Aspect Sentiment Classification for the Libyan Dialect in the Aviation Sector Using AraBERT and MARBERT

Resumo

This research aims to enhance aspect-based sentiment analysis (ABSA) for the aviation industry by developing a robust framework for identifying passenger feedback in low-resource dialects. The study is grounded in the framework of Natural Language Processing (NLP) and transformer-based architectures, specifically leveraging transfer learning fundamentals to process linguistic nuances. The methodology involves fine-tuning two sophisticated pre-trained language models, AraBERTv2 and MARBERT, using the ASALDA corpus, a specialised dataset comprising Libyan dialect tweets and comments regarding airline services. To address data scarcity, data augmentation techniques were implemented to refine the training process. The results demonstrate that the fine-tuned MARBERT and AraBERTv2 models achieved a significant accuracy score of 92% in aspect sentiment classification (ASC) to determine sentiment toward each aspect (like "food", "plane", or "customer services"). These findings provide actionable insights for improving operational efficiency and service quality in the airline sector. This study contributes to the field by bridging the gap between academic Transformer models and industrial applications, offering a scalable solution for monitoring customer satisfaction in diverse linguistic environments. The proposed models prove highly effective in processing dialectal Arabic that is low-resource by delivering aspect sentiment classification transformer-based models for the Libyan dialect.