If I add Tagging, I can eliminate the interface corpus.TagHandler in favor of ObjectHandler<Tagging>. One longer-term benefit of this is that it’s easy to write generic cross-validating corpora (which is one case where you need the pairs of inputs and to be iterated). A downside is that you can only implement ObjectHandler for one generic type — how would I handle taggings and chunkings?

If I kept string arrays as outputs instead of Tagging, it’d certainly mean less fiddling with the existing HMM classes. But there’s no way to generify the handlers because TagHandler is really tied to char sequence inputs.

I would say: go with the list. The Tagging interface doesn’t seem to provide anything interesting, unless you plan to add something to it the future. I guess one thing that would make life easier in some cases was the case if there was a way to access pairs easily, i.e. if the Tagging interface had a method that returned List<Pair> but this seems also stupid.

Question 2:
I would prefer ScoredObject. Again, unless you plan to add something to ScoredTagging in the future.

I’m not actually using LingPipe, that’s why I cannot really answer Questions 3 and 4 and also don’t take my recommendation too seriously.