Enhancing Multi-Emotion Detection in Text: A Comparative Study of Feature Extraction Techniques and Machine Learning Models
Meherunnesa Hossain Ibnath, Khan Md. Hasib, Md. Rahid Parvez, M. F. Mridha · 2025
Emotion detection in text is one of the major tasks in NLP that might have tremendous applications in sentiment analysis, human-computer interaction, and mental health diagnosis. This study evaluates the performance of various feature extraction techniques and classification models for multi-emotion detection, using a dataset of 15,996 samples divided with six emotion labels: Sadness, Joy, Love, Anger, Fear, and Surprise. Three feature extraction methods—TF-IDF, Count Vectorization, and N-grams—were applied to preprocess the text data. Six models—KNN, SVM, NB, DT, LR, and BiLSTM—were then used to classify emotions. The best results among the 18 feature-model combinations are TF-IDF with SVM at 86.15%, Count Vectorization with Logistic Regression at 86.25%, and N-grams with SVM at 86.06%. N-grams combined with LR also performed well, achieving an accuracy of 88.09%, underlining the importance of capturing contextual and sequential information in text classification. This research contributes to improving NLP-based emotion detection systems and provides insights for future work, particularly in domains requiring nuanced emotion recognition from text data.