How Does Counterfactually Augmented Data Impact Models for Social Computing Constructs?

Abstract

As NLP models are increasingly deployed in socially situated settings such as online abusive content detection, it is crucial to ensure that these models are robust. One way of improving model robustness is to generate counterfactually augmented data (CAD) for training models that can better learn to distinguish between core features and data artifacts. While models trained on this type of data have shown promising out-of-domain generalizability, it is still unclear what the sources of such improvements are. We investigate the benefits of CAD for social NLP models by focusing on three social computing constructs -sentiment, sexism, and hate speech. Assessing the performance of models trained with and without CAD across different types of datasets, we find that while models trained on CAD show lower in-domain performance, they generalize better out-of-domain. We unpack this apparent discrepancy using machine explanations and find that CAD reduces model reliance on spurious features. Leveraging a novel typology of CAD to analyze their relationship with model performance, we find that CAD which acts on the construct directly or a diverse set of CAD leads to higher performance.

Potential Pitfalls With Automatic Sentiment Analysis: The Example of Queerphobic Bias

“…As mentioned above, model bias will be a particular issue when considering data from participants belonging to different communities whereby use of identity terms by one group will systematically alter their results; unlike with lexicon-based approaches, it is not always easy to identify which terms will lead to bias without testing with a data set such as ours. There is a significant body of work looking to develop less biased language models for a range of tasks (Dixon et al, 2018; Liang et al, 2020; Schick et al, 2021; Ungless et al, 2022; Webster et al, 2021; Zhao et al, 2018), for example, using counterfactually augmented data (Sen et al, 2021), which researchers with the right technical skills may be able to adopt, where they have access to the original model. However, for those who must rely on third party tools, our findings suggest marginalised individuals continue to be impacted by predictive bias despite the likely use of debiasing strategies, and in particular the least salient identities.…”

A Survey on Gender Bias in Natural Language Processing

“…Since the model observes the same scenario in the doubled (for binary gender) sentences, it can learn to abstract away from the entities to the context [Emami et al 2019]. This method has shown encouraging results in mitigating bias in contextualised word representations such as ELMo and monolingual BERT [Bartl et al 2020;de Vassimon Manela et al 2021;Sen et al 2021;, and for hate speech detection [Park et al 2018]. Nonetheless, collecting annotated lists for gender-specific pairs can be expensive, and the method essentially doubles the size of the training data.…”