Towards a word-granularity paradigm for Chinese event detection: Targeting long-tail challenges in syntax and semantics

Event detection (ED) seeks to identify and categorize event triggers in unstructured text. A major challenge in Chinese ED is word boundary ambiguity due to the lack of explicit delimiters. Although word granularity provides clearer boundaries and shorter token sequences, it is underexplored due to challenges in granularity alignment and long-tail issues at the syntactic and semantic levels. To address these challenges, we propose Ant icipatory prototype and Syntactic-structure Enhanced E vent D etection (AntED), the first ED framework at word granularity. AntED incorporates C ontrastive L earning-based O ut-of- V ocab word representation (CLOV) module, which can generate high-quality embeddings for OOV words in a plug-and-play manner, achieving unified word granularity. We further design a T ail-aware H eterogeneous G raph AT tention Network (THGAT), ensuring equal representation of low-frequency syntactic relations. Moreover, prompt-based An ticipatory P rototype (AnP) learning is used to model event category prototypes and to enhance performance in semantic-scarcity settings. Extensive experiments on three datasets demonstrate that AntED achieves state-of-the-art performance. Especially in the trigger identification subtask, AntED outperforms other methods by over 2% F1 on DuEE and FewFC, and by more than 6% on ACE2005 compared to LLaMA3. These findings highlight the effectiveness of word-granularity ED and encourage further research into its advantages.