Enhance the Aspect Category Detection in Arabic Language using AraBERT and Text Augmentation
Miada Almasre · 2022
Research into Arabic Aspect Category Detection (ACD), a subtask of Aspect-Based Sentiment Analysis (ABSA) is primarily focused on experimenting with Machine Learning (ML) and Deep Learning (DL) models, but rarely consider methods for robust dataset creation or augmentation. Therefore, this study investigates and compares two applications of AraBERT vr2 on the Arabic SemEval2016 hotel dataset, where the objective is to predict aspect categories in the reviews. The baseline pipeline uses the AraBERT classification layer to predict 34 aspect categories, and the proposed pipeline performs a data augmentation process prior to classification, where the HARD dataset is used with a Word2Vec model to extend the original training of the SemEval2016 dataset. Both implementations of AraBERT were evaluated on Gold Standard dataset derived from SemEval2016. Results indicate that both applications achieved a closely similar f1-score of 0.663 for the baseline and 0.661 for the augmented model. The by-class results are further discussed.