Describing Differences between Text Distributions with Natural Language

Abstract

How do two distributions of text differ? Humans are slow at answering this, since discovering patterns might require tediously reading through hundreds of samples. We propose to automatically summarize the differences by "learning a natural language hypothesis": given two distributions D 0 and D 1 , we search for a description that is more often true for D 1 , e.g., "is military-related." To tackle this problem, we fine-tune GPT-3 to propose descriptions with the prompt: "[samples of D 0 ] + [samples of D 1 ] + the difference between them is ". We then re-rank the descriptions by checking how often they hold on a larger set of samples with a learned verifier. On a benchmark of 54 real-world binary classification tasks, while GPT-3 Curie (13B) only generates a description similar to human annotation 7% of the time, the performance reaches 61% with fine-tuning and reranking, and our best system using GPT-3 Davinci (175B) reaches 76%. We apply our system to describe distribution shifts, debug dataset shortcuts, summarize unknown tasks, and label text clusters, and present analyses based on automatically generated descriptions. 1

Active Programming by Example with a Natural Language Prior

“…Scalable Oversight As AI systems become more capable of generating candidate responses, an emerging line of research supervises AI systems by providing preferences over AI-generated candidates rather than providing human demonstrations (Stiennon et al, 2020;Wiegreffe et al, 2021;Askell et al, 2021;Liu et al, 2022;Ouyang et al, 2022). Therefore, to supervise AI to perform more complex tasks, it becomes increasingly important to determine human preferences over model outputs that are expensive to verify, such as full-book summaries or natural language descriptions of distributional properties (Amodei et al, 2016;Wu et al, 2021;Zhong et al, 2022). Our work presents a strategy to re-weight complex outputs (programs) from an AI system by asking simple informative questions of annotators who do not have to understand the outputs directly.…”