An Analysis of Dataset Overlap on Winograd-Style Tasks

Proceedings of the 28th International Conference on Computational Linguistics, pages 5855–5865 Barcelona, Spain (Online), December 8-13, 2020 5855 An Analysis of Dataset Overlap on Winograd-Style Tasks Ali Emami Mila/McGill University ali.emami@mail.mcgill.ca Adam Trischler, Kaheer Suleman Microsoft Research Montreal {adam.trischler, kasulema}@microsoft.com Jackie Chi Kit Cheung Mila/McGill University jcheung@cs.mcgill.ca 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. 1 Introduction The original purpose of the Winograd Schema Challenge was to serve as an alternative Turing test to evaluate an automatic system’s capacity for common-sense inference (Levesque et al., 2011). As an example: (1) a. Jim yelled at Kevin because he was so upset. (Answer: Jim) b. Jim comforted Kevin because he was so upset. (Answer: Kevin) For a number of years, models struggled to exceed chance-level performance (Kruengkrai et al., 2014; Sharma et al., 2015; Peng et al., 2015; Liu et al., 2016). The WSC task is carefully controlled, such that heuristics involving syntactic and semantic cues were ineffective, and the common-sense knowledge re- quired to correctly resolve its test instances make it particularly difficult for statistical systems to model. More recently, however, the advent of deep bidirectional transformers (e.g., BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019)) pretrained on massive amounts of data has led to near-human-level perfor- mance (Kocijan et al., 2019; Ye et al., 2019; Ruan et al., 2019). Various works have lately re-examined the challenges of the WSC, leading to the proposal of more difficult, larger variants, data and model debi- asing methods, and evaluation protocols that clarify which types of instances models excel on and which they struggle with (Trichelair et al., 2019; Emami et al., 2019; Sakaguchi et al., 2020; Abdou et al., 2020). However, little attention has been paid to studying the effects and influence of pretraining data points. While recent work has included some analysis on the effects of 13-gram overlaps between pretraining and test instances for the WSC (Brown et al., 2020), a deeper look into how the degree of overlap (and how this can be defined) affects language models’ performance is critical to revealing models’ reasoning This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http://creativecommons.org/licenses/by/4.0/. 5856 and inference functions. …

Content truncated to 500 words