Gamon & Aue combine a semantic orientation method based on Turney's pairwise mutual information approach with an approach  based on the assumption that
terms with opposite orientation tend not to occur at the sentence level (may be in contradiction with Hazy & McKewon assumption that this can occur depending on particular syntactic contexts -- "interesting and useful" versus "beautiful but boring". They test this idea in a  classification task --  which consist on classifying sentences into positive, neutral, and negative -- achieving around 50% accuracy.


Dave presents several techniques to create features (words or terms) and associated scores from training corpora for a classification task which consist on
sifting positive and negative statements associated to
product reviews from Cnet and Amazon. They investigate various mechanisms to produce 
features -- the baseline being an unigram model and more complex models
employing lexical substitution, higher n-grams, and syntactic phrases  -- and weighting mechanisms such as inverted document frequency.
Their classifier aggregates features' scores for sentences and bases 
 the classification on the sign of the 
aggregated score. The use of simple n-grams seem to perform any other investigated
feature generation technique and  n-grams grater that 1 seem to perform better than unigrams. The proposed tecnique 
achieves over 80\% classification accuracy.

Investigate the issue of generating in an objective way a lexicon of expressions for
positive and negative opinion. They note that expressions such as "good" can be considered not so positive but quite negative in some contexts such as in e-commerce.
They investigate the correlation of the monetary gains with the
occurence of particular phrases in reviews of merchants obtaining in this way an objective ranking of phrases which inflence the monetary gain a
merchant can make according wit its reputation.