Abstract: Emotion detection in text is a complex yet vital component of natural language processing, significantly contributing to enhanced human-computer interaction. This research investigates the effectiveness of various word embedding techniques—Fast Text, RoBERTa, and Glo Ve—when combined with Convolutional Neural Networks (CNN) for detecting emotions. The study categorizes emotions into five types: happiness, anger, sadness, fear, and surprise. Three datasets are analyzed: one comprising movie reviews, another consisting of customer feedback from e-commerce platforms, and a hybrid dataset merging the two. Results indicate that RoBERTa+CNN outperforms other combinations, achieving accuracy rates of 89.45%, 90.12 % and 89.87% on the respective datasets. FastText+CNN is the second-best performer, while GloVe+CNN achieves the lowest accuracy. Additionally, evaluation metrics such as Precision, Recall, and F1-Score highlight the superior performance of RoBERTa+CNN in text-based emotion detection. This study underscores the value of contextual embeddings like RoBERTa in improving the reliability of emotion recognition models.
Keywords: Emotion Detection, Text Classification, Word Embeddings, RoBERTa, FastText, GloVe, Convolutional Neural Networks.
|
DOI:
10.17148/IJARCCE.2025.14228
