Shortcutted Commonsense: Data Spuriousness in Deep Learning of Commonsense Reasoning

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1504–1521 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 1504 Shortcutted Commonsense: Data Spuriousness in Deep Learning of Commonsense Reasoning Ruben Branco and António Branco and João Silva and João Rodrigues University of Lisbon NLX – Natural Language and Speech Group, Department of Informatics Faculdade de Ciências, Campo Grande, 1749-016 Lisboa, Portugal {rmbranco,ambranco,jrsilva,jarodrigues}@fc.ul.pt Abstract Commonsense is a quintessential human ca- pacity that has been a core challenge to Arti- ficial Intelligence since its inception. Impres- sive results in Natural Language Processing tasks, including in commonsense reasoning, have consistently been achieved with Trans- former neural language models, even match- ing or surpassing human performance in some benchmarks. Recently, some of these ad- vances have been called into question: so called data artifacts in the training data have been made evident as spurious correlations and shallow shortcuts that in some cases are leveraging these outstanding results. In this paper we seek to further pursue this analysis into the realm of commonsense re- lated language processing tasks. We undertake a study on different prominent benchmarks that involve commonsense reasoning, along a number of key stress experiments, thus seek- ing to gain insight on whether the models are learning transferable generalizations intrinsic to the problem at stake or just taking advan- tage of incidental shortcuts in the data items. The results obtained indicate that most datasets experimented with are problematic, with models resorting to non-robust features and appearing not to be learning and general- izing towards the overall tasks intended to be conveyed or exemplified by the datasets. 1 Introduction Reasoning helps humans to cope with their expe- rience, whether in a complex situation such as de- vising a plan to solve a pandemic or in a simple, intuitive inference like what will happen to a cof- fee mug when dropped to the ground. It helps to foresee new events as well as to form new beliefs and justify and defend them before others (Kintsch and Van Dijk, 1978; Mercier and Sperber, 2017). Reasoning with knowledge widely shared by hu- mans, usually termed as commonsense, makes up a significant portion of our higher level cognitive skills: commonsense knowledge encompasses hu- man values and needs, and by reasoning with it we can organize sensible arguments and decide on ef- fective actions. Endowing computing devices with these knowledge and reasoning capabilities should allow them to better get access to world view of humans, their needs, capabilities, beliefs, and thus allow them to act more appropriately. To a large extent, acting on the basis of these capabilities can be verbalized, knowledge can be written down, and the inference chain of a reason- ing process can be expressed in some human idiom. In Natural Language Processing (NLP), common- sense processing tasks usually involve answering questions that, when addressed by humans, require different types of commonsense knowledge and rea- soning to be answered. Accordingly, for a model coping with this kind of tasks to achieve good …

Content truncated to 500 words