From Alignment to Assignment: Frustratingly Simple Unsupervised Entity Alignment

Abstract

Cross-lingual entity alignment (EA) aims to find the equivalent entities between crosslingual KGs (Knowledge Graphs), which is a crucial step for integrating KGs. Recently, many GNN-based EA methods are proposed and show decent performance improvements on several public datasets. However, existing GNN-based EA methods inevitably inherit poor interpretability and low efficiency from neural networks. Motivated by the isomorphic assumption of GNN-based methods, we successfully transform the cross-lingual EA problem into an assignment problem. Based on this re-definition, we propose a frustratingly Simple but Effective Unsupervised entity alignment method (SEU) without neural networks. Extensive experiments have been conducted to show that our proposed unsupervised approach even beats advanced supervised methods across all public datasets while having high efficiency, interpretability, and stability.

Multi-depth Joint Entity Alignment Based on Textual InformationMulti-depth Joint Entity Alignment Based on Textual Information

“…Character-level entity name vectors: Word-level entity name vectors perform poorly on EA tasks due to the out-of-vocabulary (OOV) problem. To this end, many methods have also explored the acquisition of character-level features, using neural networks to extract character-level features of entities [25] [26]. In our model, in order to keep the model concise and efficient, we use the character binary model to randomly generate the character-level entity name vector char H after translating the entity name.…”

Cybersecurity Entity Alignment via Masked Graph Attention Networks

“…PARIS (Suchanek et al, 2011) is a conventional method that align entities by probabilistic reasoning. Several studies (Mao et al, 2021; have noted that PARIS outperforms many "advanced" EA models. PRASE (Qi et al, 2021) is an extension of PARIS, which incorporates the translation-based EA models to boost the alignment decisions.…”

“…Further, (Pei et al, 2020) proposes reinforced training strategy to achieve noise-aware entity alignment; (Yan et al, 2021) proposes topology-invariant gates to dynamically align evolving knowledge graphs. By applying graph isomorphism, (Mao et al, 2021) learns a permutation matrix that transforms one KG to another by reordering the entity node indices. Domain-specific entity linking.…”

“…The security information is always timely updated, and it is unrealistic to retrain the KG embedding model on the whole augmented graph each time new security entities (e.g., vulnerabilities) are revealed. (Mao et al, 2021) regards entity alignment as an assignment problem between two isomorphic graphs by reordering the entity node indices. However, it's difficult to be adapted for security entity alignment because security KGs crafted from different repositories diverge in graph topology, and only a subset of entities on both KGs can be aligned.…”

Enhancing Cross-Lingual Entity Alignment in Knowledge Graphs through Structure Similarity Rearrangement

“…However, when considering a specific node, the entity selected based on distance alone may not be the optimal cross-lingual counterpart, as it does not take into account the structural similarity between nodes. We can draw inspiration from SEU [ 19 ], which transformed cross-lingual knowledge graph entity alignment into an assignment task, as the original knowledge graph and the target knowledge graph share a similar graph structure, i.e., their adjacency matrices are similar. In this paper, we propose a cross-lingual entity alignment method based on local structural rearrangement.…”