CED-BGFN: Chinese Event Detection via Bidirectional Glyph-Aware Dynamic Fusion Network

Event Detection is an essential task in information extraction. However, most existing studies on event detection are designed for English text. There is still a lack of efficient algorithm for Chinese event detection, which is expected to be greatly improved. Recent work has shown that enhanced text representation, such as introducing glyph information, can significantly improve downstream tasks in natural language processing. In this paper, we propose a novel method for Chinese Event Detection via Bidirectional Glyph-aware Dynamic Fusion Network, called CED-BGFN. We use two representations: glyph-aware information and pre-trained language model. To integrate the heterogeneous representation modules, we propose a creative fusion network Bidirectional Glyph-aware Fusion Network, named BGFN. Considering the dynamic interaction of the two expressions, BGFN adaptively learns the fusion weights for the downstream event detection task. We conduct extensive experiments to investigate the validity of the proposed method on the ACE 2005 Chinese corpus. Results demonstrate that compared with the previous state-of-the-art methods, our approach obtains transcendent performance in both event trigger identification task and classification task, with an increase of 5.48 (7.46%) and 5.03 (7.1%) in F1-score, respectively.