Probably not too many of those, but enough to be annoying?
With this kind of function, you generally need a very low error rate (less than 1%) for it to be considered acceptable.
In a scenario where the glossary is used on a broad front and the models are well-described, there will easily be half a dozen glossary highlights in any given notes field. With an error rate of, say, 5%, this means you'll find an incorrect glossary highlight every four or five elements you look at. Granted, not every highlight will be a plural form, but if you throw in participles, tenses, and a few other fun things along those lines it very quickly breaks down completely.
My point is that guesswork is not going to be good enough for this. What is needed is the ability to use a proper dictionary (which provides forms of words, not just their respective meanings) in conjunction with the glossary. The users can then select the appropriate dictionary for their language and culture.
/Uffe