Emotion Intensity Value Prediction with Machine Learning Approach on Twitter

Abstract

Recognizing the intensity of the emotions is a paramount task for an affective system. By recognizing the intensity of the emotions, the system can have better human-computer interaction. The research explores several machine learning approaches with several different feature extraction method combinations to solve the emotion intensity prediction task while also analyzing and comparing it with several previous related papers. The research uses the dataset provided through theWASSA 2017 and SemEval 2018 competition. The dataset utilizes four of the eight basic emotions that Plutchik defines (anger, fear, joy, and sadness). The total data result in 19,736 rows of entry, with a total of 10,715 (54.3%) for training, 1,811 (9.17%) for validation, and 7,210 (36.53%) for testing. Three feature extraction methods are used and compared: N-gram, TFIDF, and Bag-of-Words. Meanwhile, machine learning algorithms are Linear Regression, Ridge Regression, KNearest Neighbor for Regression, Regression Tree, and Support Vector Regression (SVR). The results show that SVR with TF-IDF features has the best result of all attempted experiments, with a Pearson correlation score of 0.755 for all data and 0.647 for gold labels data. The final model also accepts newly seen data and displays the corresponding emotion label and intensity.