Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 9275–9293, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 9275 Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics Swabha Swayamdipta† Roy Schwartz‡∗ Nicholas Lourie† Yizhong Wang♦ Hannaneh Hajishirzi†♦ Noah A. Smith†♦ Yejin Choi†♦ †Allen Institute for Artificial Intelligence, Seattle, WA, USA ‡The Hebrew University of Jerusalem, Israel ♦Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle, WA, USA {swabhas,nicholasl}@allenai.org roys@cs.huji.ac.il {yizhongw,hannaneh,noah,yejin}@cs.washington.edu Abstract Large datasets have become commonplace in NLP research. However, the increased em- phasis on data quantity has made it challeng- ing to assess the quality of data. We intro- duce Data Maps—a model-based tool to char- acterize and diagnose datasets. We leverage a largely ignored source of information: the behavior of the model on individual instances during training (training dynamics) for build- ing data maps. This yields two intuitive mea- sures for each example—the model’s confi- dence in the true class, and the variability of this confidence across epochs—obtained in a single run of training. Experiments across four datasets show that these model-dependent measures reveal three distinct regions in the data map, each with pronounced character- istics. First, our data maps show the pres- ence of ambiguous regions with respect to the model, which contribute the most towards out-of-distribution generalization. Second, the most populous regions in the data are easy to learn for the model, and play an important role in model optimization. Finally, data maps un- cover a region with instances that the model finds hard to learn; these often correspond to labeling errors. Our results indicate that a shift in focus from quantity to quality of data could lead to robust models and improved out-of- distribution generalization. 1 Introduction The creation of large labeled datasets has fueled the advance of AI (Russakovsky et al., 2015; An- tol et al., 2015) and NLP in particular (Bowman et al., 2015; Rajpurkar et al., 2016). The common belief is that the more abundant the labeled data, the higher the likelihood of learning diverse phe- nomena, which in turn leads to models that gener- alize well. In practice, however, out-of-distribution ∗Work done at the Allen Institute for AI. 0.0 0.1 0.2 0.3 0.4 0.5 variability 0.0 0.2 0.4 0.6 0.8 1.0 confidence ambiguous easy-to-learn hard-to-learn correct. 0.0 0.2 0.3 0.5 0.7 0.8 1.0 Figure 1: Data map for SNLI train set, based on a ROBERTA-large classifier. The x-axis shows variability and y-axis, the confidence; the col- ors/shapes indicate correctness. The top-left corner of the data map (low variability, high confidence) corresponds to easy-to-learn examples, the bottom- left corner (low variability, low confidence) cor- responds to hard-to-learn examples, and examples on the right (with high variability) are ambiguous; all definitions are with respect to the ROBERTA-large model. The modal group in the data is formed by the easy-to-learn regions. For clarity we only plot 25K random samples from the SNLI train set. Fig. 8b in App. §C shows the same map in greater …

Content truncated to 500 words