Training a unigram part-of-speech tagger