Python implememtation of Soft-Tfidf

Improved over Cohen's original similarity metric in a few ways:

• Appropriate weighting given to terms that don't exist in the corpus.
• Close similarity terms use the partner term idf if it is available.
• No longer a symmetry or overflow problem with the results.
• Also includes a second class called SemiSofTfidf that attempts to apply the soft-tfidf concept to information retrieval.