DE-ESD: Dual encoder-based entity synonym discovery using pre-trained contextual embeddings

Extracting synonymous entities from unstructured text is important for enhancing entity-dependent applications such as web searches and question-answering systems. Existing work primarily falls into two types: statistics- based and deep learning-based. However, these approaches often fail to discern fine semantic nuances among entity mentions and are prone to cumulative errors; thus, they inadequately represent the holistic semantics of entity synonym sets. To address these limitations, this paper introduces a novel framework, Dual Encoder-based Entity Synonym Discovery (DE-ESD). The proposed method initially uses pre-trained language models to extract multiperspective contextual embeddings of entity mentions. Then, it employs a dual encoder architecture to differentiate features between an established entity synonym set and a pseudo-set-created by adding a candidate entity mention to the synonym set. A set scorer evaluates the quality scores of both sets. By leveraging the trained dual encoder and the set scorer, DE-ESD can implement an efficient online algorithm for mining new entity synonym sets for open text streams. The experimental results obtained on two real-world datasets (NYT and Wiki) demonstrate the effectiveness of DE-ESD. Furthermore, we investigated the impact of different pre-trained language models on DE-ESD performance, particularly their ability to extract effective contextual embeddings.