Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery

• DOI: 10.48550/arxiv.2403.07955
• Publication Date: 2024-03-12
• Authors: Linan Yue; Qi Liu; Yichao Du; Wang Li; Weibo Gao; Yanqing An
• Keywords: rationalization; Boosting

The remarkable success in neural networks provokes the selective rationalization. It explains prediction results by identifying a small subset of inputs sufficient to support them. Since existing methods still suffer from adopting shortcuts data compose rationales and limited large-scale annotated human, this paper, we propose Shortcuts-fused Selective Rationalization (SSR) method, which boosts rationalization discovering exploiting potential shortcuts. Specifically, SSR first designs discovery approach detect several Then, introducing identified shortcuts, two strategies mitigate problem utilizing rationales. Finally, develop augmentations close gap number Extensive experimental on real-world datasets clearly validate effectiveness our proposed method.