TaIncBC: Topic-aware In-context Prompt with Bias Calibration for Event Causality Identification

Abstract Event Causality Identification (ECI) is to determine whether there exists a causal relation between two event mentions. Recently, the prompt learning paradigm has been applied in the ECI task to leverage a large-scale pre-trained language model through carefully designed prompts, yet their contexts do not describe causal relations. In this paper, we propose to include some topic-aware demonstration samples with their ground-truth labels into the query prompt to form an in-context input prompt, so as to facilitate PLM’s understanding and comprehending of universal causal relations. Although including demo samples into prompts may enrich in-context causality information, care must be taken for potential bias that could also be introduced into the event prediction model by such demo samples. To address this challenge, we further design a nearest neighbour-based bias calibration to help alleviating demonstration deviations for our topic-aware in-context prompt learning. Our model, called TaIncBC, is experimented on the widely used EventStoryLine and Causal-TimeBank corpus and results validate our design objective in terms of significantly performance improvements over the state-of-the-art algorithms.