Aspect Based Sentiment Analysis (ABSA) is a sentiment analysis technique that not only identifies overall sentiment, but also reveals opinions on specific aspects of an entity. To facilitate computer processing, a numerical representation of words into vectors (word embedding) is used, where each word or phrase is mapped into a vector of dimension N. Although static embedding such as Word2Vec or GloVe has been widely used, these approaches have limitations in capturing the dynamic context essential for deep sentiment analysis. This research develops and tests several deep learning algorithms, namely CNN, Bi-LSTM, CNN+BiLSTM, and CNN+BiLSTM+Attention Mechanism, which initially use static embedding and then modified by integrating BERT as contextual embedding. The results show that the use of BERT improves sentiment prediction accuracy by 15% and aspect prediction accuracy by 11% compared to models with static embedding. In particular, the combination of BERT+CNN obtained the best accuracy, which was 94% for aspect prediction and the combination of BERT+CNN+BiLSTM+Attention Mechanism 87% for sentiment prediction. These findings demonstrate the significant potential of BERT integration in improving ABSA performance, which can be applied in social media opinion analysis and sentiment-based recommendation systems.
Turangan et al. (Sat,) studied this question.