This study investigates a self-rationalization framework consisting of a cooperative game between a generator and a predictor. The generator extracts the most informative parts of the raw input, and the predictor uses the selected subset as input. The generator and predictor are cooperatively trained to maximize prediction accuracy. This paper first reveals a potential problem: the cooperative game can unintentionally introduce sampling bias during rationalization extraction. Specifically, the generator can unintentionally generate false correlations between selected rationalization candidates and labels, even if they are semantically unrelated in the original dataset. We then explain the origin of this bias using detailed theoretical analysis and empirical evidence. Our results suggest ways to examine these correlations using attacks, and based on these findings, we provide additional guidance to prevent the predictor from learning correlations. Experiments on six text classification datasets and two graph classification datasets using three network architectures (GRUs, BERT, and GCN) demonstrate that the proposed method significantly outperforms recent rationalization methods and achieves results comparable to or better than the representative LLM (llama3.1-8b-instruct).
