Multi-Label Emotion Classification Using Multi-Attribute Bidirectional Encoder Representations from Transformers-based Constructive Loss Function

Emotion classification is a key task in Natural Language Processing (NLP), with an extensive range of downstream applications such as brand management, emotional chatbots, public health etc.,.Compared with primary sentiment polarity classification, emotion classification offers a finer level of granularity, encompassing emotions of hate, love, and surprise.However, multi-label emotion classification struggles with accurately determining and categorizing overlapping emotions in different and context-rich data, which results in potential misinterpretation and nuanced emotional insights loss.Therefore, this research proposes a Multi-Attribute Bidirectional Encoder Representations from Transformers-based Constructive Loss Function (MBERT-CLF) to classify the multi-label emotion accurately.Integrating a multi-attribute with CLR in BERT enhances the model's ability to capture nuanced relationships among emotions which enables more effective management of overlapping emotional labels.In preprocessing, tokenization, hashtag removal, and hyperlink removal are applied to break down text into manageable units by removing irrelevant noise and distractions.Then, the skip-gram is performed to extract features by capturing contextual relationships among words.When compared to the existing methods like RobustBERT-Multi-Attention (RoBERTa-MA), the proposed MBERT-CLF achieves a better accuracy of 97.45%, 95.72%, and 89.20% on the SemEval-2018 Task 1:EC, RE-CECps, and GoEmotions datasets, respectively.