In this work we describe a system to perform sentiment classication based
on an unsupervised linguistic approach that uses natural language processing
techniques to extract individual words from reviews in social network
sites. Our pattern-based method applies classication rules for positive or
negative sentiments depending on its overall score calculated with the aid
of SentiWordNet. Searching for the best classication procedure, we investigated
several classier models created from a combinations of different
methods applied at word and review level; the most relevant among them
has been then enhanced with additional linguistically-driven functionalities,
such as spelling correction, emoticons, exclamations and negation detection.
Furthermore, an empirical study on Word Sense Disambiguation has been
conducted on a set of test sentences extracted from the SemCor Corpus. We
dened two gloss-centered word sense disambiguation techniques which rely
on overlaps and semantic relatedness calculated on disambiguated glosses'
denitions provided by eXtended WordNet. Experimental results conrmed
that Word Sense Disambiguation can improve sentiment classication performance;
moreover, they indicated that all the words potentially carry emotions,
including nouns.