Counterfactual Misleading Inference Generation via Reinforced Proximal Policy Optimization

Abstract In this paper, we propose to investigate Misleading Inference Generation, a new natural language generation task. The goal is to generate a counterfactual sentence for a context and a factual sentence. This paper proposes a framework based on BART and reinforcement learning for the misleading inference generation task. The experiment results show our model significantly outperforms the compared models, making our solution a necessary and strong baseline for future research toward misleading inference generation.