Sentence-Level Emotion and Valence Tagging

The paper proposes the tagging of sentence-level emotion and valence based on the word-level constituents on the SemEval 2007 affect sensing news corpus. The baseline system for each emotion class assigns the class label to each word, while the WordNet Affect lists updated using the SentiWordNet were also used as the lexicon-based system. Though the inclusion of morphology into the lexicon-based system improves the performance of the word-level emotion tagging, the Conditional Random Field-based machine-learning framework was employed for the word-level emotion-tagging system, and it outperforms both the baseline- and lexicon-based systems. Six separate sense scores for six emotion types are calculated from the SentiWordNet and applied to word-level emotion tagged constituents for identifying sentential emotion scores. Three emotion scoring methods followed by a post-processing technique were employed for identifying the sentence-level emotion tags. In addition to that, the best two emotion tags corresponding to the maximum obtained sense scores are assigned to the sentences, whereas the sentence-level valence is identified based on the total sense scores of the word-level emotion tags along with their polarity. Evaluation was carried out with respect to the best two emotion tags on 250 gold standard test sentences and achieved satisfactory results for sentence-level emotion and valence tagging.