[bbayer]

That was a great read. As I understand OP's main concern here is accuracy. What about performance? NLTK is good start point and it deemed as slow.I really like to hear about runtime performance for same functionality with Python and R.

[jkldotio]

I think NLTK is pure python because it's supposed to work as a teaching tool as well. Perhaps pypy can come into play for that, although in real world situations I've only see marginal gains from pypy so far.