Emotion Recognition From Text Using Deep Learning

Abstract

The existing research on emotion recognition from text has primarily focused on sentiment analysis, providing results in terms of polarity or numerical ratings. While these studies have significantly contributed to our understanding of sentiment in textual data, they often miss the complexity of human emotions. Our contribution begins with the Vietnamese Social Media Emotion Corpus (UIT-VSMEC), a dataset that includes 6,927 sentences annotated with various emotions. This corpus is particularly significant as it addresses a critical gap in emotion recognition research for the Vietnamese language, an area that has been underexplored until now. Building on this foundational work, we conducted an extensive evaluation of advanced deep neural network models using the UIT-VSMEC corpus. Unlike prior studies that limited their experiments to four machine learning algorithms—Support Vector Machine (SVM), Random Forest, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)—our research introduces an innovative approach. We employ Reinforcement Learning with Active Learning for the Pretrained Language Model designed for Vietnamese, known as PhoBERT and VisoBERT, in combination with a Long Short-Term Memory network. This sophisticated methodology not only leverages the strengths of these advanced models but also integrates active learning techniques to enhance the model's performance iteratively. Our approach demonstrates a significant improvement in emotion recognition accuracy. Among the evaluated models, our methodology achieved a weighted F1-score of 63%, marking a notable 6% improvement with the PhoBERT model compared to previous studies. This enhancement underscores the effectiveness of our approach and its potential for setting new criteria in the field of emotion recognition for Vietnamese text. Additionally, the integration of PhoBERT and VisoBERT with LSTM combined with Reinforcement Learning and Active Learning highlights the potential of combining non-classification methods such as Reinforcement Learning and Active Learning with advanced neural network architectures to achieve superior results in classification. This work paves the way for future research to explore more nuanced and accurate emotion detection methods, ultimately contributing to the broader field of natural language processing and its applications in sentiment analysis.