An Analysis of Dataset Overlap on Winograd-Style Tasks

TL;DR

This paper investigates how dataset overlap influences model performance on Winograd-Style tasks, revealing that high overlap inflates accuracy and introducing a new large, less-overlapping dataset for more robust evaluation.

Contribution

The authors analyze the impact of dataset overlap on Winograd-Style task performance and create the KnowRef-60K dataset with minimal overlap to improve evaluation robustness.

Findings

High overlap correlates with higher model accuracy.

Models perform significantly worse on low-overlap instances.

KnowRef-60K is the largest low-overlap dataset for WSC-style reasoning.

Abstract

The Winograd Schema Challenge (WSC) and variants inspired by it have become important benchmarks for common-sense reasoning (CSR). Model performance on the WSC has quickly progressed from chance-level to near-human using neural language models trained on massive corpora. In this paper, we analyze the effects of varying degrees of overlap between these training corpora and the test instances in WSC-style tasks. We find that a large number of test instances overlap considerably with the corpora on which state-of-the-art models are (pre)trained, and that a significant drop in classification accuracy occurs when we evaluate models on instances with minimal overlap. Based on these results, we develop the KnowRef-60K dataset, which consists of over 60k pronoun disambiguation problems scraped from web data. KnowRef-60K is the largest corpus to date for WSC-style common-sense reasoning and exhibits a significantly lower proportion of overlaps with current pretraining corpora.