Topics to Avoid: Demoting Latent Confounds in Text Classification

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4153–4163, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4153 Topics to Avoid: Demoting Latent Confounds in Text Classification Sachin Kumar♦ Shuly Wintner♣ Noah A. Smith♥♠ Yulia Tsvetkov♦ ♦Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA, USA ♣Department of Computer Science, University of Haifa, Haifa, Israel ♥Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle, WA, USA ♠Allen Institute for Artificial Intelligence, Seattle, WA, USA sachink@cs.cmu.edu, shuly@cs.haifa.ac.il nasmith@cs.washington.edu, ytsvetko@cs.cmu.edu Abstract Despite impressive performance on many text classification tasks, deep neural networks tend to learn frequent superficial patterns that are specific to the training data and do not al- ways generalize well. In this work, we ob- serve this limitation with respect to the task of native language identification. We find that standard text classifiers which perform well on the test set end up learning topical features which are confounds of the prediction task (e.g., if the input text mentions Sweden, the classifier predicts that the author’s native lan- guage is Swedish). We propose a method that represents the latent topical confounds and a model which “unlearns” confounding features by predicting both the label of the input text and the confound; but we train the two predic- tors adversarially in an alternating fashion to learn a text representation that predicts the cor- rect label but is less prone to using information about the confound. We show that this model generalizes better and learns features that are indicative of the writing style rather than the content.1 1 Introduction Text classification systems based on neural net- works are biased towards learning frequent spu- rious correlations in the training data that may be confounds in the actual classification task (Leino et al., 2019). A major challenge in building such systems is to discover features that are not just cor- related with the signals in the training data, but are true indicators of these signals, and therefore gen- eralize well. For example, Kiritchenko and Mohammad (2018) found that sentiment analysis systems im- plicitly overfit to demographic confounds, system- atically amplifying the intensity ratings of posts 1The code is available at: https://github.com/ Sachin19/adversarial-classify written by women. Zhao et al. (2017) showed that visual semantic role labeling models implic- itly capture actions stereotypically associated with men or women (e.g., women are cooking and men are fixing a faucet), and in cases of higher model uncertainty assign stereotypical labels to actions and objects, thereby amplifying social bi- ases found in the training data. We focus on the task of native language identi- fication (L1ID), which aims at automatically iden- tifying the native language (L1) of an individ- ual based on their language production in a sec- ond language (L2, English in this work). The aim of this task is to discover stylistic features present in the input that are indicative of the au- thor’s L1. However, a model …

Content truncated to 500 words