You're right, but the problem is much worse than that. Examining 200 entries from Garner's Modern American Usage at random reveals that half of them are easy to implement, the kind of thing that could be assigned as a homework problem (e.g., recognizing that “$10 USD” is redundant, that “very unique” is comparing an uncomparable adjective, or that people from Michigan are called “Michiganders”, not “Michiganites”). Thirty percent are moderately challenging, requiring a week’s effort. Fifteen percent are hard — they are entire projects, requiring advances in AI. And the remaining advice (around five percent), the best kind, is AI-complete. Consider, e.g., "John hit Peter only in the nose". Does this mean that, of all Peter's body parts that could have been hit, John hit only Peter's nose? Or is it a grammatical error that was suppose to convey that, of all the people John could have hit, it was only Peter who he did hit.

We're interested in incorporating deeper NLP. In particular, we've been eyeing https://github.com/spacy-io/spaCy.