I vaguely remember someone mentioning this (or something similar) in the past, but I never looked it up.
Based on the paper, their AI was playing Freeciv (but using the Civ II user guide to get information). Presumably this was so that they could program interactions in such a way that games could be played by the computer much more quickly.
Since their benchmark is beating the Freeciv AI, (I don't know how good it is), I don't think they came up with a "great" civ 2 playing AI. Rather, the point of their research seems to be that they have a way to train a neural net faster by using text from the manual. At least I think, I've only skimmed the article, and only have a superficial knowledge of this kind of stuff anyway.