Source-linked AI summary
WinoGrande: An Adversarial Winograd Schema Challenge at Scale
Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, Yejin Choi
TL;DR
Existing WSC-style benchmarks can yield high neural-model accuracy despite concerns that models exploit spurious dataset biases rather than robust commonsense. The paper introduces WinoGrande, combining large-scale crowdsourcing with AFLITE bias reduction. Models remain below humans on WinoGrande, while the dataset supports strong transfer results on related benchmarks and highlights risks of overestimating machine commonsense.
Problem
High accuracy on WSC variants raises whether neural language models acquired commonsense or exploit dataset biases, potentially overstating machine commonsense capability.
Method
WinoGrande combines carefully designed crowdsourcing with AFLITE, which reduces machine-detectable embedding associations in a 44k-problem WSC-inspired dataset.
Results
WinoGrande models score 59.4%-79.1% versus 94.0% for humans, while transfer learning achieves new state-of-the-art results across five related benchmarks.
Takeaways & Limitations
The results support WinoGrande as a transfer-learning resource and emphasize algorithmic bias reduction when evaluating commonsense reasoning.
Takeaways & Limitations
Existing benchmark gains may overestimate machine commonsense because models can rely on spurious statistical patterns and annotation artifacts.
Abstract
from arXiv · showhide
The Winograd Schema Challenge (WSC) (Levesque, Davis, and Morgenstern 2011), a benchmark for commonsense reasoning, is a set of 273 expert-crafted pronoun resolution problems originally designed to be unsolvable for statistical models that rely on selectional preferences or word associations. However, recent advances in neural language models have already reached around 90% accuracy on variants of WSC. This raises an important question whether these models have truly acquired robust commonsense capabilities or whether they rely on spurious biases in the datasets that lead to an overestimation of the true capabilities of machine commonsense. To investigate this question, we introduce WinoGrande, a large-scale dataset of 44k problems, inspired by the original WSC design, but adjusted to improve both the scale and the hardness of the dataset. The key steps of the dataset construction consist of (1) a carefully designed crowdsourcing procedure, followed by (2) systematic bias reduction using a novel AfLite algorithm that generalizes human-detectable word associations to machine-detectable embedding associations. The best state-of-the-art methods on WinoGrande achieve 59.4-79.1%, which are 15-35% below human performance of 94.0%, depending on the amount of the training data allowed. Furthermore, we establish new state-of-the-art results on five related benchmarks - WSC (90.1%), DPR (93.1%), COPA (90.6%), KnowRef (85.6%), and Winogender (97.1%). These results have dual implications: on one hand, they demonstrate the effectiveness of WinoGrande when used as a resource for transfer learning. On the other hand, they raise a concern that we are likely to be overestimating the true capabilities of machine commonsense across all these benchmarks. We emphasize the importance of algorithmic bias reduction in existing and future benchmarks to mitigate such overestimation.
1 Introduction
WSC benchmarks commonsense reasoning through pronoun-resolution problems, but high neural-model accuracy raises concerns about whether models use commonsense or exploit dataset biases. WinoGrande addresses scale and hardness while testing performance against humans and related benchmarks.
- Recent neural language models report around 90% accuracy on a WSC variant, prompting questions about whether they have acquired commonsense reasoning.
- WSC datasets can contain incidental word-association and other biases that let models solve problems for the wrong reasons.Reported word-association bias affected 13.5% of cases in one study.
- WinoGrande introduces 44k WSC-inspired problems designed to improve dataset scale and hardness through crowdsourcing and AFLITE-based bias reduction.AFLITE generalizes human-detectable word-occurrence biases to machine-detectable embedding associations.
- 94.0% human accuracy contrasts with 59.4%-79.1% best state-of-the-art performance, 15-35% below humans depending on training data.
- WinoGrande enables new state-of-the-art results on WSC, DPR, COPA, KnowRef, and Winogender, while these gains may still reflect spurious dataset effects.Reported accuracies are 90.1%, 93.1%, 90.6%, 85.6%, and 97.1%, respectively.
2 Crowdsourcing WINOGRANDE at Scale
WinoGrande was collected through constrained crowdsourcing and validated for answer clarity and resistance to word association. The procedure produced a large dataset, while the authors acknowledge that scaling can leave dataset-specific biases.
- WinoGrande preserves WSC’s twin-sentence structure while targeting problems that are trivial for humans but hard for AI systems.
- Crowd workers created twin sentences from randomly assigned WikiHow anchor words to increase creativity and topic diversity.The task also required WSC properties such as avoiding word association and maintaining small edit distance.
- Each question was validated by three workers for majority agreement, unambiguous answer choices, and resistance to simple word association.
- 68% of collected questions, reported as 53k, were deemed valid and the remainder were discarded.
- Crowdsourcing addresses some instance-level biases, but scaling the dataset can still leave dataset-specific biases requiring systematic reduction.
3 Algorithmic Data Bias Reduction
The paper introduces AFLITE, a lightweight adversarial filtering method that uses neural embeddings and ensembles of linear classifiers to reduce dataset-specific biases. Applied to WINOGRANDE, it substantially reduces label-separating correlations while retaining a challenging debiased dataset.
- Annotation artifacts leak target-label information and allow neural models to solve datasets correctly for incorrect reasons.
- AFLITE is more broadly applicable and lightweight than prior adversarial filtering because it neither requires overgeneration nor retraining a model at every iteration.
- AFLITE uses pre-computed RoBERTa embeddings and an ensemble of linear classifiers trained on random data subsets to identify and discard predictable instances.The method avoids manually specified lexical features and iteratively filters instances whose representations strongly indicate the correct answer.
- AFLITE proactively uncovers undesirable dataset shortcuts without assuming a particular bias source, unlike heuristic or hypothesis-only bias removal approaches.
- 2.53 → 2.51 for random reduction, →2.42 for PMI filtering, and →0.12 for AFLITE summarize the KL-divergence changes between answer-option samples.PCA still shows significant label separation after PMI filtering, whereas the much lower AFLITE divergence indicates reduced spurious correlation.
- AFLITE detects structural polarity correlations between answer options and target pronouns that lexical PMI heuristics may miss.
- 12,282 instances remain after filtering, while 31k filtered problems are released separately, producing 43,972 total WINOGRANDE problems.
4 Experimental Results
Experiments show that WINOGRANDE remains difficult for neural and retrieval-based systems, while performance improves with more training data but remains below humans. The dataset is also harder than earlier WSC-style benchmarks and supports transfer learning to related tasks.
- Results: 79.1% test accuracy is achieved by RoBERTa, compared with 94.0% human performance on WINOGRANDEdebiased.WKH and Ensemble LMs achieve chance-level performance, while the best model remains below humans.
- Results: Chance-level to below 60% accuracy is obtained by DPR-finetuned BERT and RoBERTa on WINOGRANDEdebiased, contrasting with their stronger performance on WSC-style datasets.This indicates that WINOGRANDEdebiased contains more challenging problems than WSC and existing variants.
- Learning Curve: 59% to 79% RoBERTa performance results when training data increases from 800 to 41K instances.The learning curve estimates that over 118K training instances would be needed to reach human-level performance.
- Learning Curve: 59% accuracy at approximately 800 training examples contrasts with around 90% performance on previous WSC variants at comparable training sizes.The authors conclude that the construction method produces collectively harder WSC problems than previous datasets.
5 Transfer Learning from WINOGRANDE
WINOGRANDE is used for transfer learning on WSC-related benchmarks, improving performance across related tasks and outperforming DPR-based fine-tuning. The results also include a gender-bias diagnostic on Winogender.
- Transfer-learning setup: WINOGRANDE is used as a resource by fine-tuning a model before evaluating it on WSC and related datasets.The evaluated tasks include WSC, PDP, SuperGLUE-WSC, DPR, KnowRef, and Winogender.
- Transfer-learning results: RoBERTa fine-tuned on WINOGRANDE improves accuracy on all related tasks and consistently outperforms RoBERTa fine-tuned on DPR.The comparison is reported for existing WSC-related tasks.
- Transfer-learning results: The significant improvement on COPA, which tests causes and effects, indicates that WINOGRANDE can support commonsense knowledge transfer to an unrelated task.The paper contrasts COPA with the pronoun-resolution focus of WINOGRANDE.
- Gender-bias diagnostics: Winogender measures gender bias through accuracy differences between non-gotcha and gotcha cases for female and male pronouns.Large ΔF or ΔM indicates greater gender bias, while zero absolute gaps with high accuracy are ideal.
- Gender-bias diagnostics: RoBERTa-WinoGrande has a smaller gender gap than RoBERTa-DPR while both models demonstrate very high accuracy.The comparison is reported for the Winogender diagnostics.
6 Conclusions
The paper concludes that WINOGRANDE combines a large, challenging benchmark with effective transfer learning, while warning that benchmark results may overestimate machine commonsense. It therefore supports dynamic, systematically debiased benchmarks that evolve with the state of the art.
- Contributions: WINOGRANDE contains 44k WSC-inspired problems and uses AFLITE to reduce spurious dataset-specific bias.The resulting dataset is harder for state-of-the-art models while remaining easy for humans.
- Contributions: The paper demonstrates effective transfer learning from WINOGRANDE and state-of-the-art results on several related benchmarks.This conclusion summarizes the reported cross-benchmark evaluation.
- Implications: The authors warn that existing benchmark results may overestimate machine commonsense because models can exploit spurious statistical patterns.The concern is that systems may solve problems correctly for incorrect reasons.
- Implications: The paper advocates dynamic datasets that evolve with the state of the art rather than static benchmarks maintained unchanged for many years.The proposed perspective is tied to composing challenges that remain difficult for AI systems.