For this experiment, I've randomly selected 500 hotel reviews and classified them manually. A better option might be to use Amazon's Mechanical Turk (https://www.mturk.com/mturk/) to get more reviews classified than any one person might be able to do easily. Really, a few hundred is about the minimum that we'd like to use as both the training and test sets need to come from this. I made sure that the sample contained an equal number of positive and negative reviews. (You can find the sample in the
data directory of the code download.)
The data files are
tab-separated values (TSV). After being manually classified, each line had four fields: the classification as a
- sign, the date of the review, the title of the review, ...