FLamE: Few-shot Learning from Natural Language Explanations
This addresses the challenge of effectively leveraging explanations for few-shot learning in natural language inference, though it is incremental as it builds on prior work on explanation-based methods.
The paper tackles the problem of limited utility of natural language explanations in improving classification by proposing FLamE, a two-stage few-shot learning framework that generates explanations with GPT-3 and finetunes a smaller model, resulting in accuracy increases of 17.6% over GPT-3 Babbage and 5.7% over GPT-3 Davinci on e-SNLI.
Natural language explanations have the potential to provide rich information that in principle guides model reasoning. Yet, recent work by Lampinen et al. (2022) has shown limited utility of natural language explanations in improving classification. To effectively learn from explanations, we present FLamE, a two-stage few-shot learning framework that first generates explanations using GPT-3, and then finetunes a smaller model (e.g., RoBERTa) with generated explanations. Our experiments on natural language inference demonstrate effectiveness over strong baselines, increasing accuracy by 17.6% over GPT-3 Babbage and 5.7% over GPT-3 Davinci in e-SNLI. Despite improving classification performance, human evaluation surprisingly reveals that the majority of generated explanations does not adequately justify classification decisions. Additional analyses point to the important role of label-specific cues (e.g., "not know" for the neutral label) in generated explanations.