Methods for building twitter-specific sentiment lexicon

Using mutual information

This page has quite a bit of information regarding how the mutual information should be computed. I don't think there is any difference between multi-class mutual information and binary mutual information except that the entropy of the class labels is affected.

Using Counts

I could use a simple method based upon counts of occurances of words both in the negative and in the positive tweets with normalization for frequency of terms

Using Feature Selection Methods

Mutual Information

The issue with using mutual information for this is that, ideally, we would want a two tailed statistical test, while mutual information is a one tailed test. To fix this, there are two different strategies that I am going to try.

Winner take all Mutual-Information

The winner take all mutual-information is the mutual information calculated for the positive class and the negative class, taking the larger of these two (negative mutual information will be multiplied by -1 to create a two tailed distribution of scores).

Proportional Mutual-Information

In proportional mutual information, the mutual information for the negative class si subtracted from the mutual information for the positive class.