Emotion analysis using self-training on malaysian code-mixed twitter data

Abstract

Microblogs such as Twitter has made available a vast resource of User Generated Content (UGC) on which emotion analysis may be performed. Organizations increasingly value the opinions obtained from emotion analysis. These insights help to drive decision-making activities and provide constructive inputs to engage their customers and services. In a multiracial country such as Malaysia, it is common to find that tweets are written in mixed languages of Malay, Malaysian slang and English. These tweets increase the complexity of the emotion analysis task, especially considering that there is a serious lack of labeled data available in order to make use of supervised learning techniques. This paper explores the use of self-training, a semi-supervised technique that only requires a small initial labeled dataset to conduct emotion analysis of Malaysian code-mixed Twitter data. The results are promising as the accuracy achieved is higher compared to the baseline models.