KEPT: Knowledge Enhanced Prompt Tuning for event causality identification

Abstract

Event causality identification (ECI) aims to identify causal relations of event mention pairs in text. Despite achieving certain accomplishments, existing methods are still not effective due to the following two issues: (1) the lack of causal reasoning ability, imposing restrictions on recognizing implicit causal relations; (2) the significant gap between fine-tuning and pre-training, which hinders the utilization of pre-trained language models (PLMs). In this paper, we propose a novel Knowledge Enhanced Prompt Tuning (KEPT) framework for ECI to address the issues mentioned above. Specifically, this method leverages prompt tuning to incorporate two kinds of knowledge obtained from external knowledge bases (KBs), including background information and relational information, for causal reasoning. To introduce external knowledge into our model, we first convert it to textual descriptions, then design an interactive attention mechanism and a selective attention mechanism to fuse background information and relational information, respectively. In addition, to further capture implicit relations between events, we adopt the objective from knowledge representation learning to jointly optimize the representations of causal relations and events. Experiment results on two widely-used benchmarks demonstrate that the proposed method outperforms the state-of-the-art models.