In the ERG meaning representations (and in logic-based semantics, more generally), how to deal with variation in the number and types of semantic arguments?
Survey what other broad-coverage initiatives do, e.g. EngValLex, PropBank, FrameNet, OntoNotes, AMR, etc.
Fillmore 1986 Pragmatically controlled zero anaphora
Fillmore 2007 Valency Issues in FrameNet
Johnson and Fillmore 2000 The FrameNet tagset for frame-semantic and syntactic coding of predicate-argument structure
- Grimshaw -- Dan to fill in which paper this is
(Kilgariff 1997 I don't believe in word senses)
Flickinger et al 2005 Sem-I Rational MT
Asher 2011 Lexical Meaning in Context: A Web of Words (Recommended by Alex L. as an example of flexible types in lexical semantics)
Resources to Review
- ERG Lexicon
- Prague valency lexicon
- Levin, B. (1993) English Verb Classes and Alternations: A Preliminary Investigation, University of Chicago Press, Chicago, IL.
Other Wiki Pages
Argument still there semantically even when dropped: eat
Argument gone semantically if not satisfied syntactically: kick
Argument can be NP or CP/S: report
Accepts clausal complement only: contend ... but FrameNet provides an example with and X contends the latter, so we need a better example here.
FrameNet lists two frames for deny, Affirm_or_deny and Prevent_or_allow_possession. On examining the annotated data, we found the following valence frames, and aligned them with the disavow and decline like senses:
With example arguments:
that they cheated
the petitioners; the request; Kim a kiss
a kiss to Kim
Note1: In the NP, NP frame, either of the NPs can go missing, but not both at once: *Kim denied. Note2: FrameNet has one instance of something like deny a request classified with the frame that otherwise has the disavow senses. We think this is probably an annotation error.
deny_v_1(e,x,h) deny_v_1(e,x,p) deny_v_3(e,x,x) deny_v_2(e,x,x,x) deny_v_2(e,x,x,x)
Would like to know how each of these options (as well as always getting the same predicate symbol) look to a logically-inclined semanticist. Which is preferable? Is there yet something different that would be better?
Observations: In this case at least, the analysis is consistent with the approach to having the Sem-I do the refining, since the information needed to distinguish between the predicate symbols is all apparent in the ERS. Exception: Since both ARG2 and ARG3 are syntactically optional in deny_v_2, how would we know which way to specialize NP denies NP? Alternatively, how do we know when to put in ARG3, especially given that in NP denies NP, the second NP could be either ARG2 or ARG3?
Kim denied the rumor.
Kim denied the request.
Kim denied the petitioners.
This entails two separate lexical entries, which could then have different PRED values.
Notes from discussion on 20-Apr
(Emily, Dan, Cleo)
Kim denied to Sandy that Pat arrived.
%Kim denied the rumor to Sandy. -- corpus study needed
*Kim denied to Sandy cheating.
*Kim denied cheating to Sandy.
Interesting that the addressee is available for CP complement but not the VP one. Necessarily unexpressed in VP frame, but still there in the semantics?
?He denied my having received any bribes.
*He denied me having received any bribes.
Are gerunds similar in semantic type to CPs and nouns like rumor?
I resent that he sang that song.
I resent his singing of that song.
Is rumor something like a proposition in its semantics? But:
I deny most rumors that circulate about me. -- clearly quantified
Kim denied every annoying rumor. -- what is annoying modifying? Is it the content that's annoying? The frequency of the repetition?
I deny every false rumor. -- clearly the content is false.
The whispered rumors of his affairs
The frequent rumors of his affairs
The word rumor requires something above the content, that the story had been passed from person to person --- but what deny targets is the content.
It was frequently rumored that Kim would quit. -- Suggests that the frequent rumors has frequent modifying the event of rumoring.
Are there predicates that just want the propositional content in their complements and can't take individuals? hope, insist, ... Dan has created a list.
Deny seems to want to take as its complement a proposition that has been asserted to be true --- some nouns carry this for the propositions they package, some don't. So one can deny a rumor, a claim, a conclusion but not a hypothesis, theory, hope or suspicion. The nouns that deny can take denote a communicative event with a truth value assigned. Deny asserts the opposite of that truth value. belief can't be a complement because it's an attitude about truth but not communicative. Deficient in both ways: wager, likelihood.
Deny with a CP complement presupposes that someone believes that the CP is true.
Deny a request similarly involves one actor pushing for one state of affairs (that they request) and the other pushing back the opposite (denying the request). Both rumor and request case involve flipping like this.
Request case also has the communicative act property: Kim denied Sandy a kiss suggests that Kim knew Sandy wanted a kiss.
The supervisor denied the employee a raise in the case where the person who put in the request was HR (and the employee never knew).
The guard denied the prisoner his 10 minutes in the sun. Allowed every day, standing request/expectation. Guard prevents it happening one day; deny is still good here even if prisoner said nothing.
In basketball --- player A goes up to make a shot which player B blocks at the basket: Denied! Cleo was on her way to Oslo. She had even gotten on the plane. And then: Denied! That denied is passive not past tense.
I denied him his rights. I denied her the privilege. Again: Standing expectations?
@The deputy president said for those such as MaMbeki, the struggle against apartheid was inherently a struggle to reclaim their birthright as a people, as successive racist regimes denied them their citizenship, dignity and identity. (source) --- what's being denied is the ability to claim or benefit from citizenship, dignity and identity.
NdT denied the solar system its 9th planet: What's being denied is the community's shared belief in 9 planets, but also the thing we enjoyed which was having 9 planets.
Mom denied my sister her green peas -- not okay if said sister hates peas.
Attempting to get just one sense for request/rumor cases, with all differences following from unsettled v. settled:
prop content p through coercion from NP
prop content p directly or through package
for some agent it's a good thing that p (become) +
communicative event asserts p -> +
rights? attempts? states?
flips potential for + to -
flips + to - by asserting
request type takes a much broader range of nouns, but can get that from "p (become) +" being good for agent A.
The only thing that doesn't fall out is the lack of communicative event on the unsettled/request side.
Antonym: grant, seems to work in both cases. I grant you that belief --> only works in the unsettled case, which is predicted. I grant you the ability to hold that belief.
More ex from FrameNet: deny any wrong-doing deny any mistakes ... interpreted as charges/claims that there had been wrong-doing or mistakes. Needs to be coerced into proposition with presupposed communicative event.
In settled case denied the rumor to everyone : everyone can include people who didn't assert the rumor in the first place; addressee of second assertion, not speaker of first. In unsettled case denied x y or denied y to x : y has to be the person for whom "p (become) +" is beneficial.
Would we see the same thing for grant?
Notes from WG on 24-Apr
(Emily, Dan, Cleo, Dag, Stephan, Adam)
oe, introducing the topic:
Problem statement: Variation in arity, variation in type of arguments; desire on our part to understand better the formal challenges for compatibility to logic on the one side and maybe forming a reading group to do a survey of what others who have worked in this space in broad coverage (annotating running text) have done.
Null instantiation: Does I ate mean I ate something, but I kick doesn't have the same entailment: can make a kicking movement without kicking anything. Deny shows type variation in its syntactic valency.
Side discussion of deny plus V-ing. Kim denied singing the song --- how to test whether that's a (verbal) gerund v. a participial VP. (Cf. Kim relies on singing that song the same way everytime.) Unfortunately, verbal gerunds don't always give clear indication that they've been turned into an NP, just an optional determiner.
Dag: Can you have intransitive deny?
Dag: Norwegian deny does -- but with anaphoric interpretation of the argument.
Emily: Four kinds of dropped arguments? DNI (like notice), INI (like eat), no argument there at all (kick), some syntax going on to suppress the argument that also dictates its interpretation (CNI; dropped by-PP in passive, dropped subject in imperatives...)
Dag: And reciprocals, like They kissed?
Emily: I don't know which type that is ... lexically-specific, right?
oe: Back to variation in valence of deny.
Cleo: We made an interesting observation about content bearing nouns, you can deny a rumor or a conclusion, but not an hypothesis.
Dan: Hypothesis by its lexical sense doesn't give you something with a putative truth status to deny.
Dag: Can you do deny whether?
Dan: No, and probably for the same reason.
Cleo: And it has to give you a communicative event --- can't deny a belief.
Dan: We were interested in that exploration because Cleo was teaching Emily and me about the tools in the toolkit of a formal semanticist ... what you can ask about: polarity, factivity...
Emily: The connection I see here to the main problem statement is: Is deny in fact polymorphic, or does it always take a proposition as its complement, but it knows to dig for it when the complement is a noun like rumor?
Dan: And we're curious how to build a compositional semantics that allows rumor to expose different aspects of its semantics for different things in the sentence to grab onto (package, content) in one and the same use, as in the whispered rumors of his affairs.
Dag: Co-predications: The false, whispered rumors
Dan: So it has to be just one word, can't do a lexical alternation.
Dag: As with Heavy, informative books
Adam: Pustejvosky talks about a lot of this stuff, right?
Emily: One thing that I found very interesting on Friday was that there seemed to be a space between the string and the manipulations that Cleo wanted to do. So that brings me back to the question of what would someone who wants to do a theoretical analysis of deny like what we were doing want to come out of the grammar? And how much munging is allowed while still calling ourselves compositional?
Cleo: Maybe it makes sense to start with these hard cases, methodologically.
oe: Also brings up the dichotomy between individuals and propositions, which maybe aren't in as sharp a dichotomy as logic would lead us to expect.
Dan: We might get a prediction about when we get -ing v. infinitival. Can we prdict this pattern from something?
continued making that claim continued to make that claim *continued that we made that claim continued the rumor denied making that claim *denied to make that claim denied that we made that claim denied the rumor
[Side conversation about why premise, not proposition is allowed as complement of deny]
Cleo: Even with just a plain that complement, you get the implication that someone else had held the position.
Dan: Where we have this range of slightly different syntactic packaging (nouns, gerunds, controlled complements, full clauses) it would be nice if we could predic that from the semantics.
oe: At least one necessary condition we considered was the noun taking a propositional argument.
Cleo: And it has to be packaged as a communicative event.
Emily/Dag: But that doens't answer Dan's point.
Dag: Clear contrasts: I remembered to buy the fish / I remembered buying the fish ; I forgot to vote / I forgot that I had voted.
oe: So these ones definitely need different predicate symbols, because there might not be enough info in the semantics to say we can recover it from an underspecified one. Like what we do with paper, allowing ourselves not to ambiguate when the grammar doens't tell us. For to v. -ing there's nothing left, right?
Dan: There's the aspect...
Dag: What about He forgot to be playing the piano when his teacher his came in.?
Cleo: I think what's happening with infinitivals: There's a class of verbs for which it is forward-shifted. You have to have understand the infinitival clause as having temporal reference forward-shifted from the time of the main embedding predicate, where -ing doesn't do that. In forgot to be playing the piano when the teacher arrived, there was a preceding expectation/obligation.
Dag: My worry was more that distinguishing the infinitival and the -ing in the semantics...
Emily: But in I remembered buying the fish couldn't that be a gerund?
Dan: Sure, most of those verbs can... keep, stop, began, continue...
Emily: Those are all very aspect-like main verbs.
Dan: This is a cline in every direction.
Dag: It kept raining
Dan: Maybe we can use raising to distinguish between gerunds & V-ing?
Emily: And control, too, surely. For I deny cheating
Emily: Well, it's definitely control, right? I deny cheating on that paper. I saw you cheat! You're lying. I meant I deny Dag's cheating on that paper.
Dan: And in I deny any cheating?
Emily/Dan: More plausible deniability there...
Dan: It would be nicer if we could make a sharp distinction between raising and control here.
Emily: You want things that will show up in the string, don't you?
Dan: I want to be able to avoid unnecessary ambiguation, especially when the semantics comes out very much the same.
Dag: What other verbs do this?
Dan: Verbs of perception? hear, see
I heard him sing I heard him singing
Dag: But with the infinitival marker. Will we see other sharp distinctions like with forget and remember.
Dan: [Searches his lexicon] bother (NPI), dread, fear, hate, like, love, start, continue, *stop (ing only), began, stand (NPI), try
He didn't bother to come to the party. He didn't bother coming to the party.
Emily/oe/Dag: Same meaning, both control.
Emily: There is a contrast in entailments for try
I tried eating the cake I tried to eat the cake
Cleo: Difference in how long the state of affairs hold:
I fear going to department parties I fear to go to department parties
Dan: That's older English anyway...
Cleo/Dan: Difference in how they relate to repeated events.
I hate to go to parties I hate going to parties I would hate to go parties I would hate going to parties ??I would hate to go to that party ??I hated to go to that party ??I loved to go to that party
[Scribe's notes: This isn't quite the right set of strings; didn't keep up.]
*I hated to do this to you, but I had to. I hated doing this to you, but I had to. I hate to do this to you, but I have to.
Adam: Difference between particular event and kind of event? (Cf. Chierchia.)
Dan: He hated to sing Christmas songs at Easter --- it happens multiple times and he hates it each time.
Adam: He would hate to give an after dinner speech --- no specific event. Can refer to kinds without a particular instance of this kind.
oe: So there is something that appears reasonably robust.
Dan: A property here that distinguishes ing form from the to form. Distinguishes kind from particular event.
Dan/oe: [Tongues in cheek] it's the same as mass/count
Cleo: Trying to figure out a connection with forget. In both forget and remember you get the implication that it actually happened with the -ing form. Could it be the ing form that is used to encode that something actually happened, when something in the verb requires it to happen, and that's what the infinitival resists?
Dan: Not just the infinitival --- continue to do --- maybe it fails to provide something that is needed in some cases.
Dag: Is there a contrast for continue to V and continue V-ing?
Cleo: But when there is a difference, maybe the -ing migrates to the cases where the thing happens.
oe/Emily: What was the contrast with remember again?
Cleo: Remember that is factive, remember to is implicative (implies complement except it flips under negation), and remember -ing is also factive.
Emily: And remember to and remember -ing are control.
Dan: ... for to, definitely, for -ing, maybe... Then we're back to the gerund problem.
Emily: *I remembered for Toby to bring a key. but I remembered Toby bringing a key.
Dan: That's a gerund: Toby brining a key is what I remember It's one constituent.
oe: Does that constituency test say anything about I remember bringing a key?
Dan: Still open --- could be a simple control structure or could be a gerund.
Dan: keep doesn't take an NP complement in that same sense, so it's a nice clear case of control into VP-ing complement.
Cleo/Dan: Why does it have to be I remember it raining, and not I remember raining.?
Dan: Would like a sharper test for necessarily controlled -ing-form VP complement v. gerund. keep being a nice clear case.
oe: So let's find other cases of expletive-taking predicates as gerunds, to see if we can side-step what might be an idiosyncractic property of raining blocked by rain.
Cleo: So gerunds start verbal and become nominal, right?
Dan: Yes, because they can show up after prepositions.
Cleo: So you can have: frequent raining?
Dan: Not great...
Dan: Maybe we're too far down in a corner here with worrying about when I have to posit the legit but infrequent V-ing controlled complements.
Cleo: Can I try out an argument that they're really not nominal? Because you get this implication that this thing happened --- I forgot the rain in Oslo, I forgot the frequent rains in Oslo nominal arguments need a marker of definiteness. I forgot bringing the key. Doesn't have a marker of definiteness. If it were nominal you'd expect a determiner that is by default definite.
Cleo: If you have an indefinite, it would be a specifc one.
Emily: So remember wants a specific indefinite, if the argument is indefinite?
Cleo: Yes (including partitives). The argument is that the V-ing doesn't take any marking of definiteness, so they aren't nominal in this use (with remember/forget).
Emily: So the arugment hinges on specific properties of forget and remember, that they want definite or specific indefinite.
Cleo: Yes, or every...
Dan: I remember no singing of pop songs in this cathedral. What does it means in terms of what actually happened? Still trying to understand what the constraint is on the NP meaning.
Cleo: What I thought was the descriptive content of the NP was presupposed to be instantiated... there were signings, there were rains, and then you quantify over this domain. But you're right that in present tense... there has to be something in the discourse, where someone has claimed there was some.
Dan: I see the direction you're taking and it's tempting, looking for constraints on the denotation of the NP, and whether it's definite or indefinite could be a differentiating property... could at least be a place to look for a piece of evidence for controlled complement v. NP gerund.
Dan: That's convoluted; would take a week of calisthetics to be able to try to take that on.
Dag: Hold my beer.
Dag: Even more convoluted arguments: Controlled pro is only supposed to give you de se readings, where the attitude holder is aware of the identity. Scenario: A guy watching a sports show on TV, and he doesn't know it, but it's footage of himself winning a race and he says, "Oh I remember that," but he's not aware that it's him. Can you report: John remembers winning the gold medal.?
Emily: Two weeks.
Dag: John reads a paper and really likes it, not aware that he's the author and says, "This author is a genius!" Can you report John claims to be a genius? Clearly no. Surprisingly clear judgment.
Adam: Connected to Madame Tussaud sentences. Ringo Starr goes to the wax museum and sees that the statue of himself is dusty. Can you say: Ringo Starr dusted himself? Long form v. short form reflexives differs on this in other languages.
Dag: Different to control examples, because there is only one referent in the world (and lack of knowledge about the identity).
Emily: So I'm still wondering how to take this rich discussion and use it to inform the question of whether we have to ambiguate in the grammar's output.
Dan: John bought the paper --- can still leave it underspecified. But what we've been told by logicians is that we have to change the predicate symbol as soon as the number or type of arguments change. But can we hand off the underspecified symbol anyway and let the mapping to target representations do the specialization? My problem? Seems like someone else's, but also seems like part of the grammar.
oe: But will an annotator require context to decide? I expect that will often be the case.
Dan: As they would for the paper case. The author of the text probably weren't vague --- they had a clear opinion of what was intended.
Dag: Is -ing v. to an interesting case to look at for that?
oe: Because it's not going to be black and white. Can't always underspecify --- the contrast with forget should be projected and currently isn't.
Emily: That's not a problem for the annotator.
oe: Right -- it's a problem for making available all information that comes from the grammar.
Dag: Has to be stored either in the governing predicate or in the form of the complement.
Cleo: That's where the other story becomes attractive. You get forget and what kind of complement it takes.
Emily: But we don't want to stop there. We want to get all we can out of the syntax.
Cleo: This looks like a case where I like the two-step position. It doesn't force me to do something when I shouldn't. What really matters are the properties of form, and that's what I'm going to build my inferences on.
Cleo: Lexicon says: forget+that factive, etc. And then I'll know what to do. And depending on what kind of task you have, you might want just the inferences about the complement.
oe: If you're happy to do that after parsing, then that forces the parser to include that contrast in its interface representation --- a very overtly morphosyntactic ... something we would find appalling...
Dag: Like the DRSs that Matthew and I make.
Cleo: I'm happy to play the other game too. Just trying it out in my head.
oe: The other game has the benefit and challenge of potentially forcing some separation of powers.
Dan: There is a theory of that in LFG. f-structure imposes some theory of what of syntax is available for semantics.
Cleo: It's a constraint problem; trying to satisfy multiple constraints.
Dan: With two potentially unrelated engines. One composing the interface representation, and one taking that do something with meaning constructors. Hard to have an argument with that. Just defining the rules of the game in a different way.
Cleo: If I were to teach semantics to a class, I wouldn't adopt this perspective.
Adam: For me, one argument for having a single valency frame is coordination (Kim remembered the appointment and that they had to be on time.). Also maybe co-predication.
oe: Kim remembered that it rained for an hour and the mess that it left outside the house.
Dag: How does it apply to ing and to?
Adam: I've been trying to construct an example...
Dan: I remembered buying the milk and to pick up some milk. ... so maybe that remember isn't the same remember?
Cleo: But: I remembered buying the book and that Mary was there with me? What's shared here is the factive use.
Dan: So doesn't that suggest that the ing is a proposition and not a property?
Dag: There are too many moving parts here, hard to tell. Don't have to say that a VP denotes a property.
Dan: Is that right in general for control structures?
Dag: There's a big debate about that.
Dag: And as Adam said, it's tempting to make some of these denote events.
Emily's notes on Levin 1993
She explicitly only looks at NP & PP arguments, not clausal complements, but does include diathesis alternations that change the arity.
The project is very much framed in terms of understanding the meaning components that predict the possibility of the various alternations (plus the meaning of derived nominals and adjectives) and she argues (briefly) for decomposing predicates, which seems to fly past our notion of predicate symbols. She also talks about alternations (at least in the intro) that involve predictable changes in verb meaning --- e.g. verbs like whistle can be used as verbs of motion, but not bark (I'm assuming that the X's way construction is something different still; she doesn't mention it), and concludes that in something like The train whistled past the station that sense of whistle is better treated as a verb of motion (with an accompanying sound) than as a verb of sound emission (with an accompanying motion). All this to say that she sees these different uses as the same verb (at some level of abstraction), but sees them as having different meanings, presumably differentiated by specific meaning components.
In sum, I don't think the question "Does Levin think that predicate symbols should have fixed arity?" is usefully answered in that book, but it does provide interesting further context for our discussion.