Optimized Sentiment Classification in Low-Resource Language Processing using Element-Wise Attention Gated Recurrent Unit with Bayesian Optimization

In recent days, Low-Resource Language Processing (LRLP) in sentiment classification has gained attention because of limited annotated data and tools, demanding efficient models for accurate emotional analysis in underrepresented languages. Traditional models for LRLP had faced challenges in capturing long-term dependencies in text. Therefore, this research proposed Element-Wise Attention Gated Recurrent based Bayesian Optimization (EWAG based BO) for LRLP in sentiment classification. Initially, data is collected from Arabic Sentiment Twitter Corpus dataset and preprocessed by using stemming and lemmatization which normalize words to their root and dictionary forms by reducing vocabulary size. Then, feature extraction is done by using SentiBERT which generates the context-aware embeddings by capturing sentiment-specific semantics. After that, sentiment classification is done by EWAG which enhances GRU through applying element-wise attention to focus on crucial input features that improved model’s ability in capturing context and increase classification accuracy. To further improve performance, BO is utilized hyperparameter tuning that efficiently explores search space to find optimal model settings. The proposed EWAG-BO effectively achieved better results when compared with existing LRLP-Improved Deep Learning Hunt Prey Optimization (LRLP-IDLHPO) in terms of accuracy (99.50%), precision (99.20%) and recall (99.18%) respectively.