Improve tokenization to reduce dimensionality

[karlhigley]

This removes tokens that only occur once (which are irrelevant to computing cosine similarity) and strips out non-alphabetic characters (which can lead to double-counting essentially the same token).