In attempts to develop sample-efficient and interpretable algorithms, researcher have explored myriad mechanisms for collecting and exploiting feature feedback, auxiliary annotations provided for training (but not test) instances that highlight salient evidence. Examples include bounding boxes around objects and salient spans in text. Despite its intuitive appeal, feature feedback has not delivered significant gains in practical problems as assessed on iid holdout sets. However, recent works on counterfactually augmented data suggest an alternative benefit of supplemental annotations, beyond interpretability: lessening sensitivity to spurious patterns and consequently delivering gains in out-of-domain evaluations. We speculate that while existing methods for incorporating feature feedback have delivered negligible in-sample performance gains, they may nevertheless provide out-of-domain benefits. Our experiments addressing sentiment analysis, show that feature feedback methods perform significantly better on various natural out-of-domain datasets despite comparable in-domain evaluations. By contrast, performance on natural language inference remains comparable. Finally, we compare those tasks where feature feedback does (and does not) help.

The ongoing COVID-19 pandemic resulted in significant ramifications for international relations ranging from travel restrictions, global ceasefires, and international vaccine production and sharing agreements. Amidst a wave of infections in India that resulted in a systemic breakdown of healthcare infrastructure, a social welfare organization based in Pakistan offered to procure medical-grade oxygen to assist India - a nation which was involved in four wars with Pakistan in the past few decades. In this paper, we focus on Pakistani Twitter users’ response to the ongoing healthcare crisis in India. While #IndiaNeedsOxygen and #PakistanStandsWithIndia featured among the top-trending hashtags in Pakistan, divisive hashtags such as #EndiaSaySorryToKashmir simultaneously started trending. Against the backdrop of a contentious history including four wars, divisive content of this nature, especially when a country is facing an unprecedented healthcare crisis, fuels further deterioration of relations. In this paper, we define a new task of detecting supportive content and demonstrate that existing NLP for social impact tools can be effectively harnessed for such tasks within a quick turnaround time. We also release the first publicly available data set at the intersection of geopolitical relations and a raging pandemic in the context of India and Pakistan.