Multi-Label Emotion Classification using Sparse Long Short-Term Memory with Sigt Activation Function

The most significant task in Natural Language Processing (NLP) is emotion classification, which is used to recognize and detect various textual emotions.However, the existing methods identify multiple emotions in a single text, which is challenging due to the ambiguity of language and the lack of clearly defined boundaries between emotions.The Sparse Long Short-Term Memory with Sigt (SLSTM-Sigt) activation function is proposed in this research for multi-label emotion classification.The SLSTM helps reduce the norm of the weight vector to improve generalization.The SLSTM captures all the non-linear relationships inherent in the data, which are necessary for emotion classification.Moreover, the pruning and regrowth depend on the training phase, which ensures that the model adapts to the underlying data during the training process and produces better performance.The integration of the activation function enhances the gradient flow through the network, which prevents vanishing gradients and helps to increase the learning ability of the model through longer sequences.Metrics like precision, accuracy, recall, macro F1, and micro F1 are used to estimate the SLSTM-Sigt performance.The SLSTM-Sigt achieved an accuracy of 96.53% and 98.27% for the RE-CECps and SemEval-2018 Task 1: EC datasets, respectively, which is better than Robust Bidirectional Encoder Representations from Transformers-Multi-Attention (RoBERTa-MA).