Event Causality Identification via Prompt Learning with Knowledge and Memory

Fei Cai, Wanyu Chen, Shen Wang, Guowang Li, Weijie Chen, Junxian Liu

Abstract: Event Causality Identification (ECI) aims to identify causality between events, where the issue of data scarcity is challenging. Advanced ECI methods focus on introducing external knowledge or mining the causal clues. When introducing external knowledge, they typically ignore the structural information and face a semantic bias. In addition, they often overlook the exploration of internal knowledge, fully utilizing the known causal information. Thus, we propose a prompt learning method (PKC) integrating external knowledge and internal causal memory to address the above issues. It mainly consists of three modules: a prompt learning module detecting the causal cues and generating an event representation; a knowledge fusion module with a novel interactive fusion mechanism encoding the external knowledge structurally as well as solving the semantic deviation; and a causal memory processing module enriching the event representations with the inside causal memories to further improve the quality of causal recognition. Experiments on the public EventStoryLine and Causal-TimeBank datasets show that PKC outperforms the state-of-the-art models. We also find that PKC is insensitive to the imbalanced causal datasets under different numbers of memories used and is robust within a certain range of forgetting the rates of causal memories.

External IDs:dblp:conf/ideal/CaiCWLCL25