Emotion-specific features for classifying emotions in story text

Abstract

In this work, we are attempting emotion classification in view of synthesizing story speech. We are proposing emotion-specific text features (ESF) for classifying sentences from children stories into five different emotion categories: happy, sad, anger, fear and neutral. ESF is a five dimensional feature vector, where each dimension corresponds to weight of the sentence according to each emotion class. The dataset consists of 780 Hindi emotional sentences collected from children stories belonging to three genres: fable, folk-tale and legend. Part-of-speech (POS) and proposed ESF are used as features for emotion classification. Emotion classification performance is analysed using various combinations of features with three classifiers: Naive Bayes (NB), k-nearest neighbour (KNN) and support vector machine (SVM). The effectiveness of classifiers is analysed using precision, recall, F-measure and accuracy. The classification performance of 67.9% and 67.2% is achieved using POS and ESF respectively. The fusion of both features resulted an accuracy of 71.1%. Further, the importance of story genre information in emotion classification was observed from the experiments conducted on classifying emotions within story genre. An accuracy of 73.7% was observed after adding story genre information to the fusion of POS and ESF. SVM models outperformed other models in terms of classification accuracy.