Synthesized feature space for multiclass emotion classification

Emotion classification is an active area of research. It is difficult to differentiate among different emotion categories. This paper deals with classification of emotions into seven different categories such as (i) Anger, (ii) Disgust, (iii) Fear, (iv) Guilt, (v) Joy, (vi) Sadness, and (vii) Shame using ngram (unigram, bigram, and trigram) features. The features are extracted from the ISEAR dataset and WordNet Affect database. Three feature selection techniques, Weighted Log-Likelihood, Normalized Google Distance and Mutual Information, are used to extract the relevant features from the initial large and redundant feature space. A Multinomial Naïve Bayes (MNB) classifier trained on these features extracted from ISEAR dataset is used to predict the emotion at sentence level. The proposed MNB emotion classifier provides an average accuracy of 71.35% across the seven emotion classes for 450 unigram features.