Arabic Temporal Common Sense Understanding

Natural language understanding (NLU) includes temporal text understanding, which can be complex and encompasses temporal common sense understanding. There are many challenges in comprehending common sense within a text. Currently, there is a limited number of datasets containing temporal common sense in English and there is an absence of such datasets specifically for the Arabic language. In this study, an Arabic dataset was constructed based on an available English dataset. This dataset is considered a valuable resource for the Arabic community. Consequently, different multilingual pre-trained language models (PLMs) were applied to both the English and new Arabic datasets. Based on this, the effectiveness of these models in Arabic and English is compared and discussed. After analyzing the errors, a new categorization of errors was proposed. Finally, the ability of the PLMs to understand the input text and predict temporal features was evaluated. Through this detailed categorization of errors and classification of temporal elements, this study establishes a comprehensive framework aimed at clarifying the specific challenges encountered by PLMs in temporal common sense understanding (TCU). This methodology underscores the urgent need for further research on PLMs’ capabilities for TCU tasks.