Source-linked AI summary
Counterfactually-Augmented SNLI Training Data Does Not Yield Better Generalization Than Unaugmented Data
William Huang, Haokun Liu, Samuel R. Bowman
TL;DR
Models can exploit annotation artifacts on crowdsourced benchmarks, leaving out-of-domain generalization uncertain. This study compares counterfactually augmented and unaugmented SNLI training for NLI, finding no generalization advantage over size-matched data and possible harm to challenge-set robustness.
Problem
Models can exploit annotation artifacts to perform well on crowdsourced benchmarks, making their out-of-domain generalization difficult to predict.
Source: §1 · p. 1Method
The authors train RoBERTa on CNLI, size-matched unaugmented SNLI, and CNLI seed examples, then evaluate on out-of-domain MNLI and diagnostic sets.
Source: §1 · p. 1Results
CNLI provides no advantage over size-matched unaugmented SNLI on out-of-domain MNLI, and augmented examples can make models less robust to challenge examples.
Source: §5 · p. 4Takeaways & Limitations
Standard crowdsourcing of counterfactually augmented NLU training data does not appear effective for building more robust models with better generalization.
Source: §1 · p. 2The authors leave identifying heuristics to target and determining which augmentations are useful as future work.
Source: §5 · p. 5
Abstract
from arXiv · showhide
A growing body of work shows that models exploit annotation artifacts to achieve state-of-the-art performance on standard crowdsourced benchmarks---datasets collected from crowdworkers to create an evaluation task---while still failing on out-of-domain examples for the same task. Recent work has explored the use of counterfactually-augmented data---data built by minimally editing a set of seed examples to yield counterfactual labels---to augment training data associated with these benchmarks and build more robust classifiers that generalize better. However, Khashabi et al. (2020) find that this type of augmentation yields little benefit on reading comprehension tasks when controlling for dataset size and cost of collection. We build upon this work by using English natural language inference data to test model generalization and robustness and find that models trained on a counterfactually-augmented SNLI dataset do not generalize better than unaugmented datasets of similar size and that counterfactual augmentation can hurt performance, yielding models that are less robust to challenge examples. Counterfactual augmentation of natural language understanding data through standard crowdsourcing techniques does not appear to be an effective way of collecting training data and further innovation is required to make this general line of work viable.
1 Introduction
Annotation artifacts let models score highly on crowdsourced benchmarks while making out-of-domain generalization difficult to predict. This study tests whether counterfactually augmented NLI training data improves generalization and robustness compared with size-matched unaugmented data.
1 Introduction
- Annotation artifacts enable high benchmark performance but make out-of-domain generalization difficult to predict and obscure what these tasks test.Prior work found models can exploit artifacts to achieve state-of-the-art results on standard benchmarks.Source: §1 · p. 1
- Counterfactually augmented examples are made by minimally editing seed examples to produce counterfactual labels, with the aim of building more robust classifiers.Prior work reported that this collection method mitigates artifacts in SNLI, while Khashabi et al. (2020) found size- and cost-controlled augmentation offered little benefit.Source: §1 · p. 1
- The study compares RoBERTa trained on 8.3k CNLI examples, 8.3k unaugmented SNLI examples, and 1.7k CNLI seed examples on MNLI and two diagnostic sets.The unaugmented SNLI sample controls for training-set size.Source: §1 · pp. 1–2
- CNLI matched size-controlled unaugmented SNLI on out-of-domain MNLI, while adding augmented examples helped over the smaller seed set but could hurt challenge-set robustness.The authors also report that the improvement over seed examples corresponded to increased n-gram diversity.Source: §1 · p. 2
2 Related Work
Prior NLI research documents exploitable dataset artifacts and explores adversarial or counterfactually augmented data. This paper focuses on naturally crowdsourced augmentation for training, contrasting it with prior work on evaluation and reading comprehension.
2 Related Work
- NLI models can exploit hypothesis-only cues, negation associations, and premise-hypothesis word overlap to score well on benchmarks.Hypothesis-only models have performed as much as 35 points above chance.Source: §2 · p. 2
- ANLI uses a human-and-model-in-the-loop process, whereas this study examines naturally collected human-in-the-loop augmented data.The authors note that ANLI is inherently adversarial toward the model used in its collection.Source: §2 · p. 2
- Kaushik et al. (2020) reduce hypothesis-only bias in SNLI with minimally edited counterfactuals, while Khashabi et al. (2020) report similar-to-worse results when augmentation is no cheaper and dataset size is controlled.This study differs by testing out-of-domain NLI performance rather than the prior work's focus on BoolQ or in-domain effects.Source: §2 · p. 2
- Gardner et al. (2020) use manually written contrast sets for evaluation, while this work uses crowdsourced counterfactual data for training.Their contrast sets take 1–3 minutes per example and 17–50 hours per 1,000 examples.Source: §2 · p. 2
3 Experimental Setup
The experiments train RoBERTa on SNLI, CNLI, or CNLI seed examples, then assess out-of-domain generalization and robustness to challenge examples. The setup controls training-set size for SNLI comparisons and uses early stopping during training.
3 Experimental Setup
- The study evaluates out-of-domain generalization on MNLI and robustness on challenge examples from Naik et al. (2018) and the GLUE diagnostic set.MNLI validation accuracy combines matched and mismatched examples.Source: §3 · pp. 2–3, Figure 1 · p. 3
- Training compares 8.3k unaugmented SNLI examples, 8.3k CNLI examples, and 1.7k CNLI seed examples; five SNLI subsamples control for size.SNLI contains premise-hypothesis pairs labeled for entailment, neutral, and contradiction; CNLI examples minimally edit SNLI seeds to yield counterfactual labels.Source: §3 · pp. 2–3
- McCoy et al.'s adversarial examples produced performance consistently below 3%, so the authors do not report them in detail.This limits the reported evaluation of those examples.Source: §3 · p. 3
- Each run uses RoBERTa-base for 20 epochs with early stopping, across 20 runs and a random search over learning rate, batch size, and seed.Learning rates are 1e-5, 2e-5, and 3e-5; batch sizes are 32 and 64.Source: §3 · p. 3
4 Results
On MNLI, CNLI performs no better than size-comparable unaugmented SNLI, though it improves over its seed examples. Diagnostic performance is similar on most categories but worse on Negation, while the augmented data add n-gram diversity.
Generalization to MNLI
- CNLI performs no better on MNLI than a comparably sized sample of unaugmented SNLI, though it improves over the 1.7k seed examples.The authors describe counterfactual examples as somewhat helpful when strictly additive.Source: §4 · p. 3
Robustness to Diagnostic Sets
- Diagnostic performance is similar across most categories, but CNLI performs worse on Negation than either unaugmented SNLI or CNLI seed examples.The Negation challenge appends “and false is not true” to each hypothesis; the authors speculate that augmentation may amplify an association already present in the seed examples.Source: §4 · pp. 3–4, Figure 2 · p. 4
Lexical Diversity
- Adding minimally edited examples roughly doubles CNLI’s 4-gram types relative to its seed examples, but CNLI has about 70% of the 2-, 3-, and 4-gram types in similarly large unaugmented sets.The authors link the increase in n-grams to the general improvement over seed examples and note less diversity per example under minimal edits.Source: §4 · p. 4, Table 1 · p. 4
5 Conclusion
In English NLI, counterfactually augmented training data improve over seed examples but offer no advantage over similar-sized unaugmented data, and can reduce robustness to challenge examples. The conclusion questions their value for training under current crowdsourcing protocols while retaining a role for evaluation and identifying expert guidance as a future direction.
5 Conclusion
- Counterfactually augmented data improve over seed examples but offer no advantage over similar amounts of unaugmented data, and can make models less robust to challenge examples.The conclusion relates the seed-set improvement to increased n-gram diversity.Source: §5 · p. 4
- Assuming edits and new-example writing take similar time, current protocols offer no obvious value for crowdsourcing augmentations as training data.The authors distinguish this training conclusion from the continued value they see in naturally collected counterfactual data for evaluation.Source: §5 · p. 5
- Counterfactually augmented data may still help address systematic gaps in test sets and provide longer-lasting benchmarks as benchmark performance saturates.The authors cite Gardner et al. (2020) for using this data collection approach to address testing gaps.Source: §5 · p. 5
- Expert guidance may help target training-data heuristics, but identifying which heuristics to target and which augmentations are useful remains future work.The authors leave these questions open, conditional on useful augmentations being possible.Source: §5 · p. 5