Weak Reward Model Transforms Generative Models into Robust Causal Event Extraction Systems
This work addresses the problem of robust causal event extraction for NLP researchers, offering an incremental improvement in evaluation and training methods.
The paper tackles the challenge of evaluating causal event extraction due to ambiguous cause-effect boundaries by training evaluation models that approximate human judgment and using them for reinforcement learning to align extraction models with human preference, achieving high agreement and exploring transfer across datasets to reduce reliance on annotated data.
The inherent ambiguity of cause and effect boundaries poses a challenge in evaluating causal event extraction tasks. Traditional metrics like Exact Match and BertScore poorly reflect model performance, so we trained evaluation models to approximate human evaluation, achieving high agreement. We used them to perform Reinforcement Learning with extraction models to align them with human preference, prioritising semantic understanding. We successfully explored our approach through multiple datasets, including transferring an evaluator trained on one dataset to another as a way to decrease the reliance on human-annotated data. In that vein, we also propose a weak-to-strong supervision method that uses a fraction of the annotated data to train an evaluation model while still achieving high performance in training an RL model. Our code is available at https://github.com/oyarsa/event_extraction/tree/causal-event-extraction.