search

bwbaugh / infertweet

Infer information from Tweets. Useful for human-centered computing tasks, such as sentiment analysis, location prediction, authorship profiling and more!
http://infertweet.bwbaugh.com/
Other
10 stars 1 forks source link

Confidence level for subjective documents #37

Closed bwbaugh closed 11 years ago

bwbaugh commented 11 years ago

Using our hierarchical classifier, if we are 100% sure that a document is positive, but only 60% sure that the document is subjective, aren't we really only 60% sure that the document is positive? If so, we should be updating the confidence values of the polarity classifier to incorporate the output of the subjectivity classifier.

If we do this, the color output that we show for polarity will be affected. For example, a previously bright green color for 100% positive would be closer to yellow or orange if the subjectivity classifier was only 60% sure the document was subjective. We could either leave that as it is, to show that the overall confidence is lower, or we could use just the confidence from the polarity classifier for coloring purposes.

bwbaugh commented 11 years ago

After discussing this in class with my professor, it seems like multiplying the probabilities together should be done if they are in fact probabilities. I'm more of the belief that the output is a confidence value rather than a true probability. However, I still think it would be worthwhile to make the change as a 100% confident positive document with 60% confidence in being subjective really should be different than a 100% confident positive document with 90% confidence in being subjective.