An Analysis of Dataset Overlap on Winograd-Style Tasks
This work addresses the problem of inflated benchmark performance due to data contamination for researchers in common-sense reasoning, though it is incremental as it builds on existing WSC-style tasks.
The paper analyzed how dataset overlap between training corpora and test instances affects performance on Winograd Schema Challenge tasks, finding that models perform significantly worse on instances with minimal overlap, and introduced the KnowRef-60K dataset with over 60k problems to reduce such overlaps.
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.