A heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization

Abstract

The application discloses a heterogeneous neighbor semantic perception entity alignment method based on two-dimensional convolution optimization, which is based on the original heterogeneous neighbor semantic perception (SEAM-HN) framework, and innovatively introduces a pseudo-embedding generation mechanism based on two-dimensional convolution and a generative adversarial network (GAN) training paradigm, generates high-difficulty adversarial negative samples by constructing a pseudo-embedding generator, forces the model to learn more fine-grained semantic differences in the alignment process, and thus significantly improves the discrimination ability and robustness of cross-language knowledge graph alignment. The method supports the integration of multi-source knowledge resources, the mapping of course ability elements, the semantic matching of real combat scenes and the recommendation of personalized learning paths in the process of network security police talent training, and provides a technical basis for constructing a new era of public security education intelligent teaching system with the characteristics of "real combat orientation and digital empowerment".

Description

Technical Field

[0001] This invention relates to the field of knowledge graph completion and link prediction technology, and in particular to a heterogeneous neighbor semantic awareness entity alignment method (SEAM-HN-GAN) based on two-dimensional convolution optimization. Background Technology

[0002] In the field of knowledge graph completion and link prediction, the core objective is to predict missing relationships between entities using algorithmic models, thereby improving the knowledge graph structure and supporting downstream intelligent applications (such as intelligent question answering, recommendation systems, and semantic retrieval). The known technologies in this field mainly revolve around two core aspects: "entity and relation embedding modeling" and "triple score prediction." Key known technologies include:

[0003] (1) Early mainstream methods of traditional knowledge graph embedding technology mainly used shallow models. The core idea was to map entities and relations to a low-dimensional vector space and model their relationships through simple mathematical operations. For example:

[0004] TransE model: Treats relations as translation operations between entity vectors, measures the rationality of ternary combination by calculating the distance es + r≈eo. The model is simple and parameter efficient, but it is difficult to handle complex many-to-many relations.

[0005] DistMult model: using trilinear dot product , , As a scoring function, it enhances expressive power by increasing the complex dimension of the relation vector, but the parameter scale is large, and the modeling effect on high-in-degree nodes is limited.

[0006] Graph Convolutional Networks (GCNs) and Relational Graph Convolutional Networks (R-GCNs): GCNs generalize convolution operators to graph structures and learn entity embeddings using local information, but they are limited to undirected graphs and cannot handle the directed edge semantics of knowledge graphs; R-GCNs, as an extension, are suitable for multi-relation scenarios and can capture relation semantics, but they suffer from high complexity, slow computation speed, and dependence on relation annotations.

[0007] (2) Application of Convolutional Neural Networks in Embedding Modeling With the development of deep learning, convolution operations have been introduced into knowledge graph embedding tasks. The core is to extract local interaction features of entities and relationships through convolution. Early related technologies mostly used one-dimensional convolution (such as time series operations for entity embedding), which could only capture limited interactions near the embedding splicing points, and the feature expression ability was limited. However, the exploration of two-dimensional convolution technology (represented by two-dimensional convolution models) has become a key breakthrough. By reshaping the one-dimensional embedding of entities / relationships into a two-dimensional matrix, spatial convolution is used to capture richer local associations, which greatly improves the depth and breadth of feature interactions.

[0008] (3) The core optimization logic of the existing technology for link prediction is based on “accuracy of triplet scoring” and “model efficiency”: by designing reasonable scoring functions (such as inner product, distance calculation, nonlinear mapping, etc.) to distinguish positive and negative samples, using techniques such as Dropout and regularization to avoid overfitting, and improving the training and inference speed of the model through batch calculation and parameter sharing, so as to adapt to the application scenarios of knowledge graphs with millions of entities.

[0009] Based on the current state of technology in the fields of knowledge graph link prediction and entity alignment, existing technologies (represented by two-dimensional convolutional models and related derivative solutions) have the following clear technical shortcomings:

[0010] (1) The negative samples are of low quality and lack discrimination: existing solutions generate negative samples by randomly replacing entities. These negative samples are too different from positive samples and are easy to distinguish. There is a lack of high-quality difficult negative samples that are "difficult to distinguish but not truly matched". Some solutions that attempt to construct difficult negative samples do not combine model embedding and reconstruction logic, resulting in poor semantic rationality of negative samples. They cannot effectively train the model's fine-grained feature discrimination ability and have limited generalization performance.

[0011] (2) Poor adaptability of convolution kernels and weak cross-scene transfer capability: The convolution kernel design with fixed parameters cannot adapt to the differences in feature distribution of different knowledge graphs (such as different sources, data distributions, and dimension settings), resulting in deviations in local feature capture. In complex scenarios such as cross-domain and multilingual knowledge graphs, the model performance degrades significantly.

[0012] (3) Lack of multimodal information fusion and one-sided semantic representation: It only relies on the entity structure embedding based on the knowledge graph topology, without integrating multimodal information such as entity attributes (such as "category" and "attribute value") and text descriptions (such as natural language descriptions), resulting in incomplete entity semantic feature representation. In knowledge graphs with sparse structural information and rich attributes, the accuracy of link prediction or entity alignment is insufficient.

[0013] (4) Insufficient cross-graph interaction modeling capability: It is limited to the interaction modeling of "entity-relationship-entity" triples within a single knowledge graph, and has not designed a dedicated interaction mechanism for cross-knowledge graph entity pairs, making it difficult to meet the needs of complex business scenarios such as multi-source knowledge graph fusion and cross-modal entity alignment.

[0014] (5) Fixed weights in the loss function lead to an imbalance in the optimization objective: The fixed weights are used to balance the embedding reconstruction loss and the alignment loss. This cannot be dynamically adjusted according to the training (such as the similarity deviation between pseudo-embedded and real embeddings, and the imbalance of the loss ratio). This results in the model either excessively pursuing the "realism" of pseudo-embedded and losing its ability to distinguish, or having too strong a distinguishing ability and insufficient semantic rationality of pseudo-embedded, which affects the final modeling effect. Summary of the Invention

[0015] This invention addresses the technical shortcomings of existing technologies by proposing a heterogeneous neighbor semantic awareness entity alignment method based on 2D convolution optimization. It aims to comprehensively improve the accuracy, generalization, and adaptability to complex scenarios of models in tasks such as knowledge graph link prediction and entity alignment, providing reliable technical support for practical applications such as intelligent question answering, multilingual knowledge graph fusion, and cross-modal matching. Building upon the existing heterogeneous neighbor-based semantic awareness (SEAM-HN) framework, this method innovatively introduces a pseudo-embedding generation mechanism based on 2D convolution and a generative adversarial network (GAN) training paradigm. By constructing a pseudo-embedding generator to create highly challenging adversarial negative samples, the model is forced to learn finer-grained semantic differences during the alignment process, thereby significantly improving the discriminative ability and robustness of cross-lingual knowledge graph alignment. This method supports the integration of multi-source knowledge resources, mapping of course competency elements, semantic matching of practical scenarios, and personalized learning path recommendation in the training of cybersecurity police personnel, providing a technical foundation for building a new era of intelligent teaching system for public safety education that is "practice-oriented and digitally empowered."

[0016] To achieve the above objectives, the present invention provides the following technical solution:

[0017] In a first aspect, the present invention provides a heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization, comprising the following steps:

[0018] Heterogeneous graph structure embedding encoding steps: A graph convolutional network (GCN) incorporating TransE regularization is used to extract the source knowledge graphs. and target knowledge graph The initial structured features of the entities in the middle are used to generate real entity embedding vectors;

[0019] The pseudo-embedding generation steps are as follows: the real entity embedding vector is input into the generator based on two-dimensional convolution, and it is transformed into a two-dimensional matrix through a reshaping operation. Two-dimensional convolution operation is performed on the two-dimensional matrix to extract deep spatial features. Then, it is restored to the original dimension through a fully connected layer to generate a pseudo-embedding that is highly similar to the real entity embedding.

[0020] Adversarial negative sample construction and discrimination steps: construct hard negative sample pairs using pseudo-embedding and real entity embedding, and use an alignment loss function through a discriminator to distinguish between real aligned positive sample pairs and hard negative sample pairs (pseudo-embedding aligned pairs).

[0021] Joint loss optimization steps: Construct a composite loss function that includes reconstruction loss and alignment discrimination loss, and jointly optimize the parameters of the graph convolutional network, generator and discriminator through backpropagation, and output the optimized entity alignment mapping relationship.

[0022] Furthermore, the reconstruction loss uses the Wasserstein distance to measure the difference between the true entity embedding distribution and the pseudo-embedded generation distribution, and its calculation formula is as follows:

[0023] ,

[0024] in, For true embedding distribution, Distributions generated for pseudo-embedding. As a discriminator, the reconstruction loss is used to encourage the pseudo-embedded distribution to approximate the true embedded distribution.

[0025] Furthermore, the alignment discrimination loss adopts a contrastive loss function based on Softmax, using real aligned entity pairs as positive samples and pseudo-embedded pairs as difficult negative samples, forcing the model to increase the similarity score of real aligned entity pairs and decrease the score of difficult negative sample pairs.

[0026] Furthermore, the joint loss optimization step also includes a gradient penalty term, which constrains the gradient norm of the discriminator to be close to 1 at the interpolation points between the real and generated samples, thereby satisfying the Lipschitz continuity condition of the Wasserstein distance. Its calculation formula is as follows:

[0027] ,

[0028] in, These are random interpolation points between real and generated samples. This is the similarity function.

[0029] Furthermore, the joint loss optimization step employs a dynamic weight adaptive strategy, which automatically adjusts the weight of the alignment discrimination loss in the total loss based on the magnitude of the reconstruction loss: when the reconstruction loss is too small, the weight of the alignment discrimination loss is automatically increased; when the reconstruction loss is too large, the weight of the alignment discrimination loss is decreased.

[0030] Furthermore, the two-dimensional convolution generator in the pseudo-embedding generation step adopts a dynamic convolution kernel adaptive generation mechanism. By calculating the mean and variance statistical features of the target knowledge graph entity embedding, it uses parameterized functions to generate exclusive convolution kernels for different knowledge graphs to adapt to the feature distribution of different knowledge graphs.

[0031] Furthermore, the above-mentioned heterogeneous neighbor semantic awareness entity alignment method based on two-dimensional convolution optimization also includes a cross-knowledge graph attention-enhanced interaction modeling step: concatenating the feature vectors of cross-knowledge graph entity pairs into a joint matrix and reshaping it into a two-dimensional structure, introducing an attention mechanism to learn the weights of the feature dimensions and performing weighted fusion, and then capturing deep interaction features through secondary convolution to strengthen the semantic association representation of cross-knowledge graph entity pairs.

[0032] Furthermore, the above-mentioned heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization also includes a multimodal multi-channel semantic feature fusion step: the entity's structural embedding, attribute embedding, and text embedding are reshaped into two-dimensional matrices to form multi-channel inputs, a dedicated convolution kernel is set for each modality to extract local features, and weights are calculated based on feature variance to perform weighted fusion, so as to enrich the semantic information dimension of the entity.

[0033] Secondly, this invention proposes a heterogeneous neighbor semantic-aware entity alignment system based on two-dimensional convolution optimization to implement the method described in any of the above claims, comprising:

[0034] Heterogeneous graph structure embedding encoding module: used to extract the initial structured features of entities in the source knowledge graph and the target knowledge graph, and generate real entity embeddings;

[0035] The pseudo-embedding generation module based on 2D convolution is used to receive real entity embeddings and generate pseudo-embeddings through 2D convolution operations.

[0036] Adversarial negative sample construction and discrimination module: used to construct real alignment pairs and pseudo-embedded alignment pairs as hard negative samples, and to distinguish them by a discriminator;

[0037] Joint Loss Optimization and Iteration Module: This module is used to calculate the composite loss function, which includes reconstruction loss and alignment discrimination loss, and to jointly optimize the parameters of each module.

[0038] Furthermore, the pseudo-embedding generation module based on two-dimensional convolution performs the following operations in sequence: ... The entity embedding vector is reshaped into A two-dimensional matrix, where The 2D matrix is ​​convolved using a 2D convolution kernel and then processed by an activation function to obtain a feature map. The feature map is then flattened and mapped back to its original dimensions using a fully connected layer. This results in a pseudo-embedding.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] (1) High-discrimination sample generation significantly improves the model's generalization ability.

[0041] To address the shortcomings of existing technologies, such as low negative sample quality and insufficient fine-grained discrimination ability, this invention utilizes a difficult negative sample generation mechanism based on two-dimensional convolutional reconstruction loss. It employs a Reshape-Conv-FC process to generate pseudo-embeddings highly similar to the true embeddings, constructing a difficult negative sample consisting of pseudo-embeddings and the true embedding (e.g., ...). and Simultaneously, through the synergistic optimization of reconstruction loss (controlling the similarity between pseudo-embedded and true embeddings) and alignment loss (distinguishing between true and false sample pairs), the model can effectively learn the feature differences of similar but mismatched samples. In link prediction and entity alignment tasks, the ability to distinguish similar entities is significantly improved, and the generalization performance is significantly optimized in multilingual KG and cross-modal scenarios. For example, in multilingual KG fusion tasks, the entity alignment accuracy is improved by 15%-20%.

[0042] (2) Dynamic convolution kernel adaptation to solve the feature capture bias across scenes.

[0043] To address the issue that fixed convolutional kernels cannot adapt to different KG feature distributions, this invention proposes a dynamic convolutional kernel generation method based on KG embedding statistical features (mean, variance), which is used to align the KG features. , Customized convolution kernels for each ( This design allows the convolutional kernel to adapt to the differences in entity embedding distributions of different KGs (such as differences in attribute features of cross-domain KGs and differences in semantic encoding of multilingual KGs), improving the accuracy of local feature extraction, avoiding the performance degradation problem of traditional fixed convolutional kernels, and significantly enhancing cross-scene transfer stability, such as improving the MRR (mean reciprocal ranking) index by 8%-12% in cross-domain KG link prediction tasks.

[0044] (3) Multimodal feature fusion to improve the semantic representation dimension of entities.

[0045] To address the limitation of relying solely on structural embeddings in semantic representation, this invention extends the two-dimensional convolutional input to a multi-channel feature set including structural embeddings, attribute embeddings, and text embeddings. This is achieved through a modality-specific convolutional kernel (…). , , Local features of each modality are extracted separately, and weights are calculated based on feature variance to achieve weighted fusion (e.g.) This approach supplements the semantic attributes of entities (such as "age" and "category") with textual descriptive information (such as entity descriptions encoded by BERT). In key-value pairs (KGs) with sparse structural information (such as emerging domain KGs), the semantic features of entities are more complete, effectively solving the problem of "low modeling accuracy caused by sparse structure".

[0046] (4) Cross-KG attention interaction to enhance multi-source knowledge association modeling

[0047] To address the limitations of modeling only single-KG interactions and insufficient cross-KG association capture, this invention adopts a 2D convolutional embedding-2D reshaping-convolution process. It concatenates the feature vectors of cross-KG entity pairs into a joint matrix, introduces a cross-KG attention mechanism to highlight high semantic matching feature regions (such as the embedding dimensions corresponding to entity "nationality" and "type"), and then captures deep interaction features through secondary convolution. This design breaks through the boundaries of single-KG interactions and can effectively support multi-language KG fusion, cross-domain graph alignment, and other multi-source knowledge integration tasks. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0049] Figure 1 This is an overall architecture diagram of the heterogeneous neighbor semantic awareness entity alignment method based on two-dimensional convolution optimization provided in an embodiment of the present invention.

[0050] Figure 2 This is a flowchart of pseudo-embedding generation and adversarial training provided in an embodiment of the present invention.

[0051] Figure 3 A schematic diagram illustrating the gradient penalty principle of WGAN-GP provided in an embodiment of the present invention. Detailed Implementation

[0052] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Example 1

[0054] This invention proposes a heterogeneous neighbor semantic-aware entity alignment method based on 2D convolution optimization. The overall architecture is as follows: Figure 1 As shown, the system mainly consists of four core modules: (1) Heterogeneous Graph Embedding module, (2) Generator module based on two-dimensional convolution pseudo-embedding, (3) Discriminator module for constructing and discriminating adversarial negative samples, and (4) Joint loss optimization and iteration module.

[0055] Input: Receive source knowledge graph and target knowledge graph .

[0056] Heterogeneous graph structure embedding encoding module: Adopting the SEAM-HN architecture, it includes a TransE-regularized graph convolutional network (GCN) for extracting initial structured feature vectors of entities and relations. and .

[0057] A pseudo-embedding generation module based on 2D convolution (key improvement): A convolutional embedding network (2D convolution) is introduced as the generator. It receives real entity embeddings. By reshaping Convolution The process of fully connected reconstructing a fully connected structure generates pseudo-embedded structures that are highly similar to the real embeddings. .

[0058] Adversarial negative sample construction and discrimination module (key improvement): Instead of relying solely on random negative sampling, it constructs true alignment pairs (Positive Pairs) and pseudo-embedded alignment pairs (Hard Negative Pairs). A discriminator (alignment loss function) distinguishes between the two, and training is stabilized using Wasserstein distance and gradient penalty techniques.

[0059] Output: Outputs the optimized entity alignment mapping relationship.

[0060] 1. Heterogeneous Graph Embedding Module

[0061] This module is responsible for extracting the basic features of the knowledge graph, providing high-quality ground truth for subsequent adversarial generation.

[0062] Entity embedding: Aggregating heterogeneous neighbor information using multi-layer GCN:

[0063]

[0064] Relational embedding: Combining the TransE translation mechanism The semantic space of constraint relation vectors.

[0065] This section retains the original SEAM-HN's ability to handle heterogeneous neighbor structures.

[0066] 2. A pseudo-embedding generation module based on 2D convolution (Generator)

[0067] This invention designs a generator based on a two-dimensional convolutional structure, aiming to generate sufficiently realistic pseudo-embeddings to simulate highly challenging interference samples.

[0068] The technical implementation path is as follows:

[0069] (1) Reshape: Reshape the shape of the dimension... entity embedding vector Reshape into a 2D matrix (tensor) ,in This allows the model to capture local interaction patterns of features using convolution operations.

[0070] (2) Convolution: using a set of 2D convolution kernels right Perform convolution operations to extract deep spatial features:

[0071]

[0072] (3) Fully Connected (FC): Flatten the convolutional output feature map and map it back to the original dimension through a fully connected layer. Resulting in pseudo-embedding :

[0073]

[0074] Functional significance: This process simulates the "variants" of an entity. The generated... The original entity was preserved. Core semantic features (derived from) Furthermore, nonlinear transformation differences were introduced.

[0075] 3. Adversarial negative sample construction and discrimination module (Discriminator)

[0076] To fully utilize pseudo-embedding to improve model performance, this invention constructs a composite loss function system based on the GAN (Generative Adversarial Network) concept.

[0077] (1) Reconstruction Loss – Controls the similarity of the imitation

[0078] To ensure the generated pseudo-embedding Sufficiently deceptive (i.e., resembling a true embedding sufficiently) a reconstruction loss is introduced. This invention innovatively introduces Wasserstein distance instead of traditional Euclidean distance to address the gradient vanishing problem when distributions do not overlap:

[0079]

[0080] Chinese annotation: For true embedding distribution, Distributions generated for pseudo-embedding. This is the discriminator. The loss causes the pseudo-embedded distribution to approximate the true embedded distribution.

[0081] (2) Alignment Discrimination Loss – Learning to distinguish between true and false signals

[0082] Hard negative pairs are constructed using pseudo-embedding.

[0083] Positive sample pairs: That is, the entities that are actually aligned in the two KGs.

[0084] Difficult negative sample pairs: and .in They are Pseudo-embedding.

[0085] Loss function:

[0086]

[0087] Technical implications: This formula, through Softmax, forces the model to assign extremely high similarity scores to "true-to-true" pairings, while simultaneously lowering the scores for "false-to-true" pairings. Because... and High similarity means the model must capture extremely subtle semantic differences to complete the classification, thereby significantly improving its discriminative ability.

[0088] (3) Gradient Penalty enhances robustness

[0089] To satisfy the Lipschitz continuity constraint of WGAN and prevent training crashes, in Introducing a gradient penalty term:

[0090]

[0091] Chinese annotation: These are random interpolated samples between real pairs and pseudo-embedded pairs. This forces the gradient norm of the discriminant function to be close to 1, ensuring the smoothness and stability of the optimization process.

[0092] like Figure 3 As shown:

[0093] Left ellipse: Pr (Real Distribution) → Pr (True Data Distribution)

[0094] x Sampled from the true distribution

[0095] Right ellipse: Pg (Generated Distribution) → Pg (Generated Data Distribution)

[0096] ' 'Sampling self-generated distribution'

[0097] Intermediate connecting line and point:

[0098] Interpolated Interpolation Sample

[0099] The text and arrows above:

[0100] Gradient Constraint: Gradient constraint: The gradient norm of the discriminator at the interpolated sample x^ is equal to 1.

[0101] Figure 3 This section provides an intuitive explanation of the mathematical principle behind gradient penalty in WGAN-GP. The ellipse on the left represents the space of the real data distribution, while the one on the right represents the space of the generated (fake) data distribution. To ensure training stability (satisfying the Lipschitz continuity condition), we interpolate between the real sample x and the generated sample x' to obtain an intermediate point x^. The core idea of ​​gradient penalty is to force the gradient of the discriminator D at this interpolation point x^ to be close to 1, thereby preventing gradient vanishing or exploding and making model training smoother.

[0102] 4. Joint Loss Optimization and Iteration Module

[0103] The total loss of dynamic weight adjustment is defined as:

[0104]

[0105] in Employ a dynamic adaptive strategy: when When the size is too small (the pseudo-embedding is too realistic to distinguish), it automatically increases in size. Increase attention by emphasizing certain aspects; conversely, decrease attention by emphasizing others.

[0106] This invention proposes an end-to-end knowledge graph modeling scheme based on 2D convolution optimization. Aiming at knowledge graph link prediction and cross-KG entity alignment tasks, it constructs a full-process architecture encompassing input preprocessing, core feature processing, sample augmentation, loss optimization, and result output. This integrates dynamic convolutional adaptation, cross-KG interaction enhancement, multimodal feature fusion, high-discrimination sample generation, and dynamic loss optimization. It addresses the technical problems in existing technologies, such as cross-KG feature capture bias, single-KG interaction limitations, one-sided semantic representation, low-quality negative samples, and imbalanced optimization objectives, achieving efficient, accurate, and highly generalizable knowledge graph semantic modeling. This invention also incorporates a deep semantic understanding method based on BERT multi-vector semantic matrices, offering the following innovative points.

[0107] (1) Adaptive generation method of dynamic convolution kernel

[0108] This invention proposes a dynamic convolution kernel generation method based on the statistical features of knowledge graph embedding. By calculating the mean, variance, and other statistical features of the target knowledge graph entity embedding, a parameterized function is used to generate a dedicated convolution kernel for different knowledge graphs, replacing the fixed convolution kernel design. This achieves adaptive adaptation to the feature distribution of different knowledge graphs, effectively solving the problem of local feature capture bias caused by the difference in feature distribution across knowledge graphs, and improving the accuracy of cross-scene feature extraction.

[0109] (2) Cross-KG attention-enhanced interaction modeling method

[0110] This invention proposes a cross-Knowledge Graph (KG) entity pair attention-enhanced interaction modeling method. It concatenates the feature vectors of cross-KG entity pairs into a joint matrix and reshapes it into a two-dimensional structure. It introduces an attention mechanism to learn the importance of feature dimensions and performs weighted fusion. Then, it captures deep interaction features through secondary convolution. This method breaks through the limitation of traditional two-dimensional convolution, which can only model interactions within a single KG. It strengthens the semantic association representation of cross-KG entity pairs and improves the association matching accuracy in complex scenarios.

[0111] (3) A method for generating high-discrimination, hard-to-bear samples based on reconstruction loss

[0112] This invention proposes a high-discrimination, hard-negative sample generation method based on embedding reconstruction loss. It utilizes the embedding reconstruction logic of two-dimensional convolution to generate "pseudo-embeddings" that are highly similar to the embeddings of real entities, constructing hard-negative samples of "pseudo-embeddings + real embeddings". Combining reconstruction loss to control the similarity of pseudo-embeddings and alignment loss to distinguish sample pairs, this method solves the problems of low quality of negative samples and insufficient fine-grained discrimination ability in existing technologies, thereby improving the generalization performance of the model.

[0113] (4) Multimodal and multichannel semantic feature fusion method

[0114] This invention proposes a multimodal, multi-channel knowledge graph entity semantic feature fusion method. The structural embedding, attribute embedding, and text embedding of entities are reshaped into two-dimensional matrices to form multi-channel inputs. Dedicated convolutional kernels are designed for each modality to extract local features, and weights are calculated based on feature variance to achieve weighted fusion. This method overcomes the one-sidedness of traditional solutions that rely solely on structural embedding for semantic representation and enriches the dimensions of entity semantic information.

[0115] (5) Adaptive allocation method for dynamic loss weights

[0116] This invention proposes a dynamic loss weight adaptive allocation method. By calculating the real-time ratio of reconstruction loss to alignment loss during training, an adaptive weight function is designed to dynamically adjust the loss weights, balancing the "fidelity" of pseudo-embeddings with the "discrimination ability" of the model. This solves the problem of imbalance in optimization objectives caused by fixed weights, and improves the stability and final performance of model training.

[0117] Example 2

[0118] This invention proposes a heterogeneous neighbor semantic-aware entity alignment method based on 2D convolution optimization. The specific processing flow is as follows: Figure 2 As shown, it includes the following steps:

[0119] S1 Initialization and Encoding: Acquiring data using GCN and TransE , Entity / relationship embedding.

[0120] S2 pseudo-embedding generation (G process): Input entity embeddings into a 2D convolution generator and output pseudo-embeddings.

[0121] S3 Difficult Negative Sample Construction: Combine true embeddings and false embeddings to form highly confusing negative sample pairs.

[0122] S4 calculates the reconstruction loss: constrains the distribution distance (imitation) between the pseudo-embedded and the real embedded.

[0123] S5 calculates the alignment discrimination loss: maximizes the score difference (discrimination) between real pairs and pseudo-sample pairs.

[0124] S6 Joint backpropagation: Combines the original structural loss of SEAM-HN to update the model parameters.

[0125] (1) Description of technology application scenarios

[0126] The core application scenario of this invention is multi-source knowledge graph fusion related tasks, specifically including:

[0127] 1) Cross-language knowledge graph entity alignment: such as entity matching between Chinese DBpedia and English DBpedia, solving the alignment problem caused by the difference in entity structure in different language graphs;

[0128] 2) Cross-domain knowledge graph fusion: such as the fusion of public safety public opinion knowledge graph and government service knowledge graph, to handle the unique neighbor structure differences of domain-specific entities;

[0129] 3) Long-tail entity alignment: For long-tail entities that appear infrequently and have sparse neighbors in the knowledge graph, the alignment accuracy is improved by supplementing the signal of unique neighbors.

[0130] (2) Key parameter setting instructions

[0131] The core parameters and default settings of this invention have been verified based on mainstream datasets (DBP15K ZH-EN, JA-EN, FR-EN) and are universally applicable.

[0132] GCN parameters: 2-layer GCN, hidden layer dimension 300, learning rate 0.001;

[0133] Weight and threshold parameters: α=0.6 (weight calculation tradeoff coefficient), β=2.0 (calibration item weight), γ=0.5 (reward coefficient), δe=5 (distance threshold), conflict score threshold=1.0;

[0134] Iteration parameters: maximum number of iterations 4, number of candidate entities to be screened 100.

[0135] (3) Comparison of core differences with the original RMN model

[0136]

[0137] (4) Experimental verification

[0138] This invention has been validated on mainstream datasets, and the core experimental results are as follows (compared with the original RNN model):

[0139]

[0140] Experimental results show that the present invention achieves stable performance improvements on different datasets, with the most significant improvement on the ZH-EN dataset, which has significant cross-language differences, thus verifying the effectiveness of the unique neighbor enhancement mechanism.

[0141] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization, characterized in that, Includes the following steps: Heterogeneous graph structure embedding encoding steps: A graph convolutional network (GCN) incorporating TransE regularization is used to extract the source knowledge graphs. and target knowledge graph The initial structured features of the entities in the middle are used to generate real entity embedding vectors; The pseudo-embedding generation steps are as follows: the real entity embedding vector is input into the generator based on two-dimensional convolution, and it is transformed into a two-dimensional matrix through a reshaping operation. Two-dimensional convolution operation is performed on the two-dimensional matrix to extract deep spatial features. Then, it is restored to the original dimension through a fully connected layer to generate a pseudo-embedding that is highly similar to the real entity embedding. Adversarial negative sample construction and discrimination steps: construct hard negative sample pairs using pseudo embeddings and real entity embeddings, and use an alignment loss function to distinguish real aligned positive sample pairs from hard negative sample pairs through a discriminator; Joint loss optimization steps: Construct a composite loss function that includes reconstruction loss and alignment discrimination loss, and jointly optimize the parameters of the graph convolutional network, generator and discriminator through backpropagation, and output the optimized entity alignment mapping relationship.

2. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, The reconstruction loss uses the Wasserstein distance to measure the difference between the true entity embedding distribution and the pseudo-embedded generation distribution, and its calculation formula is as follows: ， in, For true embedding distribution, Distributions are generated for pseudo-embedding. As a discriminator, the reconstruction loss is used to encourage the pseudo-embedded distribution to approximate the true embedded distribution.

3. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, The alignment discrimination loss adopts a contrastive loss function based on Softmax, which uses real aligned entity pairs as positive samples and pseudo-embedded pairs as difficult negative samples, forcing the model to increase the similarity score of real aligned entity pairs and decrease the score of difficult negative sample pairs.

4. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, The joint loss optimization step also includes a gradient penalty term, which constrains the gradient norm of the discriminator to be close to 1 at the interpolation points between the real and generated samples, thereby satisfying the Lipschitz continuity condition of the Wasserstein distance. Its calculation formula is as follows: ， in, These are random interpolation points between real and generated samples. This is the similarity function.

5. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, The joint loss optimization step employs a dynamic weight adaptive strategy, which automatically adjusts the weight of the alignment discrimination loss in the total loss based on the magnitude of the reconstruction loss: when the reconstruction loss is too small, the weight of the alignment discrimination loss is automatically increased; when the reconstruction loss is too large, the weight of the alignment discrimination loss is decreased.

6. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, The two-dimensional convolution generator in the pseudo-embedding generation step adopts a dynamic convolution kernel adaptive generation mechanism. By calculating the mean and variance statistical features of the target knowledge graph entity embedding, it uses parameterized functions to generate exclusive convolution kernels for different knowledge graphs to adapt to the feature distribution of different knowledge graphs.

7. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, It also includes a cross-knowledge graph attention-enhanced interaction modeling step: concatenating the feature vectors of cross-knowledge graph entity pairs into a joint matrix and reshaping it into a two-dimensional structure, introducing an attention mechanism to learn the weights of the feature dimensions and performing weighted fusion, and then capturing deep interaction features through secondary convolution to enhance the semantic association representation of cross-knowledge graph entity pairs.

8. The heterogeneous neighbor semantic-aware entity alignment method based on two-dimensional convolution optimization according to claim 1, characterized in that, It also includes a multimodal, multi-channel semantic feature fusion step: the entity's structural embedding, attribute embedding, and text embedding are reshaped into two-dimensional matrices to form multi-channel inputs, a dedicated convolution kernel is set for each modality to extract local features, and weights are calculated based on feature variance to perform weighted fusion, so as to enrich the semantic information dimension of the entity.

9. A heterogeneous neighbor semantic-aware entity alignment system based on two-dimensional convolution optimization, characterized in that, include: Heterogeneous graph structure embedding encoding module: used to extract the initial structured features of entities in the source knowledge graph and the target knowledge graph, and generate real entity embeddings; The pseudo-embedding generation module based on 2D convolution is used to receive real entity embeddings and generate pseudo-embeddings through 2D convolution operations. Adversarial negative sample construction and discrimination module: used to construct real alignment pairs and pseudo-embedded alignment pairs as hard negative samples, and to distinguish them by a discriminator; Joint Loss Optimization and Iteration Module: This module is used to calculate the composite loss function, which includes reconstruction loss and alignment discrimination loss, and to jointly optimize the parameters of each module.

10. The heterogeneous neighbor semantic-aware entity alignment system based on two-dimensional convolution optimization according to claim 9, characterized in that, The pseudo-embedding generation module based on two-dimensional convolution performs the following operations in sequence: ... The entity embedding vector is reshaped into A two-dimensional matrix, where The 2D matrix is ​​convolved using a 2D convolution kernel and then processed by an activation function to obtain a feature map. The feature map is then flattened and mapped back to its original dimensions using a fully connected layer. This results in a pseudo-embedding.

Images