You don't need any training data to do better than Porter2, just lexical data. If you're content to map the same surface form to the same stem regardless of context, as Porter2 would, what you need is a dictionary mapping words to their stems.

This used to be considered a lot of memory to use. It's not anymore.

This leaves you with some edge cases (is "axes" the plural of "axis" or "axe"?), but it's much better than mixing up the words "universe" and "university".

That's a good point, I think I overstated the need for training data.