Identifying Spurious Correlations for Robust Text Classification

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3431–3440 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3431 Identifying Spurious Correlations for Robust Text Classification Zhao Wang Illinois Institute of Technology zwang185@hawk.iit.edu Aron Culotta Tulane University aculotta@tulane.edu Abstract The predictions of text classifiers are often driven by spurious correlations – e.g., the term Spielberg correlates with positively reviewed movies, even though the term itself does not semantically convey a positive sentiment. In this paper, we propose a method to distinguish spurious and genuine correlations in text clas- sification. We treat this as a supervised classi- fication problem, using features derived from treatment effect estimators to distinguish spu- rious correlations from “genuine” ones. Due to the generic nature of these features and their small dimensionality, we find that the approach works well even with limited train- ing examples, and that it is possible to trans- port the word classifier to new domains. Ex- periments on four datasets (sentiment classifi- cation and toxicity detection) suggest that us- ing this approach to inform feature selection also leads to more robust classification, as mea- sured by improved worst-case accuracy on the samples affected by spurious correlations. 1 Introduction Text classifiers often rely on spurious correlations. For example, consider sentiment classification of movie reviews. The term Spielberg may be cor- related with the positive class because many of director Steven Spielberg’s movies have positive reviews. However, the term itself does not indi- cate a positive review. In other words, the term Spielberg does not cause the review to be positive. Similarly, consider the problem of toxicity classi- fication of online comments. Terms indicative of certain ethnic groups may be associated with the toxic class because those groups are often victims of harassment, not because those terms are toxic themselves. Oftentimes, such spurious correlations do not harm prediction accuracy because the same cor- relations exist in both training and testing data (under the common assumption of i.i.d. sam- pling). However, they can still be problematic for several reasons. For example, under dataset shift (Quionero-Candela et al., 2009), the testing distribution differs from the training distribution. E.g., if Steven Spielberg makes a new, bad movie, the sentiment classifier may incorrectly classify the reviews as positive because they contain the term Spielberg. Additionally, if the spurious cor- relations indicate demographic attributes, then the classifier may suffer from issues of algorithmic fair- ness (Kleinberg et al., 2018). For example, the tox- icity classifier may unfairly over-predict the toxic class for comments discussing certain demographic groups. Finally, in settings where classifiers must explain their decisions to humans, such spurious correlations can reduce trust in autonomous sys- tems (Guidotti et al., 2018). In this paper, we propose a method to distinguish spurious correlations, like Spielberg, from genuine correlations, like wonderful, which more reliably indicate the class label. Our approach is to treat this as a separate classification task, using features drawn from treatment effect estimation approaches that isolate the impact each word has on the class label, …

Content truncated to 500 words