Identifying Expressions of Emotion in Text

Features	Naïve Bayes
SVM
GI	71.45%	71.33%
WN-Affect	70.16%	70.58%
GI+WN-Affect	71.7%	73.89%
ALL	72.08%	73.89%

³ http://www.cse.unt.edu/~rada/affectivetext/data/WordNetAffectEmotionLists.tar.gz
all the features were combined. While the use of non-lexical features does not seem to affect results of SVM, it did increase the accuracy of the Naïve Bayes classifier. This suggests that a combination of features is needed to improve emotion classification results.
The results of the automatic emotion classification experiments show how external knowledge resources can be leveraged in identifying emotion-related words in text. We note, however, that lexical coverage of these resources may be limited, given the informal nature of online discourse. For instance, one of the most frequent words used for happiness in the corpus is the acronym “lol”, which does not appear in any of these resources. In future experiments, we plan to augment the word lists obtained from GI and WordNet-Affect with such words. Furthermore, in our experiments, we have not addressed the case of typographical errors and orthographic features (for e.g. “soo sweeet”) that express or emphasize emotion in text.
We also note that the use of emotion-related words is not the sole means of expressing emotion. Often a sentence, which otherwise may not have an emotional word, may become emotion-bearing depending on the context or underlying semantic meaning. Consider (8), for instance, which implicitly expresses fear without the use of any emotion bearing word.
(8) What if nothing goes as planned?
Therefore to be able to accurately classify emotion, we need to do contextual and semantic analysis as well.

6.