Domain-Specific Entity Recognition as Token-Pair Relation Classification

Abstract

Named Entity Recognition (NER) is a fundamental but crucial task in natural language understanding, aiming at identifying entity mentions from free text. Current methods mainly use sequence-labeling and span-based models, where the former ignores the importance of token interaction, and the latter pays little attention to the global inter-dependency among entity tokens. In this work, we propose a novel NER model that consists of two branches: a Token-Pair Interaction Module (TPIM) and a U-shaped Network. The TPIM models head-tail relations between token pairs while capturing intrinsic token connectivity within entity boundaries. The U-shaped Network is employed to capture the contextual dependency in the token-pair relation matrix. Furthermore, we build a typical domain-specific entity dataset CCAEE based on real-world applications in the chemical accident domain. The experimental results on CCAEE and CLUENER datasets demonstrate the effectiveness of our proposed model.

Keywords

• Natural language understanding
• information extraction
• named entity recognition
• boundary detection