Commonsense Validation and Explanation for Arabic Sentences

Commonsense understanding poses a significant challenge, especially in complex languages like Arabic. However, recent advancements in deep learning have facilitated improvements in various language tasks, including the ability to distinguish commonsense in sentences. This research focuses on participating in the SemEval 2020 Task 4 (ComVE) competition by developing classification and text generation models tailored for the Arabic language. The competition comprises three subtasks: Subtask A involves choosing the sentence that makes sense between two given sentences, Subtask B requires selecting the most appropriate reason from multiple choices for a sentence that goes against common sense, and Subtask C entails generating an explanation and reason for a sentence violating common sense. Our models leverage a set of multilingual pre-trained transformer models and have achieved remarkable performance in the competition. In Subtask A, our accuracy reached 84.7%, surpassing the performance of other works in Arabic. Similarly, in Subtask B, our approach outperformed other multilingual approaches, achieving a score of 79.3% compared to the state-of-the-art BERT model’s 61%. In Subtask C, our model generated explanations with a BLEU score of 24, which is considered acceptable in the domain of text generation, particularly in the context of Arabic.