Joint Modeling of Local and Global Semantics for Contrastive Entity Disambiguation

Abstract

Entity disambiguation (ED) is a critical natural language processing (NLP) task that involves identifying and linking entity mentions in the text to their corresponding real-world entities in reference knowledge graphs (KGs). Most existing efforts perform ED by firstly learning the representations of mention and candidate entities using a variety of features and subsequently assessing the compatibility between mention and candidate entities as well as the coherence between entities. Despite advancements in the field, the limited textual descriptions of mentions and entities still lead to semantic ambiguity, resulting in sub-optimal performance for the entity disambiguation task. In this work, we propose a novel framework LogicED, which considers both Local and global semantics for contrastive Entity Disambiguation. Specifically, we design a local contextual module, which utilizes a candidate-aware self-attention (CASA) model and the contrastive learning strategy, to learn robust and discriminative contextual embeddings for both mentions and candidate entities. Furthermore, we propose a global semantic graph module that takes into account both the local mention-entity compatibility and the global entity-entity coherence to optimize the entity disambiguation from a global perspective. Extensive experiments on benchmark datasets demonstrate that our proposed framework surpasses the state-of-the-art baselines.