Scary films good, scary flights bad - topic driven feature selection for classification of sentiment.

ABSTRACTThis paper describes preliminary work on feature selection for classification of review text by both sentiment rating and topic. The premise stems from the notion that one size does not fit all; that feature sets for sentiment analysis should be tailored to the topic of a text. Thus it naturally follows that for this to be effective it is also necessary to first determine the topic of a text. Following successful work on classification of texts by author demographics, a corpus of review texts labelled with attributed rating, topic area, and user demographics has been compiled. This collection was divided for this work into different topic groups in order to automatically classify between both text topic and subjective rating. By using a single supervised statistical approach to feature selection, it is shown that improvements can be made to classification accuracy using topic tuned features sets over more generic features.