LADUW20170110

Francis: (Background:) Fragments in the ERG are handled with an unknown-rel, which has an argument, which unifies the fragment with some other thing.

Fragments in Zhong: some verb fragments are currently handled with e.g. subject dropping. There are also adjectives (colors, "sleepy") that can do this. Currently(?), these are handled as fragments rather than dropped arguments, but Francis wants to know whether anybody has any thoughts. He might ask Dan.

There's also fragments like "to Tokyo," in which case the unknown-rel is an event.

He's also noticed that fragments in JACY are handled with arg-0 coindexing, which should probably be fixed.

Emily: Semantically, we should have entity fragments and we should have predicate fragments. If we admit noun-modifying PP fragments, we want them to assimilate to either "I am just a PP fragment" or that they are modifying unknown-rel(? NB: not sure if I caught this last bit).

Francis:

Emily: You could say that subject and comps are a non-empty list.

ZhenZhen: Currently we're treating them as having an optional subject and complement. The non-verbal ones are treated as fragments.

Emily: It might be that some verbs are happy to drop their complements and some aren't, in which case you can use a binary feature (OPT +/-) to control it. Then you will only get a fragment analysis with the ones that really want their arguments.

ZhenZhen: So the real work lies in figuring that out and separating the types of verbs.

Emily: Yes, the question lies in whether you want to model that. The gain here is that fragments are a root, so you can turn it on and off.

ZhenZhen: Having this handled in the grammar makes it a lot more robust.

Francis: But, on the other hand, adds a lot of ambiguity.

ZhenZhen: Particularly with coordination.

Francis: The next question I had, which is perhaps too vague... we find we're often not making as much use of the POS hierarchy as we could, because we're scared of changing the head type hierarchy. For example, if we want to split the preposition hierarchy into two types.

Emily: There is LISP code that generates all of that, and if you change it, you can have all the types that you could ever possibly want (automatically).

Francis: We could be making much more use of subtypes.

Emily: I agree it's not pretty, and there's a reason it's segregated off into a separate file.

Francis: Third question. If we want to add a doc string to a type, how do you do it? We want to start adding these.

(long discussion about whether doc strings were removed, encouraged, discouraged, and in which grammars they occurred or were taken out.)

Francis: So maybe I'll email the developers, just to double check.

Emily (finishing an email search): It looks like in 2007 we ended up deciding to use a single semicolon after the type name...?

Francis: No, we've moved on since then. I'll email the developers to confirm.

Glenn: Another problem with that was that those strings were inserted as a string into the grammar itself, which polluted the string list that was part of the grammar.

Francis: I think in the LKB, it's stored as a separate thing, called "comment."

(A further question about fragments gets asked here.)

Francis: In "Are you going on Tuesday?" "Possibly," we want "possibly" (the fragment) to modify "going on Tuesday." It needs to fit in the right spot.

(And now back to comments and documentation)

Glenn: The thing with checking comments using version control is that some of the types doing a lot of work have very complicated histories.

Francis: I'd like these to be easy for me to maintain. The string in the beginning is the documentation that we'll show to people normally. One or two notes with the types, and then further details afterward. What I want to make very easy for people to do is to put in documentation with the (example?) sentences, and documentation for the types.

Glenn: The biggest failing in commenting in general is that if it goes out of date, you've done yourself a detriment.

Francis: I'd like to delete some of the really old JACY comments. Strings of messages back and forth...

Emily: Even there, there might be some benefit...

Francis: Which is why I haven't removed them.