Sentence level emotion tagging

Abstract

This paper reports the mechanism of sentence level emotion identification based on emotion tagged word level constituents acquired by an automatic classifier applied on the SemEval 2007 Affect Sensing corpus. Basic set of six emotion types, namely, happy, sad, anger, disgust, fear and surprise have been selected for reliable and semi-automatic word level annotation. WordNet Affect lists have been preprocessed using SentiWordNet information for use in the semi-automatic word level emotion annotation process. The Conditional Random Field (CRF) based word level emotion classification has yielded an accuracy of 87.65% on a test set of 250 sentences. Sense based scoring mechanism has been applied for calculating scores of a sentence for each of the six emotion types. Probable sentence level emotion tags have been assigned based on the system produced ordered sense scores. Post-processing strategies have been adopted for handling negative words in sentence level emotion tagging. The best two emotion tags, with the maximum sense scores, have been assigned to 250 test sentences and an accuracy of 67.2% has been achieved. The sentence level valence has been calculated based on the total sense score of the word level emotion tags. Accuracy, precision and recall are 60.47%, 67.95 and 65.11 respectively for valence identification on 250 test sentences. ©2009 IEEE.