A Multi-granularity Similarity Enhanced Model for Implicit Event Argument Extraction

Yanhe Fu, Yi Liu, Yanan Cao, Yubing Ren, Qingyue Wang, Fang Fang, Cong Cao

Abstract: Implicit Event Argument Extraction (Implicit EAE) aims to extract the document event arguments given the event type. Influenced by the document length, the arguments scattered in different sentences can potentially lead to two challenges during extraction: long-range dependency and distracting context. Existing works rely on the contextual capabilities of pre-trained models and semantic features but lack a straightforward solution for these two challenges and may introduce noise. In this paper, we propose a Multi-granularity Similarity Enhanced Model to solve these issues. Specifically, we first construct a heterogeneous graph to incorporate global information, then design a supplementary task to tackle the above challenges. For long-range dependency, span-level enhancement can directly close the semantic distance between trigger and arguments across sentences; for distracting context, sentence-level enhancement makes the model concentrate more on effective content. Experimental results on RAMS and WikiEvents demonstrate that our proposed model can obtain state-of-the-art performance in Implicit EAE.