Multi-Classes Emotion Detection for Unbalanced Indonesian Tweets

Research concerning Twitter mining is becoming an exciting research topic in recent years. Emotion detection is one such research area that uses microblogs, such as Twitter, to discover emotions from textual data. There was a modern technique using graph-based was introduced to extract patterns that bear emotion. The system has achieved good performance in different western languages. By adopting the approach, we are motivated to enhance the performance of emotion detection for Indonesian language. We use eight emotions, i.e., joy (senang), sadness (sedih), fear (takut), surprise (terkejut), disgust (jijik), anticipation (antisipasi), trust (percaya), and anger (marah). The data distribution among the emotions is unbalanced, making the Indonesian language system has low precision. This study proposed an adjusting pattern weight to address the unbalanced data problem for the Indonesian language. The experiment results show that the proposed approach can significantly improve the precision of minority classes (i.e., fear, surprise, and disgust).