Document Level Emotion Tagging: Machine Learning and Resource Based Approach

Abstract

Abstract. The present task involves the identification of emotions from Bengali blog documents using two separate approaches. The first one is a machine learning approach that accumulates document level information from sentences obtained from word level granular detail whereas the second one is a resource based approach that considers the Bengali WordNet Affect, the word level Bengali affective lexical resource. In the first approach, the Support Vector Machine (SVM) classifier is employed to perform the word level classification. Sense weight based average scoring technique determines the sentential emotion scores based on the word level emotion tagged constituents. The cumulative summation of sentential emotion scores is assigned to each document considering the combinations of various heuristic features. The second one implements a majority based approach to classify a given document considering the Bengali WordNet Affect lists. Instead of assigning a single emotion tag to a document, in both approaches, the best two emotion tags are assigned to each document according to the ordered emotion scores obtained. By applying the best feature combination acquired from the development set, the evaluation of 110 test documents yields the average F-Scores of 59.50% and 51.07% for the two approaches respectively with respect to all emotion classes.