[util]

One thing I'm unclear on: Did they use the training data in assessing accuracy? (Sorry if I missed this in my reading.) If so, it might be interesting to see how well they do on a separate test set. Also, it would be interesting to see the results broken down conditioning both on the estimated sentiment and on the true sentiment -- "precision" and "recall" by type.

[danielhfrank]

Hey, I wrote the piece so I'm happy to address your question. First off, the training data was completely separate from test data, which was painstakingly gathered by hand. We wanted to make sure the data we were testing against was classified as accurately as possible. For training data, volume is more of a priority. Your other point is very well taken-- we'll probably use some metrics like that ourselves to look for improvements! Let me know if you've got any other questions, I'm happy to clear things up.

[util]

Cool, thanks! Very interesting stuff. Nice of you to share some of the "secret sauce."