Sentence Level Emotion Tagging on Blog and News Corpora

摘要

Here we report a sentence level emotion tagging system based on the word level constituents in Bengali blog and English news corpora. An emotion-annotated Bengali blog corpus was prepared manually with Ekman's six emotion tags using the emotion annotated English news headline corpus from SemEval 2007. The word level annotation is carried out semi-automatically. The baseline system at word level for each emotion class assigns the class label to each word. The Conditional Random Field (CRF) based classifier used for word level emotion tagging outperformed the baseline for each emotion class. Sentence-level emotion scores for each emotion class are calculated as the average word level emotion scores based on the SentiWordNet The emotion tag with the highest score is assigned to the sentence, followed by a rule based post-processing technique for handling negative words. The system demonstrated the highest overall average 65% F-Score value for Bengali and 63.26% for English.