CausalSense: Leveraging common sense knowledge and LLMs for joint event extraction and relation classification

Event Relation Extraction (ERE) aims to identify and classify semantic relationships between events expressed in texts. While existing work has mainly addressed temporal or simple causal links, fine-grained causal relations such as enable, prevent, and intend remain insufficiently explored, partly due to limited and imbalanced labeled datasets.

We present a novel framework that leverages large language models (LLMs) and common-sense knowledge to jointly perform event extraction and relation classification. Our contribution includes (1) the creation of the CausalSense large-scale dataset containing more than 500k sentences from news data and common sense knowledge extracted from ATOMIC, and enriched synthetically; and (2) the evaluation of multiple architectures, including transformer-based models and end-to-end multitask systems for extracting fine-grained causal relationships . Experimental results show that our best-performing model achieves a 32.3% improvement in average F1-score over the current state of the art. The integration of common sense knowledge substantially enhances fine-grained causal relation detection. The CausalSense dataset, along with our code and models, is released as open source to support future research on causal event relationship extraction.