An Empirical Study on Explanations in Out-of-Domain Settings

Recent work in Natural Language Processing has focused on developing approaches that extract faithful explanations, either via identifying the most important tokens in the input (i.e. post-hoc explanations) or by designing inherently faithful models that first select the most important tokens and then use them to predict the correct label (i.e. select-then-predict models). Currently, these approaches are largely evaluated on in-domain settings. Yet, little is known about how post-hoc explanations and inherently faithful models perform in out-ofdomain settings. In this paper, we conduct an extensive empirical study that examines: (1) the out-of-domain faithfulness of post-hoc explanations, generated by five feature attribution methods; and (2) the out-of-domain performance of two inherently faithful models over six datasets. Contrary to our expectations, results show that in many cases out-of-domain post-hoc explanation faithfulness measured by sufficiency and comprehensiveness is higher compared to in-domain. We find this misleading and suggest using a random baseline as a yardstick for evaluating post-hoc explanation faithfulness. Our findings also show that selectthen predict models demonstrate comparable predictive performance in out-of-domain settings to full-text trained models. 1 1 Code is attached to the submission and will be publicly released. 2 We use these terms interchangeably throughout our work. reasoning behind a model's prediction (Jacovi and 040 Goldberg, 2020) 041 Two popular methods for extracting explanations 042 are through feature attribution approaches (i.e. post-043 hoc explanation methods) or via inherently faithful 044 classifiers (i.e. select-then-predict models). The 045 first computes the contribution of different parts 046 of the input with respect to a model's prediction 047