Detecting Label Errors by Using Pre-Trained Language Models

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9074 - 9091 December 7-11, 2022 ©2022 Association for Computational Linguistics Detecting Label Errors by using Pre-Trained Language Models Derek Chong∗ Stanford University derekch@stanford.edu Jenny Hong∗ Stanford University jennyhong@cs.stanford.edu Christopher D. Manning Stanford University manning@cs.stanford.edu Abstract We show that large pre-trained language models are inherently highly capable of identifying label errors in natural language datasets: simply examining out-of-sample data points in descending order of fine- tuned task loss significantly outperforms more complex error-detection mechanisms proposed in previous work. To this end, we contribute a novel method for introduc- ing realistic, human-originated label noise into existing crowdsourced datasets such as SNLI and TweetNLP. We show that this noise has similar properties to real, hand- verified label errors, and is harder to detect than existing synthetic noise, creating chal- lenges for model robustness. We argue that human-originated noise is a better standard for evaluation than synthetic noise. Finally, we use crowdsourced verification to evalu- ate the detection of real errors on IMDB, Amazon Reviews, and Recon, and confirm that pre-trained models perform at a 9–36% higher absolute Area Under the Precision- Recall Curve than existing models. 1 Introduction Improving model performance in the presence of label errors comprises an area of active re- search (Song et al., 2022). However, existing methods focus on label errors in training data. Although seldom acknowledged, evaluation la- bel errors are at least as pernicious as training label errors: pervasive errors in commonly used NLP benchmarks have been found to destabi- lize model performance (Malik and Bhardwaj, 2011; Northcutt et al., 2021b). Such findings suggest that improving training methods does not preclude the need for improving the un- derlying data. We propose a simple method for using large, pre-trained language models ∗Equal contribution. 0.0 0.2 0.4 0.6 0.8 1.0 Recall 0.0 0.2 0.4 0.6 0.8 1.0 Precision Baseline FM Loss FM Ensemble CL Overlays Figure 1: Precision-recall curves for label error detection: Large language models detect label er- rors with high precision, and far more effectively than a baseline word vector-based neural classifier. Overlaying a state-of-the-art model-agnostic error detection method, Confident Learning, results in little to no improvement (TweetNLP-5; §7). (LLMs) to directly identify label errors for the purposes of correcting or removing them. The majority of work in identifying label errors, and in data-centric artificial intelli- gence (DCAI) more broadly, focuses on image and healthcare data (DCAI Workshop, 2021). However, the success of the foundation model (FM) paradigm in applying pre-trained lan- guage models to a variety of NLP tasks (Bom- masani et al., 2021; Reiss et al., 2020) sug- gests that FMs may be a powerful tool for de- tecting and correcting label errors in language datasets. Pre-training has been shown to im- bue models with properties such as resistance to label errors, class imbalance (Karthik et al., 2021), out-of-distribution detection (Hendrycks et al., 2018), and confidence calibration (De- sai and Durrett, 2020), while conferring ro- 9074 Dataset Text Label Sentiment IMDB …

Content truncated to 500 words