Notes from discussion of the *SEM shared task on scope of negation.
Moderator: oe Present: oe, Emily, Bec, Guy, Woodley, Glenn, Rui, Joshua, Sanghoun Scribe: Emily
oe: etc etc etc [Oslo participated with a non-semantic system.]
Woodley: Forsooth!
oe: Oslo discussion: seems like it would be helpful to have the semantic analyses from the ERG, what would it take to do that?
oe: The annotation allows substrings and mwe cues, as well as discontinuous, nested and overlapping scope.
Woodley: Isn’t that a clue that the scope trees might not have been the right way to go?
oe: Oslo discussion: came around to the idea of using the resolved scopes (being surprised), and then worked on heuristics for choosing among the resolved scopes.
Emily: Then Woodley tried working out the heuristics and came to the conclusion (and convinced me) that the fully scoped versions are not useful.
Woodley: The main thing you’re doing by picking among the scoped representations is picking which of the quantifiers are in v. out of scope of the negators, not much else. The thing the annotators were doing were basically taking all arguments and modifiers of the negated events. That is a semantic thing but it doesn’t play well with brining quantifiers inside the scope of negation logically because of the constraint that every variable has to be bound. VP coordination — all predicates involving some
{Kim} arrived and <didn't> {sing}
… want Kim inside the scope of negation in order to get that, but that brings arrive inside the scope of negation as well.
Emily: But if we don’t bother with the fully scoped version and crawl the MRS graph we can get something useful.
Woodley: Genau. Once we don’t try to pretend that the ERG’s notion of scope and the annotator’s are the same, then there are lots of interesting things to do.
Only one fully scoped reading for Kim arrived and didn’t sing.: and neg is inside proper_q.
Woodley [channeling Dan]: The annotators notion of scope was totally rubbish and should be discarded. But in terms of the idea that we want to move the quantifier scopes around to bring the NPs inside the logical scope of negation even in the case of examples without coordination: The cloud is not fluffy: not refuting the existence of a cloud. We have both of those readings, but one is much more sensible.
Woodley: If I were talking to someone uncooperative (glance at Bec) and refuting all along the notion of there being a chair and my interlocutor said The chair is fluffy, I could say NOT the chair is fluffy.
oe: Two approaches to take: trying to win the competition retroactively and trying to do a comparative analysis of notions of semantics.
Emily: One more: exploring whether a semantic dependencies based system has complementary strengths to the existing syntax-based winning system, through error analysis.
Woodley: In order to do that error analysis, you first have to try to win retroactively.
oe: I actually hold this data set in high regard as a shared task data set — someone has tried to do something deep and done so carefully. As have we, so the chances of complementarity are maybe high.
Emily: One thing we haven’t talked about is the transfer rules idea to bring in the harder cases including without and the morphological negation.
oe: Two ways to go: talking about the crawling around the MRSs, and talking about expected exceptions (transfer rules).
[discussion of terms to use for connected bit of the MRS corresponding to the scope unit.]
oe: How do we walk/crawl the MRS to find the scope unit.
Woodley: Start with the neg rel, and walk down the argument links.
oe: Suggesting making minimal examples of the phenomena we will encounter:
{it} <didn't> {rain} ;; one argument to the neg rel, and that's all there is; a short crawl
{the fierce dog} <didn't> {arrive} ;; follow argument to the neg rel, then walk its arguments
so far walking the semantic dependencies and grab everything that is reachable: reachable: ARGs and shared labels — maybe inverse arguments. (crawling arguments v. crawling functors)
The probably blue dog didn’t bark.
… Dan is hanging the handles in a helpful way, so we still don’t need functor crawling.
Emily: So far it looks like argument crawling and label sharing is enough …
Bec: … to find the *EP*s.
Kim arrived and didn’t sing.
Emily/oe: How do we get the quantifiers? Need functor crawling?
Woodley: when we crawl ARG1-3, look for all things that take those variables as ARG0 (in the case where they are variables). That gets the noun-y rels and the quantifiers.
Emily/oe: Looks like coordination is not causing problems for the crawling method.
Woodley: One we were getting wrong before, but maybe because of the wrong parse was It may be that you are not yourself luminous
… surely enough there isn’t the parse that we want.
Woodley: Another one: You will observe that he could not have been on the staff of the hospital. PP attachment mistake?
Emily: with of?
oe: Investigating — don’t see one where of attaches high.
Woodley: Current implementation probably isn’t following the shared labels, and it should be.
oe: Next case:
{Kim} promised Sandy {to} <not> {sing}
Emily: not -> sing -> Kim, but Kim’s label isn’t shared with promise and Sandy isn’t an argument of sing, so no problems. (But if we drop pronoun_q and have pron_n share its label with the verb…)
Next case:
<No> {chair arrived in Berlin}
oe: Now we functor crawl: look for all EPs that take the BV of the no_q as an argument.
Woodley: Situations where multiple EPs take the same x are coordination and relative clauses.
Emily: And control.
oe: 13108 at no great distance — only scopes over the NP inside the PP, and nothing higher (not even the P). So here the functor crawling strategy needs more refinement.
Guy: Proposes the strong hypothesis that there might be some formally well-defined substructure of the MRS that corresponds to what the annotators were doing.
Maybe problematic:
{There were certainly} <no> {chairs}.
Functor crawl out of the quantifier, but don’t functor crawl out of the verb, and so we don’t get to certain_a.
oe: Maybe even a surprising alignment between our emerging approach and what the annotators did.
Sanghoun: What about double negation?
oe/bec/rui: I think they’d do:
{{i} can} <not> {do <no>{thing}}
oe: 13108 is the only example so far that’s causing problems.
various: Maybe it’s enough to say that we only functor crawl out to certain types of EPs (e.g., with preds matching _v_rel).
Woodley: Generally seems to work pretty well — can find further funny cases by implementing these refinements.
bec: Did you do an ep-to-string thing?
Woodley: Yes … opens laptop
Last update: 2013-08-02 by EmilyBender [edit]