raw reviews
"loved this movie"
"not worth your time"
"terrible acting"
"beautiful story"
labels: + / - / - / +
tokenize
+ negate
features
[loved, movie]
[not, NOT_worth, NOT_your]
[terrible, acting]
[beautiful, story]
+ bigrams: "not_worth", ...
+ tf-idf weight
fit
classifier
Naive Bayes
or Logistic Reg
learns:
w[terrible] < 0
w[loved] > 0
w[NOT_worth] < 0
+
−
keep stopwords. handle negation. bigrams catch what unigrams miss.