An Overview Of Temporal Commonsense Reasoning and Acquisition

Temporal commonsense reasoning refers to the ability to understand the typical temporal context of phrases, actions, and events, and use it to reason over problems requiring such knowledge. This trait is essential in temporal natural language processing tasks, with possible applications such as timeline summarization, temporal question answering, and temporal natural language inference. Recent research on the performance of large language models suggests that, although they are adept at generating syntactically correct sentences and solving classification tasks, they often take shortcuts in their reasoning and fall prey to simple linguistic traps. This article provides an overview of research in the domain of temporal commonsense reasoning, particularly focusing on enhancing language model performance through a variety of augmentations and their evaluation across a growing number of datasets. However, these augmented models still struggle to approach human performance on reasoning tasks over temporal common sense properties, such as the typical occurrence times, orderings, or durations of events. We further emphasize the need for careful interpretation of research to guard against overpromising evaluation results in light of the shallow reasoning present in transformers.

Introduction. Humans generally perform well in interpreting implicit information in text and speech by leveraging commonsense reasoning. This ability is reflected in the way we communicate. For example, when we read the phrase “I couldn’t get out of bed this morning.”, we generally assume that this refers to a state of mind and not a physical inability to get out of bed. When we read “He had butterflies in his stomach.”, we understand this as a figure of speech for an anxious or nervous feeling. Rather than specifying the literal meaning, we rely on the recipient’s implicit prior understanding of certain concepts and expressions in our language. Commonsense reasoning can manifest in different forms. Datasets such as CIDER (Ghosal et al, 2021a), Cosmos QA (Huang et al, 2019), GLUCOSE (Mostafazadeh et al, 2020), and COM2SENSE (Singh et al, 2021) aim to serve as benchmarks to better understand the commonsense reasoning capabilities of current state-of-the-art machine learning models.

Discussion / Conclusion. In this section, our goal is to highlight the results of our survey and propose possible future research opportunities. In this survey, we have highlighted several similarities between early temporal algebras, temporal reasoning tasks such as temporal relation extraction and event ordering, and proposed TCS reasoning dimensions. We pose that the main difference between TCS reasoning and other temporal knowledge which a model may have (such as temporal factual knowledge or reasoning capabilities over an explicit temporal context) is the inherently probabilistic nature of common sense. While we can make assumptions about the likely order, duration, or time of occurrence of individual actions, common sense does not make any guarantees. By design, the context in a TCS task does not give us a concrete answer, but we should use our prior understanding of the world to derive a likely one.