Making Sentiment Analysis Nontrivial and Interesting
Eduard Hovy
Carnegie Mellon University 

Sentiment detection / opinion mining (and a new direction called emotion detection) have been very popular topics in NLP for almost a decade.  But the absence of careful definitions for core concepts like ‘sentiment’ and ‘emotion’ has hampered the current dominant approaches to detecting the presence and valence of sentiments / emotions.  In this talk I address mismatches between typical computational studies and real-world uses, and try to provide good definitions, sketch out the requisite intellectual framework to address this problem, and suggest an interesting direction that goes beyond what is being done today in sentiment/emotion mining that could facilitate the construction of effective advertisements.