Create a well-formed MRS structure from: "what are you holding?"

I’m trying to create a well-formed MRS structure from this simple mrs (from: “what are you holding?”):

[ TOP: h0 INDEX: e2 [ e SF: ques ]
[ unknown<0:21> LBL: h1 ARG: x4 [ x PERS: 3 ] ARG0: e2 ]
[ free_relative_q<0:4> LBL: h5 ARG0: x4 RSTR: h6 BODY: h7 ]
[ thing<0:4> LBL: h8 ARG0: x4 ]
[ _be_v_id<5:8> LBL: h8 ARG0: e9 [ e SF: prop TENSE: pres MOOD: indicative PROG: - PERF: - ] ARG1: x4 ARG2: x10 [ x PERS: 2 IND: + PT: std ] ]
[ pron<9:12> LBL: h11 ARG0: x10 ]
[ pronoun_q<9:12> LBL: h12 ARG0: x10 RSTR: h13 BODY: h14 ]
[ _hold_v_1<13:21> LBL: h15 ARG0: e16 [ e SF: prop TENSE: untensed MOOD: indicative PROG: + PERF: - ] ARG1: x4 ARG2: i17 ] >
HCONS: < h0 qeq h1 h6 qeq h8 h13 qeq h11 > ]

Normally I’d:

  1. Find handles that aren’t labels anywhere (which I’ve been calling ‘holes’): h0, h6, h7, h13, h14
  2. Find handles that aren’t used anywhere: h1, h11, h12, h15, h5, h8
  3. And start trying to assign unused handle to ‘holes’, paying attention to constraints and scope of instance variables

…except in this case I am short one ‘hole’.

The only thing that looks suspicious is: _hold_v_1... ARG2: i17 but since “i (for individual) is a generalization over eventualities and instances” I wouldn’t imagine it could hold a handle.

Any ideas?

I don’t think this is the parse you want. Note how there is an unknown rel, which the ERG introduces to allow fragments (e.g. try parsing “the dog”, or “the dog?”). The free_relative_q is for a construction like “what we saw surprised us”. So this parse means something like “the things which are you and which are holding something?” (a noun phrase being used as a question).

If you really do want this parse, the suspicious bit is h15, which should either be identified with h8 or connected via another rel such as subord.

What version of the ERG did you use to produce this MRS?

Yes, “hole” is a standard term, e.g. see Copestake et al. (2005).

And by “well-formed”, do you actually mean “scoped”? An MRS can underspecify scope and still be well-formed.

The ERG I’m using is erg-2018-osx-0.9.31.dat.

I’m trying to make sure my “MRS Solver” can create all valid fully scoped trees from any proper MRS structure so even if I wouldn’t use this in my game I want to make sure I don’t have a bug in my solver.

When you say "h15 , which should either be identified with h8" do you mean that the ACE parser should have given them both the same handle? I.e. it is a bug? Or is there some operation I need to do in addition to assigning handles to holes in this case?

Same question for “or connected via another rel such as ‘subord’” I’m not sure what you meant by that one. I thought the only thing that needed to be done to completely resolve an MRS structure was assigning handles to holes (while maintaining constraints and scope of variables). Are there more steps that need to be done sometimes?

Re: Well-formed: I was trying to use the Copestake et al (2005) definition (maybe improperly): “A well-formed MRS structure is an MRS structure that link-subsumes one or more scope-resolved MRS” which I thought meant that there was no underspecification left…

No, “link-subsumes” means that it is a possibly underspecified form that can become a scope resolved form by adding (not removing or replacing) links (or label equalities). You can imagine an MRS that can’t possibly become a valid scope-resolved form without removing some improper constraint, or perhaps the DAG is disconnected, etc. These would be ill-formed.

From section 4.1 in Copestake et al. 2005 (p294 in the PDF I’m looking at, but it’s the 14th page of the PDF):

1 Like

Remember this discussion? Resolving scope is a smaller problem than coming up with precise logical interpretations, but there’s a general point that not all of the ERG’s MRSs are of equally good quality. For well-understood constructions, you will get a good MRS; for difficult constructions, you will get a reasonable placeholder MRS that is logically questionable but good enough for many applications (e.g. modified quantifiers, like in “almost everyone left”); and in some cases, there are bugs in the grammar.

By this, I meant that the MRS is suspicious and it looked to me like a bug in the ERG, which is why I asked for the version. (Bug reports can be made here: .) It seems that v2018 has more parses than v1214. There are four readings with an unknown and free_relative_q, of which one has a subord, two have a nominalization, and the last one is the one you’ve posted about. It might be that this is a bug, or it might be that this is needed as a robust parse for some other kind of fragment.

1 Like

OK, @guyemerson I posted as a bug. Thanks for the link.

By the way, should I post any parse that doesn’t have an equal number of holes and free handles as a bug?

I wouldn’t say any parse, because: (1) it might be the same bug as before, and (2) some constructions require extending the notion of a scope tree, e.g. see here.

And as with reporting bugs for any software, the most helpful bug report gives a minimal example, not necessarily the example you originally found.

Thanks. I thought my example was minimal…I guess I don’t understand what minimal means in the case of the ERG. I assumed it meant “small sentence that doesn’t generate a huge complex MRS”. Is there something else I should shoot for?

Maybe “minimal” isn’t quite the right word. Certainly, “small sentence that doesn’t generate a huge complex MRS” is a good place to start. The trouble with this reading of this sentence is that it’s a fragment and quite weird, semantically, so it’s not immediately clear what’s going wrong (e.g. is being a fragment relevant? Or the copula? Or the pronoun? Or the question?).

@guyemerson in what sense is what are you holding a fragment?
Though I agree trying to find a similar sentence with the mismatch without as many of the question, pronoun, or auxiliary (not copula) phenomena would be good. Or at least trying to remove one and see if the problem still exists.

It’s a fragment for the intended reading (see MRS in first post), in which “what are you” is a free relative clause (so “are” is a copula, not an auxiliary), and “holding” is a modifier. The obvious reading has no problems with its MRS. In the GitHub issue, I have given another sentence illustrating the same problem but where the intended reading is clearer.

1 Like

Thanks for the bug report. There is indeed a bug in the grammar (both in the 2018 version and until now also in trunk), which Guy rightly notes has to do with a missing identity in the hook of the free relative phrase that the (reduced) relative construction modifies. More specifically, the construction for free relative clauses (hd-cl_fr-rel_c) fails to identify its LTOP with that of its free-relative pronoun daughter. I’ll document the fix on Github where it belongs once I have tested it.

So your understanding, Eric, is sound here: there should normally be the same number of holes and labels to fill them, so report a mismatch like this as a bug, I note BTW that in the upcoming next release (2020) of the ERG, the grammar no longer produces that counterintuitive analysis for your original example, since the free relative “what” does not actually like agreeing with a following plural VP (no wonder you had a hard time envisioning the motive for that analysis). But the bug remained in Guy’s variant “whoever sleeps barking naps”, so needed to be fixed in any case.

1 Like

@Dan, great. Thanks for the fix! I’ve been able to work around it by changing my MRS solver to get ignore any parses that have a hole/label count mismatch, but I’ll have it spit them out so I can report them to you guys.

For posterity: Here’s the algorithm I ended up using to scope resolve the trees in case someone stumbles on this thread and is curious.