Benchmarking Supervived Learning Models For Emotion Analysis

150 150 MIMOS Berhad


Ang Jia Pheng, Duc Nghia Pham, and Ong Hong Hoe



Emotion is the most genuine reaction of a person towards a circumstance or an object that are usually hidden between lines in their speech, text and actions. While emotion is more sophisticated and complicated to process and analyze, emotion provides more detailed and valuable insights for organizations to re-evaluate/fine-tunetheir actions and make informed decisions. This paper benchmarkedsupervised learning models for emotion analysis using sentence-level documents with six categories: anger, sadness, joy, love, surprise, and fear. We evaluatedone deep learning model: Bidirectional Long-Short Term Memory (BiLSTM), and one machinelearning model: FastText. Since FastText does not support GPU, both BiLSTM and FastText were ranon CPU for a fair time comparison. The results showed that while sacrificing speed that tookat least 9000% longer to train and validate, BiLSTM consistentlyoutperformedFastText.We also found that text pre-processing helped boost the performance of supervised learning models in emotion analysis.



6th International Conference on Artificial Intelligence and Computer Science (AICS2019), Wuhan, Hubei, China