The Unreliability of Explanations in Few-Shot for Textual Reasoning
Nov 28, 2022
Speakers
Greg Durrett
Speaker · 0 followers
About
Does prompting a large language model like GPT-3 with explanations improve in-context learning? We study this question specifically on two NLP tasks that involve reasoning over text, namely question answering and natural language inference. For these tasks, we find that including explanations GPT-3's prompt and having the model generate them only mildly improves accuracy over standard few-shot learning, contrary to recent results on symbolic reasoning tasks. Moreover, explanations generated by GPT-3 may not entail the predictions nor be factually grounded in the input, even on simple tasks with extractive explanations. However, these flawed explanations can still be useful as a way to verify GPT-3's predictions post-hoc. Through analysis in three settings, we show that explanations judged as good by humans—those that are logically consistent with the input and the prediction—usually cooccur with more accurate predictions. Following these observations, we present a framework for calibrating model predictions based on the reliability of the explanations. We train calibrators using automatically extracted scores that approximately assess the reliability of explanations, which helps improve performance across three different datasets.Does prompting a large language model like GPT-3 with explanations improve in-context learning? We study this question specifically on two NLP tasks that involve reasoning over text, namely question answering and natural language inference. For these tasks, we find that including explanations GPT-3's prompt and having the model generate them only mildly improves accuracy over standard few-shot learning, contrary to recent results on symbolic reasoning tasks. Moreover, explanations generated by G…
Organizer
NeurIPS 2022
Account · 961 followers
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
On the Epistemic Limits of Personalized Prediction
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
AutoML for Climate Change: A Call to Action
Watch later
Renbo Tu, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
When is Offline Two-Player Zero-Sum Markov Games Solvable?
Watch later
Qiwen Cui, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
ESCADA: Efficient Safety and Context Aware Dose Allocation for Precision Medicine
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Reconstruction on Trees and Low-Degree Polynomials
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
ConserWeightive Behavioral Cloning for Reliable Offline Reinforcement Learning
Watch later
Tung Nguyen, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%