The Sensitivity of Language Models and Humans to Winograd Schema Perturbations

Abdou, Mostafa; Ravishankar, Vinit; Barrett, Maria; Belinkov, Yonatan; Elliott, Desmond; Søgaard, Anders

doi:10.18653/v1/2020.acl-main.679

Cited by 24 publications

(31 citation statements)

References 36 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They showed that the success of a then-state-of-the-art LM ensemble (Trinh and Le, 2018) resulted mainly from improvements on simpler "associative" instances. Similarly, experiments by Abdou et al (2020) show that models are sensitive to linguistic perturbations of Winograd-style examples. New datasets have been proposed to circumvent issues of unintentionally easy test instances, including Winogrande (Sakaguchi et al, 2020), a scaled WSC-variant debiased against RoBERTa, and KnowRef , which consists of naturally occurring sentences that are free of WSC-specific stylistic quirks.…”

Section: Related Workmentioning

confidence: 91%

An Analysis of Dataset Overlap on Winograd-Style Tasks

Emami

Suleman

Trischler

et al. 2020

Proceedings of the 28th International Conference on Computational Linguistics

View full text Add to dashboard Cite

The Winograd Schema Challenge (WSC) and variants inspired by it have become important benchmarks for common-sense reasoning (CSR). Model performance on the WSC has quickly progressed from chance-level to near-human using neural language models trained on massive corpora. In this paper, we analyze the effects of varying degrees of overlap between these training corpora and the test instances in WSC-style tasks. We find that a large number of test instances overlap considerably with the corpora on which state-of-the-art models are (pre)trained, and that a significant drop in classification accuracy occurs when we evaluate models on instances with minimal overlap. Based on these results, we develop the KNOWREF-60K dataset, which consists of over 60k pronoun disambiguation problems scraped from web data. KNOWREF-60K is the largest corpus to date for WSC-style common-sense reasoning and exhibits a significantly lower proportion of overlaps with current pretraining corpora.

show abstract

Section: Related Workmentioning

confidence: 91%

An Analysis of Dataset Overlap on Winograd-Style Tasks

Emami

Suleman

Trischler

et al. 2020

Proceedings of the 28th International Conference on Computational Linguistics

View full text Add to dashboard Cite

show abstract

“…Probability-A (Pro-A) considers the generative probability of the choice conditioned on the question. However, it suffers from the statistical bias of choices, such as word frequency and sentence length (Abdou et al, 2020). To alleviate this, MutualInfo-QA (MI-QA) calculates the mutual information between the question and the choice.…”

Section: Related Workmentioning

confidence: 99%

“…Table 1 lists several typical score functions. However, these scores can be easily influenced by word frequencies, sentence structures, and other factors, which can mislead the models and make existing methods oversensitive to lexical perturbations (Abdou et al, 2020;Tamborrino et al, 2020). Figure 1 shows two examples.…”

Section: Introductionmentioning

confidence: 99%

A Semantic-based Method for Unsupervised Commonsense Question Answering

Niu¹,

Huang²,

Liang³

et al. 2021

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Confer

View full text Add to dashboard Cite

Unsupervised commonsense question answering is appealing since it does not rely on any labeled task data. Among existing work, a popular solution is to use pre-trained language models to score candidate choices directly conditioned on the question or context. However, such scores from language models can be easily affected by irrelevant factors, such as word frequencies, sentence structures, etc. These distracting factors may not only mislead the model to choose a wrong answer but also make it oversensitive to lexical perturbations in candidate answers.In this paper, we present a novel SEmanticbased Question Answering method (SEQA) for unsupervised commonsense question answering. Instead of directly scoring each answer choice, our method first generates a set of plausible answers with generative models (e.g., GPT-2), and then uses these plausible answers to select the correct choice by considering the semantic similarity between each plausible answer and each choice. We devise a simple, yet sound formalism for this idea and verify its effectiveness and robustness with extensive experiments. We evaluate the proposed method on four benchmark datasets, and our method achieves the best results in unsupervised settings. Moreover, when attacked by TextFooler (Jin et al., 2020) with synonym replacement, SEQA demonstrates much less performance drops than baselines, thereby indicating stronger robustness.

show abstract

“…In the examples shown throughout the paper, John and Mary are used as placeholders for the subject and object of the verb of interest. The choice of names to go in these slots can affect model predictions (Abdou et al, 2020), so we generate 200 variants of each stimulus, varying the names and the order between the two genders and we query the PLMs with all of them. The full procedure is described in Appendix B.…”

Section: Context-free Ic Biasmentioning

confidence: 99%

“…In this section we consider a seemingly minor but important consideration. Abdou et al (2020) showed that model predictions on the Winograd Schema Challenge greatly vary with changes in the gender and identity of proper nouns used in the stimuli. We alleviate this issue by marginalizing over a range of proper nouns.…”

Section: B Proper Nounsmentioning

confidence: 99%

John praised Mary because _he_? Implicit Causality Bias and Its Interaction with Explicit Cues in LMs

Kementchedjhieva¹,

Anderson²,

Søgaard³

2021

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021

Self Cite

View full text Add to dashboard Cite

Some interpersonal verbs can implicitly attribute causality to either their subject or their object and are therefore said to carry an implicit causality (IC) bias. Through this bias, causal links can be inferred from a narrative, aiding language comprehension. We investigate whether pre-trained language models (PLMs) encode IC bias and use it at inference time. We find that to be the case, albeit to different degrees, for three distinct PLM architectures. However, causes do not always need to be implicit-when a cause is explicitly stated in a subordinate clause, an incongruent IC bias associated with the verb in the main clause leads to a delay in human processing. We hypothesize that the temporary challenge humans face in integrating the two contradicting signals, one from the lexical semantics of the verb, one from the sentence-level semantics, would be reflected in higher error rates for models on tasks dependent on causal links. The results of our study lend support to this hypothesis, suggesting that PLMs tend to prioritize lexical patterns over higher-order signals.

show abstract

The Sensitivity of Language Models and Humans to Winograd Schema Perturbations

Cited by 24 publications

References 36 publications

An Analysis of Dataset Overlap on Winograd-Style Tasks

An Analysis of Dataset Overlap on Winograd-Style Tasks

A Semantic-based Method for Unsupervised Commonsense Question Answering

John praised Mary because _he_? Implicit Causality Bias and Its Interaction with Explicit Cues in LMs

Contact Info

Product

Resources

About