Enhanced BERT Approach to Score Arabic Essay’s Relevance to the Prompt

Abstract

In recent years, automated essay scoring systems have seen significant progress, particularly with the integration of deep learning algorithms. This shift marks a move away from the traditional focus on style and grammar to a more in-depth analysis of text content. Despite these advancements, there remains a limited exploration of the essay’s relevance to the prompts, especially in the context of the Arabic language. In response to this lack, we propose a novel approach for scoring the relevance between essays and prompts. Specifically, our aim is to assign a score reflecting the degree of adequacy of the student’s long answer to the open-ended question. Our Arabic-language proposal builds upon AraBERT, the Arabic version of BERT, and enhanced with specially developed handcrafted features. On a positive note, our approach yielded promising results, showing a correlation rate of 0.88 with human scores.