A Transformer-Based Approach with Data Augmentation for Multilabel Emotional Context Detection

Emotion identification in texts is becoming increasingly difficult because of the wide variety of ways emotions are represented. This study uses a fine-tuned Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa) to offer a Transformer-based model for identifying multilabel emotional context in textual data. To balance emotion categories and enhance the model's capacity for generalization, data augmentation is applied on two different datasets: Semantic Evaluation and Cross-lingual Emotion Dataset (SemEval and XED) English corpus. This stage is considered one of the most important steps in preprocessing as it greatly helps to improve the results. The RoBERTa model was then used to extract and comprehend the deep context of emotional expressions. The proposed model shows robust performance on both SemEval-2018 and XED datasets with F1-macro/F1-micro of 0.924/0.93, 0.707/0.748, and 0.729/0.77 on single, double and triple emotions respectively. Further, it gives consistent results on XED with F1-macro/F1-micro of 0.846/0.846 on single emotions, 0.819/0.833 for double emotions and 0.825/0.834 for triple emotions.