Sarcasm Detection Using Sentiment Flow Shifts

Filatova, Elena (City University of New York, New York City College of Technology)

AAAI Conferences 

One of the most frequently cited sarcasm realizations is the use of positive sentiment within negative context. We propose a novel approach towards modeling a sentiment context of a document via the sequence of sentiment labels assigned to its sentences. We demonstrate that the sentiment flow shifts (from negative to positive and from positive to negative) can be used as reliable classification features for the task of sarcasm detection. Our classifier achieves the F 1 -measure of 0.7 for all reviews, going up to 0.9 for the reviews with high star ratings (positive reviews), which are the reviews that are materially affected by the presence of sarcasm in the text.