Multi-Label Emotion Classification Using Deep Learning on the Go-Emotions Dataset

This has increased the urgency to have proper emotion recognition structures that could recognize more than one co-occurring emotion due to the proliferation of textual information on social media generated by users. The purpose of this work is to create a deep learning model that will perform multi-label emotion classification with the use of Go-Emotions dataset. The overall aim is to identify five fundamental emotions, including anger, fear, joy, sadness, and surprise, provided in the text, and overcome such problems as linguistic noise, context ambiguity, and an imbalance in classes. Its methodology includes a lot of text preprocessing, that is, normalization, lemmatization, emoji management, and balancing classes and then the feature representation with a Text Vectorization and embedding layer. BiLSTM architecture is used to obtain long-range forward and backward contextual dependencies. The model is trained on binary cross-entropy loss with the Adam optimizer and tested on the basis of accuracy, precision, recall, and loss. It has been shown that the suggested BiLSTM model has an accuracy of 90.61, a precision of 90.41, and a recall of 92.25, and performs much better than traditional machine learning models and an existing mBERT+BiLSTM model. The results validate the hypothesis that a well-developed BiLSTM model with organized preprocessing offers a scalable and able performance in relation to multi-label emotion classification, which can be applied effectively to actual emotion-sensitive NLP tasks.