Can We Improve Model Robustness through Secondary Attribute Counterfactuals?

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4701–4712 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4701 Can We Improve Model Robustness through Secondary Attribute Counterfactuals? Ananth Balashankar �Xuezhi Wang �Ben Packer �Nithum Thain �Ed H. Chi �Alex Beutel Google AI {ananthbshankar,xuezhiw,bpacker,nthain,edchi,alexbeutel}@google.com Abstract Developing robust NLP models that perform well on many, even small, slices of data is a significant but important challenge, with implications from fairness to general relia- bility. To this end, recent research has ex- plored how models rely on spurious correla- tions, and how counterfactual data augmenta- tion (CDA) can mitigate such issues. In this paper we study how and why modeling coun- terfactuals over multiple attributes can go sig- nificantly further in improving model perfor- mance. We propose RDI, a context-aware methodology which takes into account the im- pact of secondary attributes on the model’s predictions and increases sensitivity for sec- ondary attributes over reweighted counterfac- tually augmented data. By implementing RDI in the context of toxicity detection, we find that accounting for secondary attributes can signifi- cantly improve robustness, with improvements in sliced accuracy on the original dataset up to 7% compared to existing robustness meth- ods. We also demonstrate that RDI generalizes to the coreference resolution task and provide guidelines to extend this to other tasks. 1 Introduction How can we build NLP models that perform well over many slices, albeit sometimes small slices, of our data? Developing models that are robust in their performance is important for trusting these models to work well in diverse, unexpected set- tings. As a concrete running example in this paper, we will consider the task of toxicity detection: us- ing a model to predict if a comment is toxic or not (Dixon et al., 2018). In this application, for example, it is often important to ensure that models are accurate over slices of data referring to differ- ent demographic groups, as has been raised across machine learning fairness research (Hardt et al., 2016a; Blodgett et al., 2020). One significant focus of research on how to im- prove model robustness has been addressing spu- rious correlations and improving counterfactual robustness. That is, researchers have found that models often rely on features or attributes that are only spuriously correlated with the task and ac- curacy often drops when models are evaluated on counterfactual data that perturbs those attributes (Jia and Liang, 2017). To return to our example of toxicity detection, a model may learn that certain identity tokens are correlated with toxicity, but that could decrease accuracy for non-toxic comments with those terms (Dixon et al., 2018; Garg et al., 2018). Recent work has explored how counter- factual generation techniques can be used to form general checklists to test for model biases (Ribeiro et al., 2020; Bender and Koller, 2020), often com- posing many sub-problems which are hard to solve formally. Similarly, a wide breadth of research has studied how to train models to be more ro- bust. We focus on one …

Content truncated to 500 words