Don’t Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biases

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4069–4082, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4069 Don’t Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biases Christopher Clark∗, Mark Yatskar†, Luke Zettlemoyer∗ ∗Paul G. Allen School of CSE, University of Washington {csquared, lsz}@cs.uw.edu †Allen Institute for Artificial Intelligence, Seattle WA marky@allenai.org Abstract State-of-the-art models often make use of su- perficial patterns in the data that do not gener- alize well to out-of-domain or adversarial set- tings. For example, textual entailment mod- els often learn that particular key words im- ply entailment, irrespective of context, and vi- sual question answering models learn to pre- dict prototypical answers, without considering evidence in the image. In this paper, we show that if we have prior knowledge of such biases, we can train a model to be more robust to do- main shift. Our method has two stages: we (1) train a naive model that makes predictions ex- clusively based on dataset biases, and (2) train a robust model as part of an ensemble with the naive one in order to encourage it to focus on other patterns in the data that are more likely to generalize. Experiments on five datasets with out-of-domain test sets show significantly im- proved robustness in all settings, including a 12 point gain on a changing priors visual ques- tion answering dataset and a 9 point gain on an adversarial question answering test set. 1 Introduction While recent neural models have shown remark- able results, these achievements have been tem- pered by the observation that they are often ex- ploiting dataset-specific patterns that do not gen- eralize well to out-of-domain or adversarial set- tings. For example, entailment models trained on MNLI (Bowman et al., 2015) will guess an an- swer based solely on the presence of particular keywords (Gururangan et al., 2018) or whether sentences pairs contain the same words (Mc- Coy et al., 2019), while QA models trained on SQuAD (Rajpurkar et al., 2016) tend to select text near question-words as answers, regardless of con- text (Jia and Liang, 2017). We refer to these kinds of superficial patterns as bias. Models that rely on bias can perform well on in-domain data, but are brittle and easy to fool (e.g., SQuAD models are easily distracted by irrelevant sentences that contain many question words). Recent concern about dataset bias has led researchers to re-examine many popular datasets, resulting in the discovery of a wide variety of bi- ases (Agrawal et al., 2018; Anand et al., 2018; Min et al., 2019; Schwartz et al., 2017). In this paper, we build on these works by show- ing that, once a dataset bias has been identified, we can improve the out-of-domain performance of models by preventing them from making use of that bias. To do this, we use the fact that these bi- ases can often be explicitly modelled …

Content truncated to 500 words