WinoGrande: An Adversarial Winograd Schema Challenge at Scale

Abstract

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% (absolute) below human performance of 94.0%, depending on the amount of the training data allowed (2% – 100% respectively).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.

Social Bias Frames: Reasoning about Social and Power Implications of Language

“…The model generates the correct stereotypes when there is high lexical overlap with the post (e.g., examples d and e). This is in line with previous research showing that large language models rely on correlational patterns in data (Sap et al, 2019c;Sakaguchi et al, 2020).…”

Putting psychology to the test: Rethinking model evaluation through benchmarking and prediction

“…Once again, natural language processing offers an excellent example: language models are generally trained on one or more general-purpose objectives (e.g. next-word prediction), and, after (often minimal) fine-tuning, they are evaluated against composite benchmarks (e.g., Sakaguchi, Le Bras, Bhagavatula, & Choi, 2019;Wang et al, 2019). In this regard, a particularly interesting example is that of GPT-3 (T. B.…”

The Winograd Schemas from Hell

“…Recently, a much larger set of Winograd Schemas, referred to as the WinoGrande set, has been created and used as the basis of the current Winograd Challenge non-human champion, a specialized version of the UnifiedQA solver [Khashabi et al 2020]. This solver attains more than 90% accuracy in the Winograd Challenge, a truly impressive figure that is similar to human accuracy [Sakaguchi et al 2019]. An important feature of the current champion solvers is that they are based on language models learned from large textual datasets; that is, they estimate the probability for each possible solution of a Winograd Schema, and output the most likely solutions.…”

“…Indeed, we suspect that if one tries to follow the original guidelines concerning the Winograd Challenge as strictly as possible, then one will be left with Winograd Schemas that resemble the ones in WSC273. We now know that such guidelines limit too much the scope of Winograd Schemas: as demonstrated by recent results on computer solvers, Winograd Schemas that are (relatively) easy for human subjects are (relatively) easy for computers as well [Sakaguchi et al 2019]. In hindsight this is perhaps unsurprising because language must reflect established facts and rules and social conventions that must appear in large textual corpora.…”