MulTSeA: Aspect-based Sentiment Analysis using Multitask Learning from Bengali Texts

Aspect category identification and aspect sentiment classification are the two main sub-tasks of aspect-based sentiment analysis (ABSA). Most existing studies address these two issues separately, i.e., when performing aspect sentiment classification, it is assumed that the aspect categories are pre-identified. This approach needs to be more practical and adequate. This work proposes a unified multi-task learning framework to overcome these limitations. The proposed approach combines a BiGRU (Bidirectional Gated Recurrent Unit) with a CNN architecture in the shared layers of the multi-task learning model, utilizing FastText word embeddings. The proposed approach is evaluated using three benchmark datasets: the Restaurant, Cricket, and BAN-ABSA. The evaluation results demonstrate superior performance across all three datasets, achieving F1-scores of 0.59 for aspect category classification and 0.65 for sentiment classification on the Restaurant dataset, 0.58 for aspect category classification, and 0.73 for sentiment classification on the Cricket dataset; and 0.86 for aspect category classification and 0.75 for sentiment classification on the BAN-ABSA dataset. Overall, the results demonstrate the effectiveness of the proposed model in performing aspect-based sentiment analysis tasks on Bengali, achieving competitive performance across the benchmark datasets.