BERT with Reinforced Embedding and Enhanced Adapters to Score Coherence in Arabic Essay

Coherence and cohesion evaluation play a fundamental role in automated essay scoring systems, contributing to an in-depth analysis of textual quality. Although significant progress has been made in this field, the automatic evaluation of Arabic essays remains challenging due to its linguistic specificities. In this context, we introduce a multi-task model designed to simultaneously evaluate the coherence and cohesion of Arabic essays, by predicting two distinct scores. However, this study focuses exclusively on coherence analysis, without exploring the direct impact of cohesion at this stage. We leveraged the pre-trained Arabic BERT model (AraBERT) by integrating two task-specific adapters. To refine the coherence evaluation, we extracted handcrafted features and injected them into the dedicated adapter. Additionally, we enhanced the embedding layer by incorporating a holistic representation of the input, allowing for a better capture of textual relationships at a broader level to improve coherence evaluation. This approach led to promising results, achieving a correlation of 0.81 for coherence.