Semantically-Informed Domain Adaptation for Named Entity Recognition

Mariya Borovikova, Arnaud Ferré, Robert Bossy, Mathieu Roche, Claire Nédellec

Abstract: Named Entity Recognition (NER) is an important task in Natural Language Processing that involves identifying entities in unstructured text. State-of-the-art NER methods often require extensive manual labeling for training. To bridge this gap, this paper introduces a domain adaptation technique that leverages semantic information about entity types using Sentence-BERT embeddings of their textual descriptions. We conduct experiments across various datasets from both general and biological domains, evaluating our approach in standard and zero-shot settings. Our experiences demonstrate the effectiveness of our method, which outperforms existing zero-shot techniques on certain datasets. Our findings underscore the importance of accurate semantic representations for entity types. This paper contributes to the advancement of zero-shot domain adaptation for NER and opens avenues for future research in improving NER systems’ adaptability and performance across diverse domains.