[Paper Review] DQI: Measuring Data Quality in NLP

Swaroop Mishra, Anjana Arunkumar+3 more|arXiv (Cornell University)|May 2, 2020

Topic Modeling76 references7 citations

TL;DR

This paper introduces the Data Quality Index (DQI), a generic formula to measure and improve NLP dataset quality by identifying and mitigating spurious biases during data creation. Using DQI with adversarial filtering (AFLite), visualization tools, and automation (TextFooler, Autofix), the authors renovate the SNLI dataset, resulting in models that generalize better to out-of-distribution tasks despite reduced performance on the original benchmark.

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.

Motivation & Objective

• To address the overestimation of AI capabilities due to spurious biases in NLP datasets.
• To develop a proactive method for dataset creators to prevent bias during data creation, rather than correcting it post-hoc.
• To create a dynamic, evolving benchmarking framework that reflects true AI progress by systematically improving data quality.
• To enhance model generalization by generating harder, bias-free benchmarks through active learning and data renovation.
• To extend the data creation process beyond static datasets by integrating quality monitoring and real-time feedback.

Proposed method

• Proposes a generic DQI formula based on bias leads from prior literature, measuring data quality through multiple components like lexical, syntactic, and semantic features.
• Employs AFLite for adversarial filtering to identify and remove biased examples from the SNLI test set, validating DQI's effectiveness.
• Introduces a data creation paradigm with interactive visualizations to help analysts assess data quality and the impact of new samples on the dataset.
• Uses Autofix to assist crowdsource workers in generating higher-quality data in real time, reducing bias at the source.
• Applies TextFooler to generate adversarial examples that expose low-quality data, enabling targeted repair.
• Retrains BERT and RoBERTa on the renovated SNLI dataset to evaluate generalization improvements on out-of-distribution tasks.

Experimental results

Research questions

• RQ1 Can a generic, component-based Data Quality Index (DQI) effectively measure and guide the reduction of spurious biases in NLP datasets during creation?
• RQ2 How does integrating DQI with adversarial filtering (AFLite) improve dataset quality and model generalization?
• RQ3 To what extent can visualization and automation tools (Autofix, TextFooler) enhance data quality during crowd-sourced data creation?
• RQ4 Does renovating a benchmark dataset like SNLI using DQI lead to models that generalize better on out-of-distribution data?
• RQ5 Can DQI be used in an active learning setup to generate a hierarchy of increasingly harder benchmarks reflecting true AI progress?

Key findings

• The DQI formula effectively identifies and measures multiple types of spurious biases in NLP datasets, including lexical, syntactic, and semantic artifacts.
• Applying DQI with AFLite successfully filters biased examples from the SNLI test set, validating the index’s sensitivity to known bias patterns.
• The integration of visualizations and automation tools (Autofix, TextFooler) enables analysts and crowdworkers to detect and correct low-quality data in real time.
• Retraining BERT and RoBERTa on the renovated SNLI dataset leads to improved generalization on out-of-distribution benchmarks, despite lower performance on the original test set.
• The renovated dataset reveals a significant performance gap between models and humans, indicating that the original benchmark overestimated model capabilities.
• The DQI-based data creation paradigm enables the generation of a series of increasingly harder benchmarks through active learning, reflecting true progress in AI.