]

[0023] in, This represents the time encoding function. and For learnable time-encoded parameters, For encoding dimensions, This represents an aggregate function.

[0024] In step one, the adaptive fusion adopts a multi-view adaptive weighted fusion mechanism, and the formula is:

[0025]

[0026] in, For entities The final unified vector representation, , , These are learnable weight coefficients used to dynamically adjust the contribution of different perspectives to the final representation of the entity.

[0027] In step one, after generating a unified entity vector representation across the graph, the Soft-InfoNCE soft contrastive learning objective is introduced to optimize the entity representation space, specifically as follows:

[0028] Let the source entity be represented as The corresponding positive sample is represented as The candidate negative sample set is The Soft-InfoNCE loss is defined as:

[0029]

[0030] in, Represents the similarity function. Indicates the temperature coefficient. Indicates negative samples The soft weights are as follows:

[0031]

[0032] in, For source entity With negative sample entities Semantic similarity between them.

[0033] In step two, the formula used for similarity calculation based on unified entity vector representation is:

[0034]

[0035] in, and Representing the source entity respectively and target entity The final vector representation, This represents the cosine similarity function.

[0036] In step two, generating a controlled-size candidate entity set for the source entity retrieval in the target knowledge graph specifically involves: sorting the entities in the target knowledge graph in descending order based on their similarity scores, and selecting the top-scoring entities. Each entity constitutes a candidate entity set for the source entity.

[0037] Step three specifically includes the following steps:

[0038] 3.1 Construct structured semantic representations for the source entity and its candidate entities respectively. For any entity... Its structured semantic representation is as follows:

[0039]

[0040] in, Indicates the semantic features of the name. Indicates descriptive information features, Indicates structural features, Represents temporal context features;

[0041] 3.2 Source Entity and its candidate entity set The structured semantic representation is organized into a unified input sequence:

[0042]

[0043] in, to For the candidate set One candidate entity;

[0044] 3.3 The unified input sequence is fed into the large language model, which performs joint relevance evaluation on all candidate entities within the same context, outputs the matching score of each candidate entity and sorts them, and selects the candidate entity with the highest score as the optimal candidate entity. The determination formula is as follows:

[0045]

[0046] in, The optimal candidate entity output by the model. The target entity output by the model With source entity The matching probability.

[0047] In step four, specifically: let the optimal candidate matching probability output by the large language model be... The probability of the second-best candidate match is Define the source entity The selection confidence level is The marginal difference is defined as Set confidence threshold and marginal difference threshold When satisfied , If any of the following conditions are met, a rejection is triggered. When a rejection is triggered, the system outputs the result that there is no matching entity, and determines that the current source entity is a dangling entity; otherwise, the optimal candidate entity is output as the final alignment result.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows: In the candidate generation stage, the present invention effectively suppresses the propagation and accumulation of heterogeneous structural noise through multi-view controllable entity representation and dual-path controllable graph attention mechanism, and significantly improves the discriminativeness and scene adaptability of entity representation; by introducing Soft-InfoNCE soft contrast learning objective in the candidate recall optimization process, the recall rate of real matching entities is significantly improved; in the fine ranking stage, the limitation of independent matching of each candidate is overcome by the joint semantic discrimination mechanism of candidate set, and the risk of mismatch is reduced; and the double threshold rejection mechanism effectively identifies dangling entities and avoids the problem of misalignment in open scenes. Attached Figure Description

[0049] Figure 1 This is a flowchart of the steps of the present invention;

[0050] Figure 2 This is a flowchart of the method of the present invention;

[0051] Figure 3 This is a flowchart illustrating the construction process of the multi-view controllable entity representation of the present invention.

[0052] Figure 4 This is a schematic diagram of the candidate set semantic selection and rejection mechanism of the present invention;

[0053] Figure 5 This is a schematic diagram of the entity to be retrieved and its local graph structure information in Experiment Example 1. Detailed Implementation

[0054] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see the appendix Figure 1 - Appendix Figure 4 The present invention provides an embodiment of a cross-graph entity alignment method driven by controllable graph attention and candidate semantics, comprising the following steps: Step 1, construction of multi-view controllable entity representation; Step 2, generation of high-recall candidate entities; Step 3, selection of candidate set semantics; Step 4, confidence calibration and rejection.

[0056] In step one above, multi-dimensional features are extracted from entities in the source and target knowledge graphs and adaptively fused to generate a unified entity vector representation across the graphs. Soft-InfoNCE (soft contrastive learning) is introduced to optimize the entity representation space. The multi-dimensional features include name semantic representations extracted from a name semantic perspective, structural representations extracted from a relational structure perspective, and temporal representations extracted from a temporal context perspective. Specifically, semantic information is extracted from the entity's name, alias, title, and text description, and name semantic representations are generated through mapping using a pre-trained text encoding model. Based on the local neighborhood structure of entities in the knowledge graph, a controllable graph attention network is used to model structural features and generate structural representations. For knowledge graph facts with time tags associated with entities, complete the encoding and aggregation of time information to generate a time representation. Among them, the controllable graph attention network explicitly distinguishes two independent paths during message passing: the preservation of the node's own semantics and the propagation of neighborhood relations. Specifically:

[0057] Let entity The set of neighboring nodes is For any neighboring node ,entity With neighboring nodes The structural propagation weights between them are denoted as:

[0058]

[0059] in, Representing entities With neighboring nodes Relationship types Represents a relation-aware attention scoring function. , Entities With neighboring nodes The initial feature vector;

[0060] Define independent control coefficients for the semantic reservation terms of the node itself. :

[0061]

[0062] in, Indicates the strength of the retention of the entity's own information. This represents the self-loop information control function; based on this, the entity The update from a structural perspective is represented as:

[0063]

[0064] in, and These represent the transformation matrices of the semantic path itself and the neighborhood propagation path, respectively. Represents a non-linear activation function;

[0065] The time representation extracted from the perspective of time context is specifically: Let's assume the entity... The relevant time set is The time information is encoded and aggregated to obtain a time representation:

[0066]

[0067]

[0068] in, This represents the time encoding function. and For learnable time-encoded parameters, For encoding dimensions, Represents aggregate functions;

[0069] Adaptive fusion employs a multi-view adaptive weighted fusion mechanism, with the following formula:

[0070]

[0071] in, For entities The final unified vector representation, , , These are learnable weight coefficients used to dynamically adjust the contribution of different perspectives to the final representation of the entity;

[0072] The soft-InfoNCE learning objective is introduced to optimize the entity representation space as follows:

[0073] Let the source entity be represented as The corresponding positive sample is represented as The candidate negative sample set is The Soft-InfoNCE loss is defined as:

[0074]

[0075] in, Represents the similarity function. Indicates the temperature coefficient. Indicates negative samples The soft weights are as follows:

[0076]

[0077] in, For source entity With negative sample entities Semantic similarity between them;

[0078] In step two above, based on the similarity calculation using unified entity vector representation, entities in the target knowledge graph are sorted in descending order according to their similarity scores, and the entities with the highest scores are selected. There are 12 entities, which constitute the candidate entity set of the source entity; the similarity calculation based on the unified entity vector representation uses the following formula:

[0079]

[0080] in, and Representing the source entity respectively and target entity The final vector representation, Represents the cosine similarity function;

[0081] In step three above, the source entities and candidate entity sets are organized into a unified input sequence. A large language model is then used to perform joint matching, scoring, and ranking of the candidate entities, outputting the optimal candidate entity. Specifically, this includes the following steps:

[0082] 3.1 Construct structured semantic representations for the source entity and its candidate entities respectively. For any entity... Its structured semantic representation is as follows:

[0083]

[0084] in, Indicates the semantic features of the name. Indicates descriptive information features, Indicates structural features, Represents temporal context features;

[0085] 3.2 Source Entity and its candidate entity set The structured semantic representation is organized into a unified input sequence:

[0086]

[0087] in, to For the candidate set One candidate entity;

[0088] 3.3 The unified input sequence is fed into the large language model, which performs joint relevance evaluation on all candidate entities within the same context, outputs the matching score of each candidate entity and sorts them, and selects the candidate entity with the highest score as the optimal candidate entity. The determination formula is as follows:

[0089]

[0090] in, The optimal candidate entity output by the model. The target entity output by the model With source entity The matching probability;

[0091] In step four above, the matching reliability is determined based on the marginal difference between the selection confidence of the optimal candidate and the second-best candidate, and no matching result is output when the evidence is insufficient; specifically: let the optimal candidate matching probability output by the large language model be... The probability of the second-best candidate match is Define the source entity The selection confidence level is The marginal difference is defined as Set confidence threshold and marginal difference threshold When satisfied , If any of the following conditions are met, a rejection is triggered. When a rejection is triggered, the system outputs the result that there is no matching entity, and determines that the current source entity is a dangling entity; otherwise, the optimal candidate entity is output as the final alignment result.

[0092] Experimental Example 1:

[0093] For heterogeneous cross-domain knowledge graph fusion scenarios, the source knowledge graph is the ICEWS event knowledge graph, and the target knowledge graph is the WIKI encyclopedia knowledge graph. Using the method proposed in the above embodiments, the synonyms of the entity "Coalition of the Radical Left" to be aligned in the source graph are identified in the target graph, completing cross-graph entity alignment. The invention's resistance to alias noise and interference from similar structures, as well as its alignment accuracy, are verified, as shown in the appendix. Figure 5As shown, the specific implementation process is as follows: First, based on the source entity Coalition of the Radical Left and its name text, local neighborhood relationship triples, and associated event time information in the source graph; and the corresponding name, structure, and time dimension information of all entities in the target graph; through multi-dimensional feature extraction, multi-view adaptive fusion, and Soft-InfoNCE optimization, a unified cross-graph vector representation of all entities in the source and target graphs is obtained; then, based on the unified cross-graph vector representation of all entities in the source and target graphs, the cosine similarity function is used to calculate the vector similarity between the source entity and each entity in the target graph. All entities in the target graph are sorted in descending order of similarity score, and the Top-10 entities are selected to form a candidate set to obtain the source entity Coalition of the Radical Left. The Top-10 candidate entity set for the Left is shown in Table 1. Next, based on the Top-10 candidate entity set, and the complete information of the source entity and each candidate entity's name, description, relational structure, and temporal context, candidate compression is performed: combining recall scores and name semantic cues, marginalizing low-relevance candidates with type deviations or inconsistent themes, and retaining core high-potential candidates. Then, a shortcut determination is performed: fast rules such as exact matching, prefix matching, and subsequence matching are executed. Since there is no explicit surface name correspondence between the source entity and Syriza, the shortcut module does not hit. Next, a structured semantic representation is constructed: for the source entity and each candidate entity, a structured semantic representation containing name semantic features, descriptive information features, structural features, and temporal context features is constructed. Finally, joint matching scoring and ranking are performed: the structured representations of the source entity and all candidate entities are organized into a unified input sequence, input into the large language model, and matching scoring is performed sequentially according to name cues, entity type, summary description, and local relational features. The candidate entity ranking results are obtained: the best candidate entity is Syriza, with a matching score of 0.96, and the second best candidate is Radical Party of the... The left matching score was 0.72; finally, based on the candidate entity matching probability ranking results, a pre-set reliability threshold was determined. Marginal difference threshold The threshold determination results in the following output: the final aligned entity is Syriza. This result shows that the final aligned entity is completely consistent with the actual corresponding entity, verifying the effectiveness of the present invention.

[0094] Experimental Example 2:

[0095] To verify the effectiveness of this invention, several representative entity alignment models were selected as comparison methods, including embedding learning-based methods (MTransE, BootEA), graph neural network-based methods (GCN-Align, RDGCN, Dual-AMN, TEA-GNN), pre-trained language model-based methods (BERT, BERT-INT, FuAlign, Simple-HHEA), and pre-trained large language model-based methods (ChatEA, ProLEA). Performance tests were conducted on the ICEWS-WIKI dataset, and the results are shown in Table 2. The test results show that on the highly heterogeneous ICEWS-WIKI dataset, the accuracy of the 10-candidate alignment achieved by this invention reaches 0.972, exceeding most existing mainstream methods and providing a very high upper limit for the fine-tuning process. The final alignment accuracy after the entire fine-tuning process reaches 0.969, exceeding all existing comparison methods. Simultaneously, it effectively rejects suspended entities, reducing the misalignment rate in open scenarios by more than 40%, verifying the effectiveness and robustness of the method in highly heterogeneous knowledge graph scenarios.

[0096] Table 1 Top-10 candidate entity set

[0097]

[0098] Table 2 Performance comparison of different entity alignment methods on the ICEWS-WIKI dataset

[0099]

[0100] Based on the above, the advantages of this invention are as follows: When used, by constructing a multi-view controllable entity representation mechanism that integrates name semantics, relational structure, and temporal context, and combining it with a dual-path controllable graph attention network that preserves node semantics and propagates neighborhood relations, it can adaptively suppress the propagation of invalid structural information and stably preserve the highly discriminative semantic features of entity ontology in heterogeneous scenarios with highly asymmetrical structures and strong neighborhood noise. This fundamentally solves the shortcomings of traditional graph neural network methods, such as susceptibility to cross-graph structural differences and insufficient entity representation discriminative power. By introducing the Soft-InfoNCE soft contrast learning objective in the representation learning stage, it can assign differentiated weights based on the potential correlation between negative samples and source entities, effectively reducing... Excessive rejection of potential matching entities significantly improves the recall rate of real matching entities and the quality of the candidate set. The joint semantic selection mechanism for the candidate set overcomes the technical limitations of independent matching for each candidate, organizing the source entity and all candidates into a unified input sequence. With the help of a large language model, joint comparison and discrimination are completed in the same context, fully mining the relative differences between candidate entities, greatly reducing the risk of ranking instability and mismatch, and accurately identifying synonymous entities with no surface word overlap such as aliases and abbreviations. The dual threshold rejection mechanism of confidence calibration and marginal difference can effectively identify and reject dangling entities when matching evidence is insufficient, completely avoiding the misalignment problem caused by forced matching in open knowledge fusion scenarios.

[0101] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A controllable graph attention and candidate semantic co-driven cross-graph entity alignment method, comprising the following steps: Step 1: Construction of multi-view controllable entity representation; Step 2: Generation of high-recall candidate entities; Step 3: Semantic selection of candidate sets; Step 4: Confidence calibration and rejection; Its characteristics are: In step one above, multi-dimensional features are extracted from entities in the source knowledge graph and the target knowledge graph respectively, and adaptive fusion is performed to generate a unified entity vector representation across the graph. In step two above, similarity calculation based on unified entity vector representation generates a controlled set of candidate entities for source entity retrieval in the target knowledge graph. In step three above, the source entity and candidate entity sets are organized into a unified input sequence, and the joint matching, scoring and ranking of candidate entities are completed through a large language model to output the optimal candidate entity. In step four above, the matching reliability is determined based on the selection confidence of the best candidate and the marginal difference between the second-best candidate. If the evidence is insufficient, a no-match result is output.

2. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: The multi-dimensional features include name semantic representation extracted from a name semantic perspective, structure representation extracted from a relationship structure perspective, and time representation extracted from a time context perspective. ; Based on the local neighborhood structure of entities in the knowledge graph, the structure feature modeling is completed through the controllable graph attention network to generate the structure representation ; For the time-labeled knowledge graph facts associated with the entity, time information encoding and aggregation are completed to generate the time representation .

3. The controllable graph attention and candidate semantic co-driven cross-graph entity alignment method according to claim 2, characterized in that: The controllable graph attention network explicitly distinguishes between two independent paths during message passing: the preservation of a node's own semantics and the propagation of neighborhood relationships. Specifically: Let entity The set of neighboring nodes is For any neighboring node ,entity With neighboring nodes The structural propagation weights between them are denoted as: ， in, Representing entities With neighboring nodes Relationship types Represents a relation-aware attention scoring function. , Entities With neighboring nodes The initial feature vector; Define independent control coefficients for the semantic reservation terms of the node itself. : ， in, Indicates the strength of the retention of the entity's own information. This represents the self-loop information control function; based on this, the entity The update from a structural perspective is represented as: ， in, and These represent the transformation matrices of the semantic path itself and the neighborhood propagation path, respectively. This represents a non-linear activation function.

4. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 2, characterized in that: The time representation extracted from the perspective of time context specifically refers to: assuming the entity... The relevant time set is The time information is encoded and aggregated to obtain a time representation: ， ， in, This represents the time encoding function. and For learnable time-encoded parameters, For encoding dimensions, This represents an aggregate function.

5. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: In step one, the adaptive fusion adopts a multi-view adaptive weighted fusion mechanism, and the formula is: ， in, For entities The final unified vector representation, , , These are learnable weight coefficients used to dynamically adjust the contribution of different perspectives to the final representation of the entity.

6. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: In step one, after generating a unified entity vector representation across the graph, the Soft-InfoNCE soft contrastive learning objective is introduced to optimize the entity representation space, specifically as follows: Let the source entity be represented as The corresponding positive sample is represented as The candidate negative sample set is The Soft-InfoNCE loss is defined as: ， in, Represents the similarity function. Indicates the temperature coefficient. Indicates negative samples The soft weights are as follows: ， in, For source entity With negative sample entities Semantic similarity between them.

7. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: In step two, the formula used for similarity calculation based on unified entity vector representation is: ， in, and Representing the source entity respectively and target entity The final vector representation, This represents the cosine similarity function.

8. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: In step two, generating a controlled-size candidate entity set for the source entity retrieval in the target knowledge graph specifically involves: sorting the entities in the target knowledge graph in descending order based on their similarity scores, and selecting the top-scoring entities. Each entity constitutes a candidate entity set for the source entity.

9. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: Step three specifically includes the following steps: Step 3.1 Constructs structured semantic representations for the source entity and its candidate entities respectively. For any entity... Its structured semantic representation is as follows: ， in, Indicates the semantic features of the name. Indicates descriptive information features, Indicates structural features, Represents temporal context features; Step 3.2 will transfer the source entity and its candidate entity set The structured semantic representation is organized into a unified input sequence: ， in, to For the candidate set One candidate entity; Step 3.3 inputs the unified input sequence into the large language model. The model performs a joint relevance evaluation on all candidate entities within the same context, outputs a matching score for each candidate entity, and sorts them. The candidate entity with the highest score is selected as the optimal candidate entity. The determination formula is as follows: ， in, The optimal candidate entity output by the model. The target entity output by the model With source entity The matching probability.

10. The cross-graph entity alignment method driven by controllable graph attention and candidate semantics according to claim 1, characterized in that: In step four, specifically: let the optimal candidate matching probability output by the large language model be... The probability of the second-best candidate match is Define the source entity The selection confidence level is The marginal difference is defined as Set confidence threshold and marginal difference threshold When satisfied , If any of the following conditions are met, a rejection is triggered. When a rejection is triggered, the system outputs the result that there is no matching entity, and determines that the current source entity is a dangling entity; otherwise, the optimal candidate entity is output as the final alignment result.

Images