DQI: Measuring Data Quality in NLP

Abstract

Neural language models have achieved human level performance across several NLP datasets. However, recent studies have shown that these models are not truly learning the desired task; rather, their high performance is attributed to overfitting using spurious biases, which suggests that the capabilities of AI systems have been over-estimated. We introduce a generic formula for Data Quality Index (DQI) to help dataset creators create datasets free of such unwanted biases. We evaluate this formula using a recently proposed approach for adversarial filtering, AFLite. We propose a new data creation paradigm using DQI to create higher quality data. The data creation paradigm consists of several data visualizations to help data creators (i) understand the quality of data and (ii) visualize the impact of the created data instance on the overall quality. It also has a couple of automation methods to (i) assist data creators and (ii) make the model more robust to adversarial attacks. We use DQI along with these automation methods to renovate biased examples in SNLI. We show that models trained on the renovated SNLI dataset generalize better to out of distribution tasks. Renovation results in reduced model performance, exposing a large gap with respect to human performance. DQI systematically helps in creating harder benchmarks using active learning. Our work takes the process of dynamic dataset creation forward, wherein datasets evolve together with the evolving state of the art, therefore serving as a means of benchmarking the true progress of AI.

Don’t Blame the Annotator: Bias Already Starts in the Annotation Instructions

“…To mitigate the bias, prior works have focused on priming crowdsourcing annotators with minimal information to increase their imagination (Geva et al, 2021; to avoid recurring patterns. Arunkumar et al (2020) develops a real time feedback and metric-in-the loop (Mishra et al, 2020b) workflow to educate crowdworkers in controlling dataset biases. provides an iterative protocol with expert assessments for crowdsourcing data collection to increase difficulty of instances.…”

To Find Waldo You Need Contextual Cues: Debiasing Who's Waldo

“…Dasgupta et al (2018) incorporate compositional information into sentence embeddings for Natural Language Inference. DQI (Mishra et al, 2020) offers quantitative metrics to assess biases in automated dataset creation in Natural Language Processing. Le Bras et al ( 2020) introduce adversarial measures to mitigate biases in various Natural Language Processing and Computer Vision tasks.…”