What Can We Learn from Collective Human Opinions on Natural Language Inference Data?

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 9131–9143, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 9131 What Can We Learn from Collective Human Opinions on Natural Language Inference Data? Yixin Nie Xiang Zhou Mohit Bansal Department of Computer Science University of North Carolina at Chapel Hill {yixin1, xzh, mbansal}@cs.unc.edu Abstract Despite the subjective nature of many NLP tasks, most NLU evaluations have focused on using the majority label with presumably high agreement as the ground truth. Less atten- tion has been paid to the distribution of human opinions. We collect ChaosNLI, a dataset with a total of 464,500 annotations to study Collective HumAn OpinionS in oft-used NLI evaluation sets. This dataset is created by col- lecting 100 annotations per example for 3,113 examples in SNLI and MNLI and 1,532 ex- amples in αNLI. Analysis reveals that: (1) high human disagreement exists in a notice- able amount of examples in these datasets; (2) the state-of-the-art models lack the abil- ity to recover the distribution over human la- bels; (3) models achieve near-perfect accuracy on the subset of data with a high level of hu- man agreement, whereas they can barely beat a random guess on the data with low levels of human agreement, which compose most of the common errors made by state-of-the-art models on the evaluation sets. This questions the validity of improving model performance on old metrics for the low-agreement part of evaluation datasets. Hence, we argue for a detailed examination of human agreement in future data collection efforts, and evaluating model outputs against the distribution over col- lective human opinions.1 1 Introduction Natural Language Understanding (NLU) evalua- tion plays a key role in benchmarking progress in natural language processing (NLP) research. With the recent advance in language representa- tive learning (Devlin et al., 2019), results on pre- vious benchmarks have rapidly saturated. This leads to an explosion of difficult, diverse propos- als of tasks/datasets for NLU evaluation, including 1The ChaosNLI dataset and experimental scripts are avail- able at https://github.com/easonnie/ChaosNLI Natural Language Inference (e.g., SNLI, MNLI and ANLI) (Bowman et al., 2015; Williams et al., 2018; Nie et al., 2020), Grounded Commonsense Inference (Zellers et al., 2018), Commonsense QA (Talmor et al., 2019), Social Interactions Rea- soning (Sap et al., 2019), Abductive Commonsense Reasoning (αNLI) (Bhagavatula et al., 2020), etc. One common practice followed by most of these recent works is to simplify the evaluation of vari- ous reasoning abilities as a classification task. This is analogous to asking objective questions to a hu- man in educational testing. This simplification not only facilitates the data annotation but also gives interpretable evaluation results, based on which behaviors of the models are studied and then weak- nesses are diagnosed (Sanchez et al., 2018). Despite the straightforwardness of this formal- ization, one assumption behind most prior bench- mark data sourcing is that there exists a single pre- scriptive ground truth label for each example. The assumption might be true in human educational settings where prescriptivism …

Content truncated to 500 words