Hypothetical-Deductive Reasoning for Event Causality Identification

ACL ARR 2025 May Submission1406 Authors

Abstract: Event Causality Identification (ECI) is the task of identifying causal relations between two events. Most works mainly enhance event encoding with pre-trained language models (PLMs), often neglecting the implicit and long-text reasoning capabilities needed for ECI tasks. Large language models (LLMs) have recently revealed substantial reasoning potential through chain-of-thought (CoT). Inspired by Pearl's Causal Hierarchy, we first introduce CoT into the ECI task and propose Causal Progressive Reasoning CoT (CPR). CPR uses a progressive reasoning approach, guiding the model step by step to explore the causal relation between two events. More importantly, we find that CoT may generate incorrect intermediate steps that propagate to the next ones, leading to error results. To deal with this problem, we propose a Hypothetical-Deductive Reasoning framework (HYDRO). HYDRO is based on hypothetical-deductive reasoning, where each step is independently reasoned. Extensive experiments have demonstrated that our methods achieve state-of-the-art performance (17.8\% and 6.8\% F1 score gains on EventStoryLine and Causal-TimeBank) on two benchmark datasets. Additionally, it exhibits significant advantages only using Flan-T5-Base (250M) in zero-shot settings.

Paper Type: Long

Research Area: Syntax: Tagging, Chunking and Parsing

Research Area Keywords: Information Extraction, Efficient/Low-Resource Methods for NLP

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English

Submission Number: 1406