Analysis of Emotion Detection from Code Mixed or Code-Switched Social Media Text using Deep Learning

Emotion detection plays a vital role in understanding user sentiment specially in this the era of digital communication. Due to exponential growth of internet and social media, social media became a huge source of information which gives insight in varieties of applications. Emotion detection is an emerging field of research. Users express their implicit feelings, thoughts on social media. Analyzing such data becomes challenging due to linguistic diversity, especially in code-mixed or code-switched content involving transliterated Hindi and English. This paper addresses the problem of emotion detection from such complex textual data. We have explored and compared performances of different deep learning models in handling code-mixed social media text. Our experiments demonstrate that among all tested architectures, the hybrid LSTM-CNN model shown the highest mean test accuracy of 79.60% and 0.4216 F1-score without data balancing. For balanced data CNN gives the highest accuracy of 77.79%, while the Bi-LSTM model gives the highest F1-score of 0.4978. This research demonstrates the effectiveness of deep learning for emotion detection in transliterated Hindi-English social media posts.