How rationals boost textual entailment modeling: Insights from large language models

Abstract

This study introduces an innovative methodology for rationale-based distillation in textual entailment. Central to our methodology is the use of diverse and deep rationale types generated by large language models, eliminating the need for explicit feature engineering between text-hypothesis pairs. Through extensive experimentation, we demonstrate the effectiveness of our rationale-enhanced distillation process, underpinned by a comprehensive study of rationales. Remarkably, our model, which utilizes the most impactful rationales and operates with 795 times fewer parameters, exhibits competitive performance, especially in contexts limited by resource availability. Specifically, our model significantly surpasses RoBERTa, FLAN-T5 Large, and GPT-3.5 in performance. Our findings underscore the potential of rationale-based approaches, which improve textual entailment modeling and pave the way for future research. These techniques could be applied to other areas of NLP and beyond. This study’s contributions are poised to benefit researchers and practitioners seeking to leverage the power of large language models in augmenting reasoning capabilities with rationales. The novelty of our approach lies in its ability to deliver high performance with significantly fewer computational resources, making it a valuable advancement in the field.