Enhancing Document Ranking with Task-adaptive Training and Segmented Token Recovery Mechanism

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 325–344 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 325 How Does Counterfactually Augmented Data Impact Models for Social Computing Constructs? Indira Sen Mattia Samory Fabian Flöck GESIS – Leibniz Institute for the Social Sciences {indira.sen, mattia.samory, fabian.floeck}@gesis.org Claudia Wagner GESIS and RWTH Aachen University claudia.wagner@gesis.org Isabelle Augenstein University of Copenhagen augenstein@di.ku.dk Abstract As NLP models are increasingly deployed in socially situated settings such as online abu- sive content detection, it is crucial to ensure that these models are robust. One way of im- proving model robustness is to generate coun- terfactually augmented data (CAD) for train- ing 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 improve- ments are. We investigate the benefits of CAD for social NLP models by focusing on three social computing constructs — sentiment, sex- ism, and hate speech. Assessing the perfor- mance 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 dis- crepancy 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 perfor- mance, we find that CAD which acts on the construct directly or a diverse set of CAD leads to higher performance. 1 Introduction Dataset design is receiving increasing attention, especially in response to concerns related to the generalizability of machine learning-based NLP models. Recent critiques argue that models trained for NLP tasks may end up “learning the dataset” rather than a particular construct (Bras et al., 2020), i.e, the intangible measure like sentiment or stance that is the ultimate goal of the learning task (Jacobs and Wallach, 2021). In particular, in the process of inferring the mapping between an input space and output space, models may learn cues in the dataset which are spuriously correlated with the construct (Schlangen, 2020). For example, senti- ment models trained on movie reviews tend to learn construct original counterfactual sentiment I thought this movie was very well put together. I thought this movie was very haphazardly put together. sexism Females should not commentate on sport AI should not commentate on sport hate speech Lets talk about the antithesis of hard work: immigrants Lets talk about the antithesis of hard work: my brother Table 1: Examples of original/counterfactual pairs for sentiment, sexism, and hate speech. As pairs of data with minimal textual differences (color-coded here) but different labels, counterfactual examples can improve NLP models’ focus on consequential features of the construct over dataset artifacts. more about movies than about sentiment, thereby failing to measure it as accurately in e.g., news me- dia (Puschmann and Powell, 2018). This potential learning of spurious cues over meaningful manifes- tations of the …

Content truncated to 500 words