NLTK word_tokenize fix for Python 3