Deep-EmoRU: mining emotions from roman urdu text using deep learning ensemble

Detecting emotions play a vital role in our lives. In various ways, people convey their feelings, i.e., facial expressions, movements, speech, and text. This study aims to classify the emotions from Roman Urdu’s text. Much research has previously been done on different emotion detection languages, but there is minimal work done in Roman Urdu. There is also a need to explore Roman Urdu, as it is the most widely used social media site for communication. The absence of a benchmark corpus for emotion detection from text is a significant problem for Roman Urdu because language assets are essential for various tasks of natural language processing (NLP). The emotional analysis has many practical applications, such as optimizing product quality, dialog systems, investment patterns, and mental health. In this research, we build a corpus of 18k sentences collected from different domains and annotate it with six other classes to concentrate on the emotional polarity of the Roman Urdu text. We also proposed a Deep-EmoRU model for emotion detection from Roman Urdu text. Our proposed model is based on Long short-term memory (LSTM) and Convolutional neural network (CNN) feature learners. We applied different baseline algorithms like LSTM, Adaboost, XGboost, Random Forest, MLP, SVM, Decision tree, and KNN on our corpus. After experimentation and evaluation, the results showed that our model achieves a better F-measure score than LSTM, KNN, SVM, Adaboost, XGboost, MLP, Decision tree, and Random Forest. We achieve an accuracy of 82.2% and an F-measure of 0.82 on Emotion Detection for Roman Urdu.