Extracting emotion from resource poor language through transfer learning

The development of textual emotion detection has coincided with the expansion of social media and communication networks, where a wide range of emotions are often conveyed. These expressions not only correspond to human emotions but also provide useful insights in a variety of circumstances. People all around the world can now express themselves through social media platforms like Twitter and Facebook. However, determining the emotional impact of textual information remains a difficult task due to the writing style of the language. In this study, we provide a novel method for categorizing Roman Urdu text into six different emotional categories, revolutionizing the field of emotion detection. To achieve the goal of our research, the proposed technique extracts emotions from Roman Urdu text using XLM-R (Roberta). The results from our study show that the proposed model outperforms other machine learning models for emotion detection when applied to the Roman Urdu corpus. With 18,000 reviews, our study has the most substantial Roman Urdu corpus to date, highlighting the uniqueness and relevance of our work. The comparison shows the XLM-R technique’s advantage over existing approaches. Our results show that our model has excellent capabilities, with an F1-score of 85 and an accuracy of 84 percent on the Roman Urdu Emotion Detection corpus.