Automatic classification of emotions in news articles through ensemble decision tree classification techniques

摘要

Emotions form a major role in human life. As human interactions with online systems have increased drastically, emotion prediction from online text, which otherwise can be monotonous, would help to provide a better environment to the users. Identification of emotions from a normal text itself is very complicated while news text that does not explicitly convey emotions adds more intricacy to it. Data mining methods can be utilized in this context. In this work, the potential of decision tree classifiers in emotion classification is explored. The advocated methodology incorporates two segments towards emotion identification. The first segment deals with data preparation and involves dataset elicitation, translation, HTML tag removal, stop word elimination and stemming. The second segment that implements data mining takes the output of the first segment as its input and applies feature vector formulation, correlation based feature selection, building of bagged Grafted C4.5 learning model and performance evaluation. Based on the evolved classification rules, the emotions are categorized into joy, surprise, fear, sadness, disgust, neutral and mixed kind. Experiments have been conducted to analyse the effect of feature selection methods and ensemble methods in generating efficient rules. The accuracy is compared against eight other decision tree classifiers and also the support vector machine learning model. The proposed methodology achieves the maximum accuracy of 87.83% justifying its utilization in the real time applications.