Multi-view Knowledge Graph Embedding for Entity Alignment (1906.02390v1)

Abstract: We study the problem of embedding-based entity alignment between knowledge graphs (KGs). Previous works mainly focus on the relational structure of entities. Some further incorporate another type of features, such as attributes, for refinement. However, a vast of entity features are still unexplored or not equally treated together, which impairs the accuracy and robustness of embedding-based entity alignment. In this paper, we propose a novel framework that unifies multiple views of entities to learn embeddings for entity alignment. Specifically, we embed entities based on the views of entity names, relations and attributes, with several combination strategies. Furthermore, we design some cross-KG inference methods to enhance the alignment between two KGs. Our experiments on real-world datasets show that the proposed framework significantly outperforms the state-of-the-art embedding-based entity alignment methods. The selected views, cross-KG inference and combination strategies all contribute to the performance improvement.

Multi-view Knowledge Graph Embedding for Entity Alignment

The paper "Multi-view Knowledge Graph Embedding for Entity Alignment" presents an innovative framework, MultiKE, designed to tackle the challenges of entity alignment between diverse knowledge graphs (KGs). Entity alignment is crucial in unifying knowledge from different sources, facilitating applications such as semantic search and recommendation systems. The current approaches predominantly focus on the relational structures of entities, often ignoring or inadequately integrating other features like attributes or entity names, thereby limiting the robustness of entity alignment processes.

Framework and Contributions

MultiKE addresses these limitations by proposing a comprehensive multi-view embedding approach that incorporates three distinct views for learning entity embeddings:

1. Entity Names
2. Relations
3. Attributes

The framework distinguishes itself by employing specific embedding techniques suitable for each view and strategically combining these embeddings to improve entity alignment accuracy.

• Literal Embedding: Utilizes pre-trained word embeddings for names and textual attributes to capture semantic similarities.
• Name View Embedding: Directly leverages the literal embeddings to represent entity names.
• Relation View Embedding: Adopts translational models like TransE to maintain relational structure information.
• Attribute View Embedding: Applies convolutional neural networks to derive attribute-driven embeddings.

A key element of MultiKE is its cross-KG training mechanism. It includes:

• Entity Identity Inference: Enhances alignment precision by inferring identities through the consistent swapping of aligned entities across relation facts.
• Relation and Attribute Identity Inference: Introduces a novel, soft alignment technique, which autonomously identifies and processes relation and attribute correspondences, reducing dependencies on predefined alignments.

The paper explores three methodologies for integrating view-specific embeddings:

• Weighted View Averaging: Weighted aggregation of view embeddings based on calculated importances.
• Shared Space Learning: Employs orthogonal transformations to map view-specific embeddings into a common embedding space.
• In-training Combination: A joint training strategy that integrates learnings from multiple views in real-time, enhancing the adaptability and effectiveness of embeddings.

Experimental Evaluation

Experiments conducted on datasets DBP-WD and DBP-YG illustrate the superior performance of MultiKE against state-of-the-art methods. Notably, MultiKE achieves remarkable improvements in Hits@1 and MRR scores, signifying enhanced precision and reliability over other embedding-based approaches. The paper reveals that the synergy of diverse, multi-view embeddings and the capability to handle semantic heterogeneity are central facets contributing to this success.

Implications and Future Prospects

MultiKE's framework has demonstrated significant potential in advancing entity alignment accuracy by effectively capturing the multifaceted nature of KGs through multi-view representations. This has broad implications in KG construction, fusion tasks, and downstream applications reliant on integrated knowledge from disparate sources.

Future research avenues suggested by the paper include the exploration of additional views, such as entity types, and the adaptation of the framework for cross-lingual entity alignment tasks. These expansions promise to further enhance the versatility and applicability of the approach in handling complex real-world datasets with varying linguistic contexts.

In summation, the paper makes a compelling case for the adoption of multi-view embedding techniques in entity alignment, providing a robust solution that addresses current methodological gaps and sets a groundwork for future innovations in knowledge graph processing and integration.