Dynabench: Rethinking Benchmarking in NLP

Kiela, Douwe; Bartolo, Max; Nie, Yixin; Kaushik, Divyansh; Geiger, Atticus; Wu, Zhengxuan; Vidgen, Bertie; Prasad, Grusha; Singh, Amanpreet; Ringshia, Pratik; Ma, Zhao; Thrush, Tristan; Riedel, Sebastian; Waseem, Zeerak; Stenetorp, Pontus; Jia, Robin; Bansal, Mohit; Potts, Christopher; Williams, Adina

doi:10.18653/v1/2021.naacl-main.324

Cited by 84 publications

(81 citation statements)

References 81 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Adversarial filtering There has been a trend toward adversarial data collection paradigms, in which annotators are asked to produce examples on which current systems fail (Le Bras et al, 2020;Kiela et al, 2021;Talmor et al, 2021;Jia and Liang, 2017;Zellers et al, 2019, i.a.). While this eliminates examples containing artifacts that models have already learned, it does not prevent the creation of new ones on which current models fail.…”

Section: Related Workmentioning

confidence: 99%

WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation

Liu¹,

Swayamdipta²,

Smith³

et al. 2022

Preprint

View full text Add to dashboard Cite

A recurring challenge of crowdsourcing NLP datasets at scale is that human writers often rely on repetitive patterns when crafting examples, leading to a lack of linguistic diversity. We introduce a novel paradigm for dataset creation based on human and machine collaboration, which brings together the generative strength of language models and the evaluative strength of humans. Starting with an existing dataset, MultiNLI, our approach uses dataset cartography to automatically identify examples that demonstrate challenging reasoning patterns, and instructs GPT-3 to compose new examples with similar patterns. Machine generated examples are then automatically filtered, and finally revised and labeled by human crowdworkers to ensure quality. The resulting dataset, WANLI, consists of 108,357 natural language inference (NLI) examples that present unique empirical strengths over existing NLI datasets. Remarkably, training a model on WANLI instead of MNLI (which is 4 times larger) improves performance on seven out-of-domain test sets we consider, including by 11% on HANS and 9% on Adversarial NLI. Moreover, combining MNLI with WANLI is more effective than combining with other augmentation sets that have been introduced. Our results demonstrate the potential of natural language generation techniques to curate NLP datasets of enhanced quality and diversity.

show abstract

Section: Related Workmentioning

confidence: 99%

WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation

Liu¹,

Swayamdipta²,

Smith³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…We convert natural language explanations to oracle importance scores instead of collecting oracle importance scores directly from naïve annotators for two reasons. First, there already exist data sets of natural language explanations, where annotators were required to reason about models' decision making in an adversarial setting (Nie et al, 2020), and more such data sets are being generated (Kiela et al, 2021). Second, we contend that for most nonexpert annotators, asking them to provide verbal descriptions is easier and more natural than asking them to answer a question like, "For which words do you think the model's prediction would change the most if that word was blanked out?".…”

Section: Computing Oracle Importance Scoresmentioning

confidence: 99%

To what extent do human explanations of model behavior align with actual model behavior?

Prasad¹,

Nie²,

Bansal³

et al. 2021

Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP

Self Cite

View full text Add to dashboard Cite

Given the increasingly prominent role NLP models (will) play in our lives, it is important for human expectations of model behavior to align with actual model behavior. Using Natural Language Inference (NLI) as a case study, we investigate the extent to which human-generated explanations of models' inference decisions align with how models actually make these decisions. More specifically, we define three alignment metrics that quantify how well natural language explanations align with model sensitivity to input words, as measured by integrated gradients. Then, we evaluate eight different models (the base and large versions of BERT, RoBERTa and ELECTRA, as well as an RNN and bag-of-words model), and find that the BERT-base model has the highest alignment with human-generated explanations, for all alignment metrics.Focusing in on transformers, we find that the base versions tend to have higher alignment with human-generated explanations than their larger counterparts, suggesting that increasing the number of model parameters leads, in some cases, to worse alignment with human explanations.Finally, we find that a model's alignment with human explanations is not predicted by the model's accuracy, suggesting that accuracy and alignment are complementary ways to evaluate models.

show abstract

“…Previous work has shed light on several dimensions of hate speech data that prevents generalisation, such as imprecise construct specification (Samory et al, 2021), biased data collection (Ousidhoum et al, 2020), and annotation artifacts (Waseem, 2016). Several solutions have been proposed for these issues such as adversarial data generation (Dinan et al, 2019), dynamic benchmarking (Kiela et al, 2021) and debiasing techniques .…”

Section: Related Workmentioning

confidence: 99%

How Does Counterfactually Augmented Data Impact Models for Social Computing Constructs?

Sen¹,

Samory²,

Floeck³

et al. 2021

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

As NLP models are increasingly deployed in socially situated settings such as online abusive content detection, it is crucial to ensure that these models are robust. One way of improving model robustness is to generate counterfactually augmented data (CAD) for training models that can better learn to distinguish between core features and data artifacts. While models trained on this type of data have shown promising out-of-domain generalizability, it is still unclear what the sources of such improvements are. We investigate the benefits of CAD for social NLP models by focusing on three social computing constructs -sentiment, sexism, and hate speech. Assessing the performance of models trained with and without CAD across different types of datasets, we find that while models trained on CAD show lower in-domain performance, they generalize better out-of-domain. We unpack this apparent discrepancy using machine explanations and find that CAD reduces model reliance on spurious features. Leveraging a novel typology of CAD to analyze their relationship with model performance, we find that CAD which acts on the construct directly or a diverse set of CAD leads to higher performance.

show abstract

Dynabench: Rethinking Benchmarking in NLP

Cited by 84 publications

References 81 publications

WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation

WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation

To what extent do human explanations of model behavior align with actual model behavior?

How Does Counterfactually Augmented Data Impact Models for Social Computing Constructs?

Contact Info

Product

Resources

About