TheAbbey_Chrysalis2014DeverbalNouns

---

The starting point for this discussion was the question of whether words like performance should be given a decomposed semantic representation so that there is an ARG1 position into which to slot the NP in a by PP on the ARG1 interpretation. Along the way, we decided that this is actually two orthogonal questions: should the semantic representation for deverbal nouns have an argument position one way or another for the argument that corresponds to the verbal subject, and should the semantic representation be _performance_n_rel or _perform_v_rel + nominalization_rel.

On the long walk to coffee, we started this topic, and noted the following interesting examples:

1. The performance and cafe by the symphony

2. Kim's arrival and rejection happened within 5 minutes

3. The performance and dance by Mary

4. The arrival and dance by Mary

(1) shows that though by the symphony can in principle be either a by-adjunct or the selected by that marks ARG1, the coordination with a non-relational noun forces only the adjunct reading. This means that there are cases where the grammar does disambiguate for us, so there is some gain to be had by ambiguating these two different uses of by PPs with relational nouns (as we already do with passive verbs).

My notes for (2) say "no hope for resolving poss_rel, but maybe hope for resolving by.

For (3), we believe that both performance and dance admit both the ARG1 and adjunct readings of the by phrase, but that the coordinated structure forces them to both combine with the by phrase semantically in the same way. So this is another case of the grammar constraining the interpretation that might otherwise be ambiguous.

My notes for (4) says that coordinating with arrival forces the COMPS list of dance down to one complement, though I can't just now reconstruct how this example demonstrates that.

Woodley asked whether these examples should best be analyzed with RNR.

---

[ Much more was of course discussed on that walk, but notes resume when we got back to the meeting room:]

Francis: Concern with performance as nominalization + perform_v_rel, not sure they mean the same things. Sometimes more than one nominalization with clear different meanings: replacement --- thing/person that is the new stand in, v. the act of replacing.

oe: Isn't Kim replaces Abrams ambiguous in the same way?

Francis: The two replacements have arrived. these are persons, not events. Not _v_rels, don't have ARG1 and ARG2.

Dan: The replacement had a long beard./The replacement took two months. Two different senses, not related to the verb having two senses. But: just an accident of the language. Need two entries. It's all handwork anyway.

Francis: Have somewhere a hierarchy of meaning relations giving a link between performance_n_rel and preform_v_rel; would be nice to have the argument sets harmonized.

Francis/Dan: If we say -ing and -ance and -ment forms are all perform_v_rel plus nominalization_rel we're losing some information, but possibly could be done by ambiguating nominalization_rel. Plain nominalization_rel is transparent, but nominalization_ment_rel flags that there might be some drift.

oe: And it's a way of having that extra EP do some work for us. Records which derivational process has occurred productively or historically.

Emily: Generator inputs?

Dan/oe: They can underspecify which nominalization_rel. Just like leaving the subsense field underspecified.

Woodley: On the assumption that you still want a nominalization_rel seems nice.

Francis: Other consumers might want to say that instruction is similar to pedagogy. Their lives become harder under this regime because there no longer is an instruction_n_rel.

Emily/Francis: nominalization_rel + perform_v gets us to a transparent link to perform_v, but that can be gotten by an explicit link between perforamance_n and perform_v, while maintaining the current ease of linking those to other things.

Emily: NomBank might be a useful resource for looking at relationship of deverbal noun argument sets to that of verbs.

Dan: Are there words like pedagogy that might behave like instruction:

The pedagogy/instruction of students by inexperienced teachers never goes very well

Francis: pedagogy isn't that word, but we may be able to find one.

Dan: I overstate the challenge. We don't have any particular reason not to have pedagogy take of and by complements.

Dan: Convinced myself along the walk that finding the guys that can take the by is the same problem is finding the set of relational nouns, if intransitive reverbal nouns never take by phrase complements. (contra Emily's judgments of the dive by the Estonian competitor.) Though there are relational nouns that are not deverbal.

Woodley: Do all deverbal nouns become derived, or do they stay in the lexicon?

Dan: They stay in the lexicon. Spelling is unpredictable, just making the semantics more transparent.

Woodley: Can't inherit from the verbs, because they're not in the same place in the lexical type hierarchy.

Woodley: Can't do that via a big irregs.tab?

Dan: Makes a much stronger claim, because the irregs table is just about spelling.

oe: Does this call for the LKB's old mechanism to record pairs lexical entry and a chain of lexical rules as first class information?

Dan: We could invoke that chain of rules and do something pretty artful with topicalization --- and do all of that in the semantics, but I'm not that excited about going down this road yet.

Francis: Possibly existing resources already have the links between those forms.

Dan: If we agree we want this kind of link for deverbal nouns, what about deadjectival nouns?

Emily: Kim's ability to sing --- relation to complement seems the same.

Dan: Sara's ability to admire herself. seems to support reflexivization.

Emily: Deadjectival nouns might be more straightforward than the deverbal nouns because they just use the same marking of the arguments. (Aside from possessive.)

Dan: Do we imagine that our imaginary consumers would thank us for this decomposition in the same way the do for deverbal nouns?

Emily: So you're already parsing ability is a relational noun. So the original motivation isn't there (can link the arguments), but once you have the decomposition for verbs, nice to have the symmetry for adjectives.

Woodley: The foolishness by Woodley was surprising.

Emily: I think that's a sentence, but I don't think that the by is marking ARG1 of foolish, but rather that there's an extra layer of something in there.

Woodley: The absence of Ann

Emily: Can that be deverbal? absent yourself from

Dan: Surely it's deadjectival.

Woodley: The absence of Ann seems syntactically clearly Ann = ARG1. by is harder.

Dan: I'm fine with those being argument taking guys, but do we want _absent_a_rel in the decomposition.

oe: Any reason to do it?

Dan: I guess I'm going to find some places where I'm not sure if it's a verb or an adjective, and there will be a sharp distinction.

Emily: But you have to make the choice either way, since you have to choose what to decompose it to.

Dan: True, but the difference will be smaller.

Dan/Woodley: Different subtype(s) of nominalization_rel for adjectives.

Francis: runnable --- and it's relationship to run Which I don't think we should decompose, but I still want to say there's some relation between them.

Dan: retry

Francis: I can reretry so unless the meaning has changed, want to handle that productively.

Emily: Feels like another jungle.

Francis: The question is how far do we want to go?

Dan: Doing the decomposition of these nominalizations opens that pandora's box, and we're never going to satisfy people entirely.

Francis: If we've opened the box, it seems that almost as regular and equally as popular is the causative/inchoative alternation…

Emily/Francis: Decomposition and more argument structure for relational nouns are separate questions.

Woodley: Seems related to red deer from the other day. performance is almost compositionally a nominalization of peform_v_rel but not quite. Why wouldn't you have both? (Schrödinger's MRS.)

Emily: But those were cases where there was something we wanted from each one? What are the things we want in this case?

Woodley: There must be something because we're arguing about it.

oe: That suggests it must be aesthetic, so why are we arguing about it.

Woodley: Can't tell what's at the bottom of the slippery slope --- ice shards or…

Emily: White striped no tailed tigers.

Dan/Woodley: We've been looking for those all day. Let's go!

Francis: There is no bottom.

Woodley: Really?

Francis: Every time I've tried to do this in dictionaries, I've never been able to see how to do it thoroughly. Never happy with the result. Like I said, I don't think that performance and peform_v_rel mean the same thing. By removing the distinction or making it harder to see, I think our grammar is not better. (Even understanding that we're just spelling performance_n_rel as nominalization_rel+perform_v_rel.) Worry about antidisestablishmentarianism anti- refers to the specific movement that disestablishmenttarianism referred to, not fully compositional. … whereas linking performance to perform, and is different from performing to perform.

Dan: I was persuaded by your reasoning that if the nominalization has a different sense, we have to write down a different predicate name, even if it's perform_v_1 v. perform_v_3. Why privilege that link over the link between peform_v and performance_n. WordNet is about to do these kinds of links across categories without batting an eye. Maybe we're better off focusing on capturing the information we can, and leave this carving up of the names and declaration of the hypothesis to later or someone else. If there is some well-behaved subset, can decompose that at a later date. But for now it's not really our problem because we could imagine that the resource that someone needs to make the link is one we can supply or point to, rather than making it look like we have a theory of something we don't.

oe: We had identified another argument that would support that position: the by can mark the ARG1, while the of can mark ARG1 or ARG2 as a complement. If you decompose the noun into something that has the verbal root, and pick up an of PP as a complement, …

Emily: That's skipping at step. The first observation was that by PPs, when marking arguments, can only mark ARG1, while of PPs, while marking arguments can mark either ARG1 or ARG2. We realized this could be a possible use for RMRS-style underspecification of roles, and then said that this could even be implemented in MRS by allowing (in external MRSs) hierarchies of role names. The grammar could give us what we need if we add features like ARG12 which are then interpreted in MRS-land as the disjunctive underspecification between ARG1 and ARG2. In some contexts, other information in the interpretation (such as the presence of an unambiguous ARG1) will constrain that further, but this is done externally to the grammar. We then realized that to do this for the _performance_n_rel representation, we would only need ARG1 and ARG12 as features, as there is (by hypothesis) never any unambiguous marking of ARG2 in nominal complements. [EMB notes later: marked by of. When there are other prepositions involved, it can be unambiguous, but maybe those are different nouns?] On the other hand, if we wanted to reuse _perform_v_rel, then we'd have to have ARG1, ARG2 and ARG12 for that one, because there is unambiguous marking of ARG2 for the verb.

Dan: Not so: Even if they have the same predicate symbol, the verb could have ARG1/ARG2, while the noun ARG1/ARG12.

oe: pick up of PP as complement to the relational noun, is that on the reading that rules out the agentive interpretation, and we leave the ARG1 role as part of the general poss_rel messiness, so maybe all of that was misguided.

Dan: Not sure I see the objection.

oe: The performance of the sonata complement/The performance of the orchestra not complement.

Dan: So far that sounds like religion. I'd like to see the test.

oe: Why do we have both here?

Dan: In general: picture of me/picture of mine: of mine --- adjunct/of me --- complement.

Emily: But can't say that performance of me, so how to we get from here to there.

Dan: It only shows that there are at least some places where that distinction is motivated.

Woodley: the rejection of his/the rejection of him

oe: of him can only be ARG2 of reject and cannot be the ARG1, and so there's no need for the underspecification in that case (with the pronouns).

Dan: You haven't shown that.

oe: I'm also assuming that the poss_rel interoperation is available from the adjunct parse, which can be interpreted either way.

Dan: If we can do something like Tuesday's rejection of Bill

Emily: Does that mean we can't have poss_rel for Bill?

Dan: Tuesday's rejection of his was unexpected

Emily: Sounds fine.

Woodley: Really?

Dan: him disambiguated, his remains ambiguous.

Francis: I'd like to reluctantly withdraw my objection to dive by: @The longest recorded dive by a whale (corpus example) plus lots of other examples easily turned up.

Woodley: I've never dove the Hawai`ian Islands.

Emily: Choking on the morphology. How about We finally dove the Hawai`ian Islands for the first time? Still sounds bad.

Dan: Too little too late. (Trying to make dive transitive to preserve the earlier generalization that deverbal nouns built from intransitives can't take by-marked ARG1.)

Francis: No sleep bys though.

Dan: The ARG12 is only a cuteness thing. Can still ambiguate in the grammar. If of can't mark ARG1, don't put that in, be happy.

Francis: 21 nap bys in 1.9 billion words in the new BYU corpus, but none are an agent.

Dan: We have picture of Bill -- ARG1, or should it be an ARG2.

oe: I picture it as an ARG2…

Dan: Then a violation of the generalization that we never have ARG2 without of ARG1.

Woodley: That picture by Ansel Adams --- picture as a deverbal noun from photograph.

Emily: Isn't that the authorial by? Keep the generalization, and say that picture has only ARG1.

Dan: Worried about difference between picture and depiction.

Emily: We are capturing just that difference.

Dan: Possible source of bewilderment for someone trying to create inputs for generation. Will have to always be on the lookout for nouns that have been verbs at some point. Whether we decompose or not, will not be able to get the name of the argument right.

oe: Is that almost an argument in favor of the decomposition, in that it would give us reliable argument mapping?

Emily/Francis: There could be other ways of doing it. In the links Francis creates, could also link the argument structures.

oe: So what are the strong arguments against the decomposition?

Dan: The slippery slope. How far back to do we go?

oe: The right amount.

oe: Wasn't it the more the merrier where we actually have a systematic correspondence of argument structure.

Emily: Either way they have to look up in the SEM-I or otherwise, to see if we decomposed, so can't they see then how many arguments there are.

oe: But wouldn't having the verbal rel be a good guide?

Dan: Someone giving us a generator input for The performance of the musicians was inept would have to know to use poss_rel.

oe: They'll just get performance by the musicians.

Dan: They'll have to look in the SEM-I either way.

Emily/Francis: Then if we have explicit links between the arguments of the _n_rels and _v_rels that helps to.

oe: Helps with creating MRSs but not with underspecification. Don't we want performance/performing to both come out?

Emily: The point about the semantic drift is that that makes sense for performing/performance, but that we won't be so happy for other non-productive nominalization examples.

oe: Would be helpful to go through a couple dozen examples and conceptually try out both universes.

Francis: I can list a few thousand and randomly sample if you like.

Francis: Before that though, completely independently if we decompose or not, what is the current plan for argument roles for 's or of relations? Leave them unspecified?

Dan: All you know is that it's a poss_rel.

Francis: So consumers of this won't get that link.

oe: So far we haven't' been able to find grammatical contrasts that would force that.

Francis: It's a principled and probably defensible decision, but not necessarily one that people would expect.

Woodley: If there are no argument positions then you know it's not filling one.

Francis: But poss_rel is still very underspecified.

[ Emily briefly gone ]

Francis: Why does the current grammar ambiguate by adjunct from passive by. To see the difference to the possessive arguments of relational nouns.

oe: The story remains the same as in the Japanese relative clause discussion. There's no reason for the grammar to do it, but the tree bankers should.

Francis: In practice in DELPH-IN, if we don't ambiguate in the grammars, it doesn't get ambiguated. I don't see the point of saying that performance is the same as perform, and that it has the same argument structure, if you don't tell me anything about that argument structure.

oe: In some cases we can.

Francis: I'm happy to get what I can get, I'd like to keep on the table the desire for more.

Dan: In a slightly better world you would be happy if whenever you see poss_rel, there was a tool that would pop and say tell me if that poss_rel corresponds to argument or adjunct.

Francis: Looking the other way: if I could be an argument, I want to know if it's in an argument.

Woodley: The notion that poss_rel is underspecified between adjunct and argument is incompatible with the idea that underspecification is implemented by subsumption of MRSs. MRS with poss_rel doesn't subsume MRS without it.

Emily/oe: Same as the hope of using RMRS to underspecify argument v. adjunct uses of PPs (more generally)

oe: Valid point, but doesn't invalidate the point about design principle that we use for deciding the division of labor. We have to invoke meaning postulates to solve that.

Francis: Worried that if we're trying to make our system more useable, if there's a very ununiform distribution of what's ambiguated and what isn't, it'll be unsettling.

Emily/Woodley: So if we're doing it for poss_rel, then why not do the same for the other argument/adjunct ambiguities involving PPs?

oe: Jump in the water in the Pacific NW. Only the first of those can be the directional.

Woodley: I don't see which ambiguity is the problem.

oe: This is the reason I think why the grammar ambiguates because there are cases where we know it has to be one or the other. Adjunct is specialized to stative and complement to directional.

Dan: I sent a letter to Paris either Paris is the name of the recipient (ARG3 of send) or the destination then it is an adjunct.

Woodley: So why not have an ep for that and just say 'meaning postulate' to get to those two possible readings.

oe: Three-place send always takes a handle-valued ARG3.

Emily: What about I sent Kim a letter. --- three x arguments.

Dan: Don't want to have users see a difference between I gave a letter to Kim and I gave Kim a letter. send seems a lot like give, but it isn't just like give, it's also like put. It can't be like both of them unless I ambiguate it.

Woodley: How about I wrote a letter to Kim

Dan/oe: Doesn't have the resultative, while send does. It ends up with Kim having the letter.

Woodley: I wrote a letter --- predicate symbol is the same, though the arity varies.

oe: Three readings, two relevant here: three place write, or two place write with unexpressed recipient and an adjunct.

Woodley: Seems like an ambiguity that should show up a lot, why bother putting them both in rather than having some to_rel meant to be interpreted as underspecified between these two.

oe: I wrote a letter to Kim to Paris in the sense of on the train to Paris.

All: star

oe: So what's the to adjunct then?

Dan: A perfectly fine modifier of nouns, harder to see just now for verbs.

Francis: I officially withdraw this topic.

Dan: You're asking us to peal back 20 years of our work and 50 years of syntax before that.

Woodley: What if someone brings a generator input that has an actually linked up ARG1 or ARG2.

oe: They get the by variant for the the ARG1.

Emily: and of for ARG2

Dan: I already assert that performance of the symphony takes an of complement, which links to an argument (currently ARG1, but rather ARG2 in the new proposal).

Francis: So what does the annotator do?

Dan/Emily: Choose between ARG2 (marker of) and adjunct, if both are available. (Sometimes grammar disambiguates with pronouns in accusative case).

Francis: But for possessives?

Dan: Just underspecified. But could ambiguate to possessive or ARG1, but grammar would never disambiguate. Compare of, where the grammar almost never disambiguates, so why not do it for the possessive for the deverbal nouns as well? Could say that the ARG1 marking one is only a possessive determiner for deverbal nouns.

Francis: Then I think you're giving me everything I want!

oe: I think Dan's already giving you the reasonable requests.

Francis: I only made reasonable requests!