Multi-source heterogeneous data entity alignment method and device and medium

Abstract

The invention provides a multi-source heterogeneous data entity alignment method and device and a medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: merging two knowledge maps according to a predicate alignment algorithm to obtain a merged knowledge map; learning semantic abstract embedding and relation structure embedding of each entity in the combined knowledge graph; according to the dependency relationship of words in the attribute text of each entity, performing part-of-speech learning on attribute semantic embedding of each entity; determining a combined embedding vector of each entity according to semantic abstract embedding, relation structure embedding and attribute semantic embedding of each entity; and aligning the entities in the combined knowledge graph according to the combined embedding vectors of the entities. According to the method, the accuracy of knowledge graph entity alignment in the scene of insufficient seed quantity and sparse attribute is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device and medium for aligning multi-source heterogeneous data entities. Background Art

[0002] Knowledge graph alignment is the process of establishing semantic mappings between different knowledge graphs. The goal is to align entities, attributes, and relationships across different knowledge graphs, thereby forming a unified and comprehensive knowledge system. This process is crucial for improving the coverage and accuracy of knowledge graphs.

[0003] Existing knowledge graph entity alignment solutions often require seed alignment, but the number of seeds in real-world scenarios is often very limited. Furthermore, existing knowledge graph entity alignment solutions often struggle with sparse attribute scenarios, directly impacting the accuracy of knowledge graph entity alignment. Therefore, there is an urgent need for a knowledge graph entity alignment solution that can handle scenarios with a limited number of seeds and sparse attributes. Summary of the Invention

[0004] The present invention provides a multi-source heterogeneous data entity alignment method, device and medium to solve the defect of low alignment accuracy of knowledge graph entity alignment scheme in the prior art when the number of seeds is limited and the attributes are sparse, and to achieve high-accuracy alignment of knowledge graph entities.

[0005] The present invention provides a method for aligning multi-source heterogeneous data entities, comprising the following steps.

[0006] Merge the two knowledge graphs according to the predicate alignment algorithm to obtain a merged knowledge graph; Learn to merge the semantic summary embeddings and relational structure embeddings of each entity in the knowledge graph; Learn the attribute semantic embedding of each entity based on the dependency relationship and part of speech of the words in the attribute text of each entity; Determine the combined embedding vector of each entity based on its semantic summary embedding, relational structure embedding, and attribute semantic embedding. The entities in the merged knowledge graph are aligned according to their combined embedding vectors.

[0007] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the multi-source heterogeneous data entity alignment method as described above is implemented.

[0008] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the multi-source heterogeneous data entity alignment methods described above.

[0009] The multi-source heterogeneous data entity alignment method, device, and medium provided by the present invention improves the accuracy of alignment by learning and merging the semantic summary embedding, relational structure embedding, and attribute semantic embedding of each entity in the knowledge graph, and aligns the entities based on the combined embedding vector after integrating the semantic summary embedding, relational structure embedding, and attribute semantic embedding. This compensates for the poor alignment accuracy caused by insufficient seed quantity and sparse attributes. Furthermore, the present invention also conducts in-depth learning based on the dependency relationships and parts of speech of words in the attribute text of the entity to obtain the entity's attribute embedding vector, further effectively improving the accuracy of knowledge graph entity alignment in attribute-sparse scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0011] Figure 1 This is one of the flow charts of the multi-source heterogeneous data entity alignment method provided by an embodiment of the present invention.

[0012] Figure 2 This is the second flow chart of the multi-source heterogeneous data entity alignment method provided by an embodiment of the present invention.

[0013] Figure 3 It is a flowchart of the attribute semantic embedding learning method provided by an embodiment of the present invention.

[0014] Figure 4 This is a schematic diagram of the merging of heterogeneous knowledge graphs provided by an embodiment of the present invention.

[0015] Figure 5 It is a structural diagram of a multi-source heterogeneous data entity alignment device provided by an embodiment of the present invention.

[0016] Figure 6 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0018] The following combination Figures 1-6 The present invention describes a method, device and medium for aligning multi-source heterogeneous data entities.

[0019] As a method for representing structured knowledge, knowledge graphs are widely used in fields such as semantic search and information retrieval, and are becoming a core tool for organizing complex data and supporting intelligent applications. Entity alignment, a key task in knowledge graph fusion, aims to identify entities across different knowledge graphs that point to the same real-world object. Knowledge graphs construct multi-relationship directed graphs with entities as nodes and relationships as edges, storing real-world facts in the form of triples.

[0020] Taking the railway industry as an example, the integration of multi-source data is particularly critical. From logistics and transportation to operational optimization, how to effectively integrate these heterogeneous data and build a unified railway knowledge system has become one of the core challenges in promoting the development of intelligent railways.

[0021] Figure 1 This is one of the flow charts of the multi-source heterogeneous data entity alignment method provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 101: Merge the two knowledge graphs according to the predicate alignment algorithm to obtain a merged knowledge graph.

[0022] Step 102: Learn to merge the semantic summary embedding and relational structure embedding of each entity in the knowledge graph.

[0023] Step 103: learn the attribute semantic embedding of each entity based on the dependency relationship and part of speech of the words in the attribute text of each entity.

[0024] Step 104: Determine the combined embedding vector of each entity based on the semantic summary embedding, relation structure embedding, and attribute semantic embedding of each entity.

[0025] Step 105: Align the entities in the merged knowledge graph based on the combined embedding vectors of the entities.

[0026] It should be noted that there is no order restriction between step 102 and step 103.

[0027] In the embodiments of the present invention, alignment is performed based on the integrated vector of the semantic summary embedding, relationship structure embedding, and attribute semantic embedding of entities, improving the alignment accuracy and compensating for the problem of poor alignment accuracy caused by insufficient seed quantity and attribute sparsity. Additionally, in the learning process of the above-mentioned attribute embedding vector in the embodiments of the present invention, the dependency relationship and word nature of words in the entity attribute text are deeply learned, further improving the alignment accuracy, and thus effectively enhancing the accuracy of entity alignment in the knowledge graph under the scenarios of insufficient seed quantity and attribute sparsity.

[0028] In some embodiments of the present invention, step 101 merges two knowledge graphs according to the predicate alignment algorithm, and the specific implementation process can be as follows: To achieve the semantic consistency of the cross-system knowledge graph, it is necessary to align the relationship predicates of the source graph and the target graph

[0029] to eliminate the semantic heterogeneity across graphs. This embodiment is based on a predicate alignment framework with a hierarchical alignment strategy, and constructs a unified predicate space through semantic similarity calculation and mapping rule derivation. The specific process includes: similarity modeling, candidate alignment screening, hierarchical mapping derivation, and global identifier generation. Predicate [[ID=?]] and The semantic similarity of is calculated by weighting the cosine similarity of name embedding and the domain dictionary matching degree. First, use the pre-trained word vector model FastText to obtain the character-level embedding representation of the predicate name: Among them, and respectively represent the number of characters of name and name , and is the pre-trained embedding vector of character . Calculate the cosine similarity according to the embedding vector: At the same time, construct a railway industry dictionary , which contains standard predicate names and their multilingual equivalent terms (such as the Chinese "field of belonging" and the English "domain"). Define the dictionary matching degree: The final semantic similarity is linearly weighted and fused: Among them, is the balance coefficient, and the default setting is 0.7 to strengthen the exact matching of domain terms. It should be noted that there is an unclear symbol "?" in the original text at line 14 which is retained as is in the translation for the purpose of maintaining the original content integrity.

[0030] (2) Candidate Alignment Screening and Hierarchical Mapping Based on the semantic similarity score, potential alignment predicate pairs are screened. A dynamic threshold is set . Among them, is the average similarity of all predicate pairs, is the standard deviation. If , it is marked as a candidate alignment pair .

[0031] To handle complex predicate hierarchies (such as "primary business" and "business activities"), this embodiment adopts a hierarchical alignment strategy of parent class alignment and subclass refinement. First, the domain identifiers (such as "business", "model") in the predicate name are extracted, and the parent class similarity is calculated. Then, for the predicates aligned at the parent class level, the subclass modifiers (such as "primary" and "activity") are further matched, and the subclass similarity is calculated. Finally, the hierarchical alignment score is generated by integrating the parent class and subclass similarities: Among them, , and the semantic consistency of the parent class is ensured first. If , it is determined as a valid alignment.

[0032] (3) Global Predicate Identifier Generation The normalized splicing strategy is adopted to generate a globally unique identifier for the aligned predicate pair . First, language-specific symbols (such as Chinese parentheses) are removed and converted to lowercase. Then, the normalized name is subjected to SHA-256 hashing, and the first 8 bits are intercepted as the unique identifier: For example, the predicate "field of belonging" (Chinese) and "domain" (English) generate the identifier d3a8f2c1 after normalization. This method ensures the unambiguous mapping of cross-system predicates in a unified space and avoids semantic drift in subsequent embedding learning.

[0033] (4) Knowledge Graph Merging Based on the global predicate identifier, the triple sets of and are merged: The merged knowledge graph is , where is the set of all global predicate identifiers. Knowledge graph merging provides a structurally consistent input for multi-source combined embedding learning and significantly reduces the interference of cross-graph semantic heterogeneity on entity alignment.

[0034] See also Figure 2 ,After the knowledge graphs are merged, the alignment method of this ,embodiment captures different features of entities through three ,embedding representations: semantic summary embedding, relational structure ,embedding, and attribute semantic embedding.

[0035] There are a large number of long-tail entities in the railway knowledge graph (such as "STP wireless shunting locomotive signal and monitoring system"). These entities have sparse attribute information but significant semantic value. In some embodiments of the present invention, in order to enhance the representation ability of such entities, a multi-level embedding method that combines syntactic structure and semantic parsing is used in attribute semantic embedding learning. Figure 3 , step 103 further includes: Step 201: Perform dependency syntax analysis on the attribute value text in the attribute text, resolve inter-word dependencies, and extract part-of-speech tags.

[0036] Step 202: Assign weights to words according to part-of-speech tags.

[0037] Step 203: Concatenate the attribute name text and the attribute value text in the attribute text into a text sequence, and input the text sequence into a pre-trained Word2Vec word model to generate an initial word vector for the word.

[0038] Step 204: Calculate enhanced word vectors based on the initial word vectors, inter-word dependencies, and word weights.

[0039] Step 205: Based on the enhanced word vector, a bidirectional GRU (BiGRU) model is used in conjunction with a BERT model to learn the attribute semantic embedding of the entity.

[0040] This embodiment sets weights for words according to parts of speech, and the specific rule is: noun weight ≥ verb weight ≥ modifier weight.

[0041] Specifically, the above-mentioned multi-level embedding method that integrates syntactic structure and semantic parsing mainly includes dependency analysis of attribute text, part-of-speech weight optimization and pre-training model fine-tuning, which realizes fine-grained semantic representation and effectively improves the robustness of entity alignment in attribute sparse scenarios.

[0042] (1) Attribute text preprocessing and dependency analysis Attribute text is usually composed of "attribute name-attribute value" pairs (such as "Function: Real-time monitoring of track circuit status"). In order to capture its internal grammatical logic, the attribute value text must first be subjected to dependency syntactic analysis, inter-word dependencies resolved, and part-of-speech (POS) tags extracted. Use the Stanford CoreNLP tool to build a dependency tree and resolve grammatical relationships between words such as subject-predicate, verb-object, etc. For example, in "Monitoring track circuit status", "monitoring" as the core verb governs the two noun dependency components "track circuit" and "status". Then, create a dependency matrix ,in Expressive words and There is a dependency relationship, otherwise it is 0. Next, based on the contribution of different parts of speech in the entity alignment task, the part-of-speech weights are dynamically assigned. For example, the weight of nouns (such as "track" and "status") is set to 0.8, verbs (such as "monitoring") to 0.6, and modifiers (such as "real-time") to 0.3, in order to strengthen the semantic expression of core vocabulary.

[0043] (2) Word vector generation and feature fusion The attribute name With attribute value Splicing into text sequence , generate words through pre-trained Word2Vec word model The initial word vector Then, the enhanced word vector is further calculated based on the dependency relationship and part-of-speech weight: in, Expressive words The enhanced word vector of Represents element-wise multiplication and strengthens the semantics of important parts of speech. 、 Represents the words generated by the pre-trained Word2Vec word model and words The initial word vector of Expressive words and The dependency relationship, n Indicates the number of words, represents the normalized dependency association weight, is a learnable parameter that represents the strength of inter-word dependencies. This step helps strengthen the semantic contribution of core vocabulary (for example, the weight of "monitoring" is significantly higher than that of "real-time") while preserving the grammatical information of the dependency structure.

[0044] (3) Syntactic feature extraction and sequence modeling First, the enhanced word vector is input into the bidirectional GRU for contextual semantic modeling to capture the semantic features of the context (including word vectors, part-of-speech tags, etc.): Concatenate the forward and backward hidden states to obtain a context-aware representation .

[0045] Next, a bi-affine classifier is used to predict the dependencies between words. Generate dominant words separately and dependent words The projection representation of is: in, , is the projection dimension.

[0046] Calculation Word As a dominant word, word Dependency relations as dependent words Score: in, is a learnable parameter matrix that represents the strength of the dependency relationship, is the bias term. For each word All candidate governing words Perform normalization and calculate the probability distribution: Select the dominant word with the highest probability , construct the dependency mask matrix .in, Indicates selection As The dominant word of . Weighted aggregation of dependency-associated hidden states to generate syntactic-aware context representation : (4) Fine-tuning the pre-trained language model Will generate , aligning BERT’s input dimensions by projection : in, Syntactic features that align dimensions Concatenate with the original text sequence and input into the BERT model for deep semantic encoding: Subsequently, the global semantic information is dynamically aggregated through the self-attention mechanism to output the final attribute embedding .

[0047] (5) Loss function design The loss function is composed of dependency classification loss and masked language modeling loss It consists of two parts.

[0048] Cross entropy loss is used to optimize the dependency parsing task, measuring the difference between the predicted probability of the classification model and the true label: in, For words The true dependency word label of Expressive words The predicted probability of the true dependency-governed word.

[0049] Following BERT’s masked language modeling objective, we randomly mask some input words and calculate the correct probability of the model predicting the masked words: in, represents the set of positions of words that are randomly masked; Indicates the obscured position corresponding real words; Represents the context words other than the masked word.

[0050] The optimization direction of weighted balance grammatical parsing and semantic reconstruction is obtained to obtain the total loss function of attribute embedding : in, is the weight parameter, the default , The optimal weight combination can be determined through experiments.

[0051] In some embodiments of the present invention, step 102 learns the relational structure embedding of the entity, specifically, adopts a translation model combined with a graph convolutional neural network model to learn the structural features and neighborhood information of the entity from the relation triples of the merged knowledge graph to obtain the relational structure embedding of the entity.

[0052] The specific details of relational structure embedding learning based on the TransER-GCN joint framework are as follows: In entity alignment tasks, the relationship information between entities reflects the underlying connections between them and is crucial for capturing semantic associations. The TransER-GCN joint framework, through the joint embedding of a translation model and a graph convolutional network (GCN) model, learns the structural features and neighborhood information of entities from relation triplets, thereby enhancing the robustness of alignment.

[0053] (1) Improved translation model TransER Traditional translation models address spatial heterogeneity by projecting entities into different semantic spaces using relationship-specific matrices. However, these models suffer from two key flaws: they fail to explicitly distinguish between head and tail entity types and they lack the ability to model symmetric and antisymmetric relationships. To address this, embodiments of the present invention propose an improved translation model, TransER (Translation-based Entity-Relation Model), that optimizes the representation of entities and relationships, generating semantic representations of entities.

[0054] First, perform entity type aware mapping. , respectively, the head entity and tail entity Project it into an independent dynamic semantic space to avoid semantic confusion in symmetric relationships. First, define and The dynamic mapping matrix of the head and tail entities: in, is the identity matrix; is the entity type perception vector, corresponding to the type features of the head and tail entities respectively; is the relationship vector, which is used to model the dynamic impact of the relationship on the entity. The projected entity representation is obtained through matrix transformation: The second step is to construct a dual scoring function. and symmetric loss function The collaborative application of , solves the limitation of a single scoring function. Using the loss function It can handle antisymmetric relations, but when dealing with symmetric relations, the relation vector Tends to 0. In order to improve the processing efficiency of symmetric relationship entities, the TransER model proposes a new loss function . and is defined as follows: Among them, the constraints The overall loss function of the TransER model is is the weighted sum of the two, defined as follows: in, is the weight parameter.

[0055] (2) GCN model enhancement While the TransER model can model complex semantic relationships between entities, it cannot capture the neighborhood structure of entities. To address this, this example introduces the GCN model, which uses the entity semantic representation output by the TransER translation model as input and enhances the entity representation by aggregating features of adjacent nodes. This example describes the design and implementation of the GCN in detail.

[0056] First, define and normalize the graph structure. , is a collection of entities, is a set of relations, is a set of triples. Construct an adjacency matrix , if the entity and There is a relationship between ,but , otherwise it is 0. In order to avoid the node feature losing its own information during the propagation process, the adjacency matrix Add the identity matrix , and obtain the corrected adjacency matrix . Construct a diagonal matrix ,satisfy , characterizes the degree distribution of nodes after self-connection enhancement. The adjacency matrix is ​​symmetrically normalized to obtain the normalized matrix The adjacency weight is adjusted by the node degree to ensure that the information propagation strength of nodes with different degrees is consistent during feature propagation, thus avoiding gradient explosion or disappearance.

[0057] Secondly, design the feature propagation and update mechanism. GCN realizes the layer-by-layer propagation and fusion of features through multi-layer stacking. The input feature matrix of the layer is ,in It is The feature dimension of the layer, then The output feature matrix of the layer is calculated as follows: in, Indicates the The trainable weight matrix of the layer; Integrate the first-order neighborhood features of each node; As the activation function, nonlinear expression capability is introduced.

[0058] The third step is to ensure the consistency of features across the graphs. Target map , construct independent GCN models respectively, recorded as and . and Both use a two-layer GCN architecture: the first layer weight matrix , receiving the initial features ; Second layer weight matrix , the output relationship is enhanced Dimensional entity embedding . and Shared weights and , constraining the structural similarity of the two embedding spaces by parameter binding. After two layers of propagation, the source graph entity and target graph entity Output embedding vectors respectively By sharing weights, the model can learn common structural patterns across graphs and enhance the robustness of alignment.

[0059] Finally, calculate the loss function of GCN. Define the positive sample (aligned entity pair) triple And the negative sample triplets generated by randomly replacing the head and tail entities , and get the loss function as follows: in, represents the entity embedding output by GCN, is a boundary parameter that controls the distance threshold between positive and negative samples. Based on contrastive learning, this loss function optimizes the distinguishability of the embedding space by minimizing the distance between positive pairs and maximizing the distance between negative pairs. This loss function forces the embedding distance between positive pairs to be smaller than that between negative pairs, thereby optimizing the distinguishability of entity representations.

[0060] (3) Joint optimization strategy Combining the TransER and GCN models, defining the overall loss function for relational structure embedding for and The weighted sum of: in, , is the weight coefficient for balancing the two types of losses. Grid search verification shows that when The model achieves optimal performance on the validation set.

[0061] In some embodiments of the present invention, for learning semantic summary embedding, the specific implementation process of step 102 is as follows: Textual descriptions of entity associations (e.g., "Road Network Monitoring System: Real-time Collection of Railway Traffic Data") are crucial for understanding entity semantics. Embedding learning methods based on semantic summarization improve the robustness of entity representations through deep semantic modeling and multi-view constraint optimization.

[0062] (1) Semantic summary embedding generation The core of semantic summary embedding lies in leveraging a pretrained language model to capture contextual dependencies within text descriptions. Specifically, the entity summary is fed into the BERT model, where its multi-layer Transformer encoder generates word-level context vectors. Unlike traditional word embeddings (such as Word2Vec), BERT dynamically resolves polysemy. For example, "traffic" in a railway context may refer to "vehicle traffic density," but in a communications system it might represent "data transmission rate." This context-aware capability enables the model to accurately extract domain-specific semantic features.

[0063] To further aggregate global semantics, we perform an average pooling operation on the word vectors output by BERT. This operation takes the average of all word vectors output by BERT to generate a fixed-dimensional summary embedding vector. This method not only preserves the deep semantic information of the text, but also reduces computational complexity through dimensionality reduction.

[0064] (2) Embedding Optimization under Multiple Scoring Function Constraints While semantic summarization effectively represents the core semantics of entities, relying solely on textual descriptions can lead to two problems: First, the summary information for some entities is too brief (e.g., "Monitoring equipment: no description available"), making reliable embedding difficult. Second, the correlation between structured relationships (e.g., "affiliation" and "function") and textual semantics is not explicitly modeled. This model employs a joint optimization mechanism for multiple scoring functions to improve representation quality through the synergistic constraints of semantics and structure. The scoring function design incorporates three complementary perspectives: pure semantic constraints for the summary and hybrid constraints for summary and structure.

[0065] Represents a scoring function based on the head and tail entities of the summary: Where: and are the summary head entity and tail entity representations, for The norm is used to measure the distance between vectors. This formula enforces that the summary embeddings of the head and tail entities are translation-invariant in the relational vector space. For example, if the relation "function" exists, then the summary embedding of "monitoring system" + the relation vector of "function" must be close to the summary embedding of "traffic collection."

[0066] represents a scoring function based on the summary head entity and relation embeddings, where the summary tail entity is represented by the relation embedding. and Similar, but the representation of head and tail entities is reversed.

[0067] The above two formulas jointly optimize the summary embedding and the structured embedding (generated by TransER-GCN). For example, It requires that the summary header entity and the structured tail entity be aligned in the relational space, thereby establishing an explicit association between text semantics and business logic.

[0068] The optimization goal of semantic summary embedding is achieved by weighted summation of multiple scoring functions. is the summary-based entity representation loss function: in, , , is the adaptive weight coefficient, which dynamically adjusts the influence of each constraint based on attention.

[0069] Through the synergistic effect of three sets of scoring functions, the model is able to adaptively balance the contributions of different information sources. When the summary information is sufficient, pure semantic constraints dominate the optimization process; when the structured relationships are dense, hybrid constraints enforce the consistency of semantics and topology.

[0070] This approach is applicable not only to general scenarios but also to scenarios involving zero-shot entities (entities not present in the training set). Traditional models rely on relation triplets in the graph to generate entity embeddings. If an entity does not appear in the training set (e.g., a newly added device), valid embeddings cannot be generated. The model directly uses the summary embeddings generated by BERT to represent new entities. Even if these entities have no associated triplets, semantic similarity can still be used to align them, thus resolving the issue of traditional models failing due to the lack of structural information.

[0071] In some embodiments of the present invention, steps 104 and 105 perform entity cross-graph entity alignment based on semantic summary embedding, relational structure embedding, and attribute semantic embedding. The specific implementation process is as follows: To improve the semantic perception of entity alignment, this embodiment proposes a dynamic weighted fusion mechanism that collaboratively optimizes the multi-view embedding of the knowledge graph (semantic summary embedding, relational structure embedding, and attribute semantic embedding) to generate a more robust combined embedding representation. The core process is as follows: (1) Multi-combination embedded integration For entities , linearly weight its embedding in three different views to generate a combined embedding vector : in, Indicates the number of embedded views, , respectively represent the embedding vectors of each view; is a dynamic weight and satisfies .

[0072] (2) Dynamic weight allocation First, calculate the center vector of the multi-view embedding , reflecting the global distribution characteristics. The center vector is dynamically calculated within each training batch to avoid the distribution shift caused by global averaging.

[0073] Then, the cosine similarity between each view embedding and the center vector is calculated and the generated weights are normalized. : in, is the cosine similarity calculation function, , to prevent division by zero errors; Represents the temperature coefficient, which controls the sharpness of the weight distribution.

[0074] This strategy gives greater weight to views with higher consistency with the central vector (such as structurally stable relation embeddings) while suppressing the influence of noise (such as sparse attribute embeddings). Because the weights of the semantic summary embedding, relation structure embedding, and attribute semantic embedding are recalculated within each training batch, the weights of the semantic summary embedding, relation structure embedding, and attribute semantic embedding are dynamically adjusted during training. This generates a more robust combined embedding representation and improves the semantic perception capabilities of entity alignment.

[0075] (3) Cross-graph alignment and matching Based on the combined embedding vector , calculate the source map Entities in Target map All candidate entities in The cosine similarity of , and the target entity with the highest similarity is selected as the alignment result.

[0076] This invention can improve the accuracy and robustness of entity alignment in complex multi-source heterogeneous railway data fusion scenarios: (1) The semantic summary embedding method based on BERT effectively captures the core semantic information of entities and effectively alleviates the semantic gap problem in attribute sparse scenarios; (2) The relational structure embedding method based on the TransER-GCN joint framework takes into account both complex semantic relationship modeling and neighborhood structure feature extraction, enhancing the ability to capture semantic associations across graphs; (3) Syntactic-aware attribute embedding representation method, combined with dependency analysis and pre-trained language model, achieves fine-grained semantic representation of long-tail entities; (4) A dynamic weight allocation strategy based on cosine similarity adaptively strengthens key features, suppresses redundant noise, improves the model's adaptability to heterogeneous data distribution, and enhances the model's scenario adaptability.

[0077] The effect of the multi-source heterogeneous data knowledge graph entity alignment solution (hereinafter referred to as the EAMEC model) provided by an embodiment of the present invention is verified below.

[0078] In this embodiment of the present invention, knowledge graphs CH-DT, JA-DT, and AM-DT were constructed using railway data from China, Japan, and the United States, respectively. Their statistical characteristics are shown in Table 1. The datasets were constructed using OpenStreetMap, a public transportation data platform, and annotations from domain experts. The training, validation, and test sets were split in a 6:2:2 ratio, with seed alignment accounting for 30%.

[0079] The experiment was conducted under Ubuntu 18.04 system, the processor was AMD EPYC 7742 processor, and the memory was 512G.

[0080] Table 1 Knowledge graphs CH-DT, JA-DT, and AM-DT

[0081] The three datasets have very different encoding formats, are highly professional, and have many unique semantic entities. To achieve entity alignment, the following problems need to be solved: (1) Entity naming has national characteristics (for example, Japan's road network coding includes administrative division codes), and the semantic gap needs to be eliminated; (2) Business level design is significantly different (for example, CH-DT uses a four-level business classification, while JA-DT only defines business activities); (3) Attribute redundancy and missing, such as the ratio of the number of attribute triples to relationship triples in AM-DT is close to 1:0.83, which restricts the effectiveness of attribute embedding. Taking CH-DT as an example, the constructed dataset is introduced. Figure 4 As shown, we select two heterogeneous knowledge graphs contained in CH-DT and record them as and , the two align entity pairs through predefined seeds (such as <CDSD, MD >,<CD_SUR, DAA03> ) to establish preliminary semantic associations.

[0082] and The relationships and attributes are shown in Table 2. Among them, Adopting four-level business classification (level 1 to level 4), it can support fine-grained railway management scenarios; It is simplified to a single level of "business activities" to adapt to the needs of rapid response operation and maintenance. Contains engineering constraint attributes such as "is it non-empty" and "data source and description," making it suitable for supporting compliance checks for railway data governance. Adding field attributes like "width," "decimal," and "filling requirements" enables real-time validation of data collection systems. Furthermore, to distinguish complex business relationships, some relationships are explicitly defined using tail entity types. For example, the "Business Domain" relationship connects a head entity (such as "Road Network Monitoring System") with tail entity types (such as "Model" and "Table"), enabling cross-layer mapping between business and technical architectures.

[0083] Table 2 and Relationships and attributes

[0084] (1) Experimental setup and evaluation indicators Three sets of comparative experiments were conducted, namely: multi-view embedding combination analysis experiments based on CH-DT and JA-DT datasets; mainstream model comparison experiments based on CH-DT and JA-DT datasets; and attribute sparse scenario verification based on AM-DT dataset.

[0085] The comparison model parameters are uniformly set to: embedding dimension , learning rate , the loss function margin , training rounds = 200.

[0086] use Hits@k , that is, the correct alignment entity is in front The ratio of bits to MR (Average rank) is used as the core indicator, and the calculation formulas are as follows. Experimental settings 1 and 10 respectively.

[0087] in, To query entity sets, is the indicator function.

[0088] (2) Experimental analysis of multi-view embedding combinations To validate the complementary effect of multi-view information in the EAMEC model, this study conducted comparative experiments using six embedding combinations on the CH-DT and JA-DT datasets. As shown in Table 3, when combining summary, attribute, and relation embeddings, the EAMEC model achieved optimal performance on both datasets (CH-DT: Hits@1 = 92.04%; JA-DT: Hits@1 = 90.86%), an average improvement of 23.6 percentage points compared to single-view embedding.

[0089] Table 3 Comparison of embedding combination performance

[0090] Experimental results reveal three key findings: First, multi-view feature fusion significantly alleviates the limitations of a single view. When using summary embeddings alone, while BERT effectively captures textual semantics (CH-DT Hits@1 = 70.40%), the lack of structural information leads to significant cross-lingual entity ambiguity. When fused with relational embeddings, the model strengthens semantic constraints through structural features such as foreign key associations, improving CH-DT's Hits@1 to 74.40%. Further introducing attribute embeddings to form a three-view combination, the model achieves a deep integration of semantic and business logic features, ultimately achieving a Hits@1 of 90.86% on the JA-DT dataset. Second, a dynamic weighting mechanism effectively balances the contributions of different views. In the JA-DT dataset, attribute embeddings alone achieve a Hits@1 of 62.52%, significantly outperforming relational embeddings' 33.49%. However, when the two are combined, the model suppresses redundant attribute interference through dynamic weighting, improving performance to 56.34%. The traceability weight distribution revealed that relational embeddings contributed 38.7% to the combined model. By depicting primary-foreign key relationships between entities, they effectively bridged textual semantics and business characteristics. Finally, multi-view combination demonstrated strong scenario adaptability. In the cross-business hierarchical alignment task, the model successfully achieved precise mapping between the CH-DT "four-level business classification" and the JA-DT "business activity" nodes through semantic matching of summary embeddings and structural constraints of relational embeddings.

[0091] (3) Baseline model comparison experiment To comprehensively evaluate model performance, this study selected seven mainstream entity alignment models as baselines for comparative analysis: MTransE, IPTransE, JAPE, BootEA, KDCoE, AttributeE, and MultiKE. As shown in Table 4, EAMEC achieved Hits@1 rates of 92.04% and 90.86% on the CH-DT and JA-DT datasets, respectively, an improvement of 0.59-2.81 percentage points over the optimal baseline model, MultiKE.

[0092] Table 4 Model performance comparison

[0093] EAMEC's ​​performance advantages stem from the following aspects. At the model architecture level, existing methods have significant limitations: (1) Translation models (MTransE and IPTransE) rely solely on relational triple modeling, resulting in a hits@1 of less than 30% in attribute-intensive JA-DT scenarios; (2) Alignment-sensitive models (JAPE and BootEA) experience a 14.2% performance drop under 30% seed data, exposing their strong dependence on labeled data; and (3) Due to differences in design goals, the cross-lingual model KDCoE achieves a hits@1 of only 42.71% in cross-graph scenarios.

[0094] Comparison of multiple view models further validates the innovativeness of EAMEC: (1) The attribute enhancement model AttributeE achieves only 68.77% Hits@1 in the attribute-sparse CH-DT dataset (attribute / relation ratio ≈ 0.87), indicating its difficulty in handling the uneven distribution of attributes. (2) The MultiKE model accumulates name feature noise due to cross-language naming differences, resulting in a 1.59% decrease in Hits@1 compared to EAMEC. Through a dynamic fusion mechanism, this model achieves a 62% weighted summary embedding in the AM-DT dataset (attribute / relation ratio = 0.83), effectively compensating for the sparse attribute information.

[0095] (4) Validation of attribute sparse scenarios To verify EAMEC's ​​robustness in attribute-sparse scenarios, this experiment selected the AM-DT dataset (attribute / relation ratio = 0.83) for testing. The baseline models include TransE, MTransE, JAPE, and AttributeE. The experimental results and comparison charts are shown in Table 5.

[0096] Table 5 Comparison of model effectiveness in attribute sparse scenarios

[0097] EAMEC significantly outperforms AttributeE (91.02%) and traditional models (<35%) with a Hits@1 of 93.35%, validating the robustness of multi-view fusion. Mechanism analysis shows that: (1) in scenarios where there is a terminological discrepancy between "data item name" and "field name," summary embedding achieves 83.6% cross-term matching through semantic generalization; (2) when attribute values ​​differ across countries (e.g., mixing "kilometer" and "mile"), the dynamic weighting mechanism suppresses the average contribution of noisy attributes to 18.3%, effectively ensuring alignment accuracy.

[0098] The multi-source heterogeneous data entity alignment device provided by the present invention is described below. The multi-source heterogeneous data entity alignment device described below and the multi-source heterogeneous data entity alignment method described above can refer to each other.

[0099] See also Figure 5 , the multi-source heterogeneous data entity alignment device includes: The knowledge graph merging module 501 is used to merge two knowledge graphs according to a predicate alignment algorithm to obtain a merged knowledge graph; A summary relationship embedding learning module 502 is used to learn the semantic summary embedding and relationship structure embedding of each entity in the merged knowledge graph; Attribute semantic embedding learning module 503, used to learn attribute semantic embedding of each entity based on the dependency relationship and part of speech of words in the attribute text of each entity; An embedding combination module 504 is configured to determine a combined embedding vector for each entity based on the semantic summary embedding, the relational structure embedding, and the attribute semantic embedding of each entity; The entity alignment module 505 is used to align the entities in the merged knowledge graph according to the combined embedding vectors of the entities.

[0100] As for the more specific functions of the module, please refer to the multi-source heterogeneous data entity alignment method described above, which will not be repeated here.

[0101] Figure 6 An example of a physical structure diagram of an electronic device is shown below. Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communications bus 640. The processor 610 may call logic instructions in the memory 630 to execute a multi-source heterogeneous data entity alignment method, which includes: merging two knowledge graphs according to a predicate alignment algorithm to obtain a merged knowledge graph; learning semantic summary embeddings and relational structure embeddings for each entity in the merged knowledge graph; learning attribute semantic embeddings for each entity based on the dependency relationships and parts of speech of words in the attribute text of each entity; determining a combined embedding vector for each entity based on the semantic summary embedding, relational structure embedding, and attribute semantic embedding of each entity; and aligning each entity in the merged knowledge graph based on the combined embedding vectors of each entity. As for a more specific method, please refer to the multi-source heterogeneous data entity alignment method described above, which will not be repeated here.

[0102] Furthermore, the logic instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0103] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-source heterogeneous data entity alignment method provided by the above methods, the method including: merging two knowledge graphs according to a predicate alignment algorithm to obtain a merged knowledge graph; learning the semantic summary embedding and relational structure embedding of each entity in the merged knowledge graph; learning the attribute semantic embedding of each entity according to the dependency relationship and part of speech of the words in the attribute text of each entity; determining the combined embedding vector of each entity according to the semantic summary embedding, relational structure embedding and attribute semantic embedding of each entity; aligning each entity in the merged knowledge graph according to the combined embedding vector of each entity. As for a more specific method, please refer to the multi-source heterogeneous data entity alignment method described above, which will not be repeated here.

[0104] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the multi-source heterogeneous data entity alignment method provided by the above methods, the method comprising: merging two knowledge graphs according to a predicate alignment algorithm to obtain a merged knowledge graph; learning the semantic summary embedding and relational structure embedding of each entity in the merged knowledge graph; learning the attribute semantic embedding of each entity according to the dependency relationship and part of speech of the words in the attribute text of each entity; determining the combined embedding vector of each entity according to the semantic summary embedding, relational structure embedding and attribute semantic embedding of each entity; aligning each entity in the merged knowledge graph according to the combined embedding vector of each entity. As for a more specific method, reference may be made to the multi-source heterogeneous data entity alignment method described above, which will not be repeated here.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0106] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for aligning multi-source heterogeneous data entities, characterized in that: include: Merge the two knowledge graphs according to the predicate alignment algorithm to obtain a merged knowledge graph; Learn to merge the semantic summary embeddings and relational structure embeddings of each entity in the knowledge graph; Learn the attribute semantic embedding of each entity based on the dependency relationship and part of speech of the words in the attribute text of each entity; Determine the combined embedding vector of each entity based on its semantic summary embedding, relational structure embedding, and attribute semantic embedding. The entities in the merged knowledge graph are aligned according to their combined embedding vectors.

2. The multi-source heterogeneous data entity alignment method according to claim 1, characterized in that: The attribute semantic embedding of each entity is learned based on the dependency relationship and part of speech of the words in the attribute text of each entity, including: Perform dependency syntax analysis on the attribute value text in the attribute text, resolve the inter-word dependency relationship and extract the part-of-speech tags; Assign weights to words based on their part-of-speech tags; The attribute name text and attribute value text in the attribute text are concatenated into a text sequence, and the text sequence is input into the pre-trained Word2Vec word model to generate the initial word vector of the word; Calculate enhanced word vectors based on initial word vectors, inter-word dependencies, and word weights; Based on enhanced word vectors, a bidirectional GRU model is combined with a BERT model to learn the attribute semantic embedding of entities.

3. The multi-source heterogeneous data entity alignment method according to claim 2, characterized in that: Word weight distribution includes: The weights are assigned to words according to the rule that noun weight ≥ verb weight ≥ modifier weight.

4. The multi-source heterogeneous data entity alignment method according to claim 2, characterized in that: Calculate the enhanced word vector based on the initial word vector, inter-word dependencies, and word weights, including: The enhanced word vector is calculated according to the following formula: in, Expressive words The enhanced word vector of represents element-wise multiplication, represents the word weight, 、 Represents the words generated by the pre-trained Word2Vec word model and words The initial word vector of Expressive words and The dependency relationship, n Indicates the number of words, represents the normalized dependency association weight, is a learning parameter used to characterize the strength of inter-word dependencies.

5. The multi-source heterogeneous data entity alignment method according to claim 2, characterized in that: Based on enhanced word vectors, a bidirectional GRU model is used in conjunction with a BERT model to learn the attribute semantic embedding of entities, including: The enhanced word vector is input into the bidirectional GRU model for contextual semantic modeling, capturing the semantic features of the previous and next contexts and obtaining context-aware representations. Based on the context-aware representation, a bi-affine classifier is used to predict the dependencies between words and generate the projected representations of the dominant and dependent words. Calculate the dependency score of the word based on the projection representation of the collocation and dependent words; According to the dependency scores of the words, the hidden states of the dependency associations are weighted and aggregated to generate a syntactic-aware context representation. Based on the syntax-aware context representation, the BERT model is used for deep semantic encoding; Based on deep semantic encoding, the self-attention mechanism is used to dynamically aggregate global semantic information and output attribute semantic embedding.

6. The multi-source heterogeneous data entity alignment method according to claim 2 or 5, characterized in that: The loss function of the bidirectional GRU model combined with the BERT model is as follows: in, represents the total attribute embedding loss, Represent the weight parameters of the bidirectional GRU model loss and the BERT model loss, respectively. represents the GRU model loss, Represents the BERT model loss.

7. The multi-source heterogeneous data entity alignment method according to claim 1, characterized in that: Learning to merge the relational structure embeddings of entities in the knowledge graph, including: A translation model is combined with a graph convolutional neural network model to learn the structural features and neighborhood information of entities from the relation triplets in the merged knowledge graph to obtain the relational structure embedding of the entities; Among them, the translation model projects the head entity and the tail entity into independent dynamic semantic spaces for the relationship triples in the merged knowledge graph, and obtains the projected head entity representation and tail entity representation through matrix transformation.

8. The multi-source heterogeneous data entity alignment method according to claim 7, characterized in that: The loss function of the translation model combined with the graph convolutional neural network model is as follows: in, represents the total loss of the translation model combined with the graph convolutional neural network model, is the weight parameter, Expressing opposition to loss, represents the symmetric loss, Represent the projected representations of the head entity and the tail entity respectively.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the multi-source heterogeneous data entity alignment method as described in any one of claims 1 to 8 is implemented.

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, the multi-source heterogeneous data entity alignment method as described in any one of claims 1 to 8 is implemented.

Images