Infer information from Tweets. Useful for human-centered computing tasks, such as sentiment analysis, location prediction, authorship profiling and more!
Instead of letting the Naive Bayes classifier generate a score by multiplying the conditional probabilities together, I read somewhere (I can't remember at the moment) that they count the number of features (words) that are highly confident for a particular class. For example:
8 highly positive features minus3 highly negative features, the result is +3, so the label is positive.
If the value is within some range around 0, then it could be labeled as neutral. Could also consider moderately confident values (or even low confidence), but in any case would need to determine the thresholds for pigeonholing the conditional probabilities into confidence levels.
Instead of letting the Naive Bayes classifier generate a score by multiplying the conditional probabilities together, I read somewhere (I can't remember at the moment) that they count the number of features (words) that are highly confident for a particular class. For example:
8highly positive features minus3highly negative features, the result is+3, so the label is positive.If the value is within some range around
0, then it could be labeled as neutral. Could also consider moderately confident values (or even low confidence), but in any case would need to determine the thresholds for pigeonholing the conditional probabilities into confidence levels.Idea was originally recorded 3/12/13 at 23:01.