What is the "discourse" predicate trying to say?

If I parse “Hi Lexi!” I get this MRS:

[ LTOP: h0 INDEX: e2 [ e SF: prop ] 
RELS: < 
[ greet<0:2> LBL: h4 CARG: "hi" ARG0: i6 ]  
[ discourse<0:2> LBL: h1 ARG0: i7 ARG1: h4 ARG2: h8 ]  
[ unknown<3:8> LBL: h8 ARG0: e2 ARG: x9 [ x PERS: 3 NUM: sg IND: + ] ]  
[ proper_q<3:8> LBL: h10 ARG0: x9 RSTR: h11 BODY: h12 ]  
[ named<3:8> LBL: h13 CARG: "Lexi" ARG0: x9 ] > 
HCONS: < h0 qeq h1 h11 qeq h13 > ICONS: < > ]

The greet predicate is nice since it removes having to deal with all the greeting words separately.

However, it is unclear to me what the discourse predicate is trying to tell me. Since it is tying together “Hi” and “Lexi” I assume it is being used as a predicate that represents “a thing that is said at someone that doesn’t have any meaning other than chit chat”???

Still trying to understand what the discourse predicate is signifying if anyone has any pointers.

I tried doing a search of the Redwoods treebank using this query:

h:discourse[]

Unfortunately every example but one that I found was the grammar marking a double-dash “m-dash” (i.e. “–”) as “discourse” so it wasn’t very helpful. In fact, it was more confusing since this is the first I’ve seen punctuation explicitly codified in a predicate.

Any tips on the meaning of the predicate or how to build a query that excludes all the “–” discourse predicates in the treebank would be appreciated!

OK, since there are a few things that are confusing about this parse I thought I’d list them out in case anyone has insight on any of them, or that some are just a grammar bugs:

  1. (as above) it is unclear to me what role the discourse predicate is playing here
  2. discourse has an i variable as ARG0. I understand this happens sometimes for scopal operators, but I guess I expected it to have an e intrinsic variable, like subord[ehh] does (or to have them both have i). Maybe the grammar is just in transition and everything isn’t consistent yet?
  3. greet has an i variable for ARG1 which in my experience so far would indicate a dropped argument but in this case I expected this to be x9 (i.e. the person being said “Hi” too). Maybe this is a side effect of the fact that discourse is in there?
  4. The unknown predicate has an arg called “ARG” without a number after it. First I’ve seen that…

It’s like a conjunction.

The greet relation needs to be connected to the rest of the MRS. In this case, the rest of the MRS is a fragment (indicated by unknown), but you can also add “Hi” to the beginning of a normal sentence, e.g. “Hi can I help?” – then you’ll get the semantics of “Can I help?” connected to the greet rel via the discourse rel.

So this is the same as with an en-dash or em-dash, where we need to connect the two parts of the MRS. There will be some kind of discourse relation between them, but it’s highly underspecified, hence the discourse rel.

1 Like

Ahhh…I was trying to figure out how discourse could possibly be used for “hello” and an m-dash. Thinking about it as a kind of conjunction helps, thanks @guyemerson .

The discourse seems to be related to both vocative and discourse dependency relations In Universal Dependencies guideline:

as @guyemerson, it is highly underspecified.

I think UD’s vocative is closer to the ERG’s addressee rel.

I didn’t know about addressee… Thank you.

Thanks @arademaker! I haven’t spent any time browsing Universal Dependencies but it seems like it could be a great way to describe at least some of the ERG predicates with very low overhead. A page that simply shows a mapping of ERG predicates to UD tags where there is a close mapping would be really useful to me at least!