In recent years, emotion detection in text has become increasingly popular because of its many potential applications in a range of areas, such as marketing, political science, psychology, human-computer interaction, artificial intelligence. Access to huge amounts of textual data, especially opinionated and self-expression text, has also contributed to bringing attention to this field. Here, we first review work that has already been done in identifying emotion expressions in text and then proceed to argue that existing techniques, methodologies, and models are incapable of capturing the nuance of emotional language. This is mostly because by using handcrafted features and lexicons, they lose the sequential information inherent in the text and are unable to capture the context. Because existing methods cannot grasp the intricacy of emotional expressions, they are insufficient for creating a reliable and generalizable methodology for emotion detection. Understanding these limitations, we developed a deep neural network model with bidirectional Gated Recurrent Units (GRUs) and an attention mechanism that does consider the sequential information of text and that can capture the contextual meaning of words. Because our emotion detection model captures a more informative representation of the text, its performance is significantly better than conventional machine learning models. Specifically, our model increases the F-measure on the test data by 26.8 points, and by 38.6 points on a dataset never seen before. We also compared our model to fine-tuned transformer model (BERT), and found that the performance was slightly better specially using emotional embeddings, and importantly, required only a fraction of the computational power. In addition to this model, we also developed a new methodology for creating emotionally fitted embeddings, and showed that they can perform up to 11 percent better compared to standard embedding models in cosine similarity metrics, and furthermore, can improve performance of emotion detection models.
