Lexical type prediction

Here and here we talked about the possibility of applying neural approaches to lexical type prediction.

I am now in a good position to actually try something like that, however I need to step back an understand better when and how we do this lexical type prediction currently, in our current processes (e.g. in the LKB or in ACE).

Was it described somewhere in enough detail? I looked at everything I could find on e.g. packing under subsumption, but those papers have a different angle (they desribe the bidirectional subsumption check).

What do we mean by this lexical type prediction, where and when does it happen, and how would I observe it? @johnca , @sweaglesw , I am assuming you have some debugging procedures in place which allow you to do that?

(Or am I on the wrong track completely and whenever we talked about lexical type prediction, we were not talking about exhaustive parsting at all and were talking instead about supertagging?)

Your final sentence is correct, I think. The aim of supertagging (or ubertagging) is to predict lexical types. In that case, parsing will not explore the full search space. Here are a couple of references to DELPH-IN work on this topic:

R Dridan, V Kordoni, J Nicholson
Enhancing performance of lexicalised grammars
Proceedings of ACL-08: HLT, 613-621

R Dridan
Ubertagging: Joint segmentation and supertagging for English
Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, 1201-1212

ACE and PET support ubertagging; at present the LKB does not.

1 Like