Knowing the No-match: Entity Alignment with Dangling Cases

Abstract

This paper studies a new problem setting of entity alignment for knowledge graphs (KGs). Since KGs possess different sets of entities, there could be entities that cannot find alignment across them, leading to the problem of dangling entities. As the first attempt to this problem, we construct a new dataset and design a multi-task learning framework for both entity alignment and dangling entity detection. The framework can opt to abstain from predicting alignment for the detected dangling entities. We propose three techniques for dangling entity detection that are based on the distribution of nearest-neighbor distances, i.e., nearest neighbor classification, marginal ranking and background ranking. After detecting and removing dangling entities, an incorporated entity alignment model in our framework can provide more robust alignment for remaining entities. Comprehensive experiments and analyses demonstrate the effectiveness of our framework. We further discover that the dangling entity detection module can, in turn, improve alignment learning and the final performance. The contributed resource is publicly available to foster further research.

Prix-LM: Pretraining for Multilingual Knowledge Base Construction

“…In another line of research, multilingual KG embeddings (Chen et al, 2017Sun et al, 2020aSun et al, , 2021 have been developed to support cross-KG knowledge alignment and link prediction. Such methods produce a unified embedding space that allows link prediction in a target KG to be processed based on the aligned prior knowledge in other KGs (Chen et al, 2020).…”

Semi-constraint Optimal Transport for Entity Alignment with Dangling Cases

“…To detect the identical entities, prior works of entity alignment mainly focus on embeddinglevel alignment by learning a representation vector for each entity, so that the similar entities can locate closely to each others in the embedding spaces (Chen et al, 2017;Sun et al, 2018;Wang et al, 2018;Zhu et al, 2021a;Liu et al, 2021;Lin et al, 2021). These embedding-based EA methods always require an impractical assumption, that for each entity there always exists a counterpart in the other KG (Sun et al, 2021). Moreover, the embedding-level EA methods evaluate their performances only on the existing entity pairs between the testing KGs.…”

“…Moreover, the embedding-level EA methods evaluate their performances only on the existing entity pairs between the testing KGs. However, in various real-world scenarios, the identical pair of one entity is not guaranteed to exist between different KGs (Zhao, 2020;Sun et al, 2021), which limits the application range of embedding-based EA methods.…”

“…Recently, the aforementioned unmatchable entities in KGs have been recognized as the dangling entities (Sun et al, 2021). To improve the EA perfomance, it is necessary to identify the dangling entities and then align the remaining matchable entities to their counterparts.…”

“…For example, a dangling entity could have an available distance (smaller than the threshold) to a target entity, while the target one has even closer distance to the ground-truth counterpart. Sun et al (2021) constructs a labeled training set of dangling entities and trains a neural classifier to predict whether each entity is matchable. However, this supervised method has limited practicability for the difficulty of creating training sets in realworld large-scale KGs.…”