Enhancing dialectal Arabic aspect based sentiment analysis through a novel Algerian dialect telecommunication dataset using a unified end-to-end approach

Abstract

This study introduces a novel, publicly available balanced Algerian Arabic Aspect Based Sentiment Analysis (ABSA) dataset consisting of 11,338 comments written in the Algerian dialect. The dataset follows the semantic evaluation 2016 annotation guidelines and includes both explicit and implicit aspects within the telecommunications domain. Furthermore, the study proposes a unified End-to-End (E2E) framework for Arabic ABSA based on the newly introduced dataset and the Arabic ABSA hotels dataset, where aspect term extraction and sentiment classification are integrated into a single sequence labeling task. Transfer learning was leveraged by fine-tuning the Arabic Bidirectional Encoder Representations from Transformers (AraBERT) model, which was further enhanced with Bidirectional Gated Recurrent Units (BiGRU) and a Softmax output layer, forming the fine-tuned AraBERT-BiGRU-Softmax model. Using the unified E2E ABSA approach, the proposed model achieved an overall accuracy of 88.07% on our dataset, along with a macro precision of 72.96%, a macro recall of 60.79%, and a macro F1-score of 65.48% across all labels. When excluding the O label, the model obtained a micro-precision of 54.94%, a micro-recall of 53.85%, and a micro-F1 score of 54.39%. Evaluated on the Arabic ABSA hotel reviews dataset, the model obtained an overall accuracy of 92.91%, with a macro precision of 57.79%, a macro recall of 47.04%, and a macro F1-score of 50.69% across all labels. In addition, when excluding the O label, it reached a micro-precision of 64.74%, a micro-recall of 55.74%, and a micro-F1 score of 59.90%.