The “Problem” of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10671 - 10682 December 7-11, 2022 ©2022 Association for Computational Linguistics The “Problem” of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation Barbara Plank Center for Information and Language Processing (CIS), MaiNLP lab, LMU Munich, Germany Munich Center for Machine Learning (MCML), Munich, Germany b.plank@lmu.de Abstract Human variation in labeling is often consid- ered noise. Annotation projects for machine learning (ML) aim at minimizing human la- bel variation, with the assumption to maximize data quality and in turn optimize and maxi- mize machine learning metrics. However, this conventional practice assumes that there ex- ists a ground truth, and neglects that there ex- ists genuine human variation in labeling due to disagreement, subjectivity in annotation or multiple plausible answers. In this position pa- per, we argue that this big open problem of hu- man label variation persists and critically needs more attention to move our field forward. This is because human label variation impacts all stages of the ML pipeline: data, modeling and evaluation. However, few works consider all of these dimensions jointly; and existing research is fragmented. We reconcile different previ- ously proposed notions of human label varia- tion, provide a repository of publicly-available datasets with un-aggregated labels, depict ap- proaches proposed so far, identify gaps and suggest ways forward. As datasets are becom- ing increasingly available, we hope that this synthesized view on the “problem” will lead to an open discussion on possible strategies to devise fundamentally new directions. 1 Introduction In Natural Language Processing (NLP) much progress today is driven by fine-tuning large pre- trained language models using an annotated dataset, assumed to be representative for a target language task of interest (Schlangen, 2021). This is analo- gously so in Machine Learning (ML) and Computer Vision (CV), where the target tasks differ, yet the conceptual pipeline remains the same: data, mod- eling, evaluation. Despite the importance of anno- tated data—as it fuels all steps in this pipeline—a crucial assumption of today’s learning systems is to rely on a single gold label per instance. The gold Figure 1: We propose the term human label variation to capture the fact that inherent disagreement in annotation can be due to genuine disagreement, subjectivity or simply because two (or more) views are plausible. label is obtained by aggregation (e.g. majority vote) of labels crucially provided by humans. The assumption of a ground truth (and taking the majority vote or the ‘mode’ of the human judge- ment distribution) makes sense when humans in- volved in labeling highly agree on the answer to the questions, such as “Does this image contain a bird?”, “Is ‘learn’ a verb?”, “What is the capital of Italy?". However, this assumption often does not make sense—especially when language is in- volved. For example, on questions determining a word sense, questions such as “Is this comment toxic?” or questions involving understanding indi- rect answers to polar questions like “Q: Hey. Every- thing ok?” “A: …

Content truncated to 500 words