You're discussing part-of-speech tagging of natural language here, which is considered "solved" (almost solved) and for which there exist very robust programs already. I linked to some part of speech tagging methods in post #31, namely Brill and HMM. They basically have a forest and maximize the probability of several trees producing the given phrase.
This already works very well (depending on the vocabulary, slang, out-of-dictionary forms etc). What the discussion should focus on is, if in natural language, this a) is useful and b) takes away the subtlety of our language, like jokes, literature or even "sexual" ambiguous remarks.