Positional Embedding and Context Representation for Indonesian Targeted Aspect-based Sentiment Analysis

Inspired by some research on entity level and aspect level sentiment analysis, we build a two-step classification system to tackle the challenges of targeted aspect-based sentiment analysis (TABSA). The system consists of aspect categorization and sentiment classification. We employ a Bi-LSTM network, position features i.e. context representation and positional embedding on both classifiers, and one-hot aspect vectors in sentiment classifier to give additional information. We conducted experiments using 900 Indonesian reviews related to automotive. Despite not using complex feature because of limitation in low-resource language, the system can still make decent prediction. We found that the use of positional embedding and context representation in our model is proven to be able to capture the relation between entity and aspect-sentiment. The use of one-hot aspect vectors was also found to be slightly helpful for sentiment classifier to capture polarity related to a specific aspect. Experiment results showed that our approach outperformed the self-training baseline in terms of F1 and accuracy scores.