Recent advances in NLP systems, notably the pretraining-and-finetuning paradigm, have achieved great success in predictive accuracy. However, these systems are usually not well calibrated for uncertainty out-of-the-box. Many recalibration methods hav
e been proposed in the literature for quantifying predictive uncertainty and calibrating model outputs, with varying degrees of complexity. In this work, we present a systematic study of a few of these methods. Focusing on the text classification task and finetuned large pretrained language models, we first show that many of the finetuned models are not well calibrated out-of-the-box, especially when the data come from out-of-domain settings. Next, we compare the effectiveness of a few widely-used recalibration methods (such as ensembles, temperature scaling). Then, we empirically illustrate a connection between distillation and calibration. We view distillation as a regularization term encouraging the student model to output uncertainties that match those of a teacher model. With this insight, we develop simple recalibration methods based on distillation with no additional inference-time cost. We show on the GLUE benchmark that our simple methods can achieve competitive out-of-domain (OOD) calibration performance w.r.t. more expensive approaches. Finally, we include ablations to understand the usefulness of components of our proposed method and examine the transferability of calibration via distillation.
Data in general encodes human biases by default; being aware of this is a good start, and the research around how to handle it is ongoing. The term bias' is extensively used in various contexts in NLP systems. In our research the focus is specific to
biases such as gender, racism, religion, demographic and other intersectional views on biases that prevail in text processing systems responsible for systematically discriminating specific population, which is not ethical in NLP. These biases exacerbate the lack of equality, diversity and inclusion of specific population while utilizing the NLP applications. The tools and technology at the intermediate level utilize biased data, and transfer or amplify this bias to the downstream applications. However, it is not enough to be colourblind, gender-neutral alone when designing a unbiased technology -- instead, we should take a conscious effort by designing a unified framework to measure and benchmark the bias. In this paper, we recommend six measures and one augment measure based on the observations of the bias in data, annotations, text representations and debiasing techniques.
