[developers] Difference between neg_rel/modifiers and modals

Ann Copestake aac10 at cl.cam.ac.uk
Sat May 20 14:06:02 CEST 2017


Sorry I don't have time to properly engage with this.

I just wanted to say that, at a DMRS level, the problem is just about 
making sure that the compositionality rules can give you the right 
readings without over-constraining.  What the DMRS post-slash 
constraints give you is information about the scopal relationships - you 
can interpret /= with events, as we usually do in ERS, but you can also 
interpret it with no events at all, as in non-Davidsonian accounts.

so:

not ---------> talk <----------- for-three-hours

      ARG1/H             ARG1/=

just guarantees that `talk' and `for three hours' are under the scope of 
`not' and you could interpret the = in different ways depending on 
whether or not you're using events.  In this view

for-three-hours ---------> not  ----------------> talk

                            ARG1/=                ARG1/H

is a perfectly good DMRS as long as one isn't using an event-based 
representation to show what the = means, because that forces one into 
having `not events'.

Conventional notation is not very helpful here, because it forces one to 
write something that looks like there's a scopal relationship if one 
doesn't use events, but that's actually a notational issue, not a 
genuine semantic one.  For instance, I can perfectly well express what 
the two alternatives mean with a temporal logic without events.

Anyway, this is just a brief indication of what the situation is from my 
own current perspective.  I can try and help work things out from an 
event-based point of view, but I can't say my heart is in it!

All best,

Ann

On 19/05/2017 21:16, Guy Emerson wrote:
> So with the modals and attitude, we're saying that there's a state, 
> which bears some relation to the embedded event, and both the state 
> and embedded event can be modified.
>
> I think this kind of situation can also happen with habituals, without 
> introducing an extra predicate:
>
> "I didn't used to brush my teeth for long enough, but for the past 
> year, I've been brushing for two minutes"
>
> In terms of the denotation, we have many brushing events lasting two 
> minutes, and a habitual state lasting a year.  A more minimal example:
>
> "Kim brushed for two minutes for a year"
>
> Some word orders seem weird (but maybe okay with the right intonation):
>
> ?"Kim brushed for a year for two minutes"
> "For a year, Kim brushed for two minutes"
> ?"For two minutes, Kim brushed for a year"
>
> So if want to introduce an extra event variable for modals, it seems 
> to me that we might want do the same for habituals, too, even though 
> they can be unmarked in English.
>
>
> For (2a) and (2b), I think we can capture the difference without 
> necessarily referring to the neg "event", by using label sharing to 
> control whether the adverbial scopes above or below negation (my 
> "option 2" below).  So for (2b), the claim would be that fronted 
> adverbials have to take high scope, i.e. share a label with the LTOP 
> rather than the INDEX.  In MRS, this would be immediate, since the 
> INDEX's label is no longer available.  In DMRS, it would have to be a 
> syntactic constraint (you know, maybe this reading is fine in Turkish).
>
> In this analysis, the negation has to act on an expression with a free 
> variable (the event), rather than a proposition, so that the event is 
> still available for the adverbial when it takes high scope.  Negation 
> is still easy to define, but rather than inverting truth values, it's 
> inverting truth-conditional functions.  (In type-theoretic terms, it's 
> of the form <<e,t>,<e,t>>.)
>
> That is, rather than not(speak(e)), we have not(speak)(e), so that we 
> can write not(speak)(e)&for-a-long-time(e), which would contrast with 
> the other reading not(speak&for-a-long-time)(e).  But then I don't 
> think it really matters whether "not" or "speak" is introducing this 
> "e".  If we say that "not" is introducing an event, it's effectively 
> just wrapping the event from "speak", anyway.
>
> If we're happy to do that, then we could perhaps extend this approach 
> to modals - can(close) takes a truth-conditional function that's true 
> of closing events, and returns a truth-conditional function that's 
> true of able-to-close states.  The embedded verb's event never gets 
> quantified, which is perhaps reasonable - "Kim can close the window" 
> can be true even if Kim never closes the window.
>
> This would make negation and modals look formally very similar, even 
> though the modal states look very different from the embedded verb events.
>
> So maybe these events aren't so problematic after all?  And I'm sure 
> there's a literature on this that I should read.
>
>
> 2017-05-19 1:45 GMT-07:00 Ann Copestake <aac10 at cl.cam.ac.uk 
> <mailto:aac10 at cl.cam.ac.uk>>:
>
>     my intuition is that events (eventualities) with modals make some
>     sense in an object language:
>
>     Kim can close the window.
>
>     can(e,k,close(e',k,w))
>
>     e refers to the state of Kim having the ability, much as in
>
>     Kim believes Sandy slept.
>
>     believe(e,k,sleep(e',s))
>
>     we can talk about the state of Kim having the belief.
>
>     Kim could close the window for an hour.
>
>     has a reading where it's Kim's ability that lasts for an hour
>     (e.g., follow up with "and then was too weak") - that seems OK in
>     terms of eventuality modification.
>
>     Originally event semantics didn't include states and people argued
>     both ways, and off the top of my head, I can't remember who ... 
>     Still, states make a certain amount of sense in terms of a
>     collection of properties or potentialities associated with a
>     spatio-temporal location, in a way that the not "event" and the
>     probably "event" don't.  I think one might find discussion of why
>     not events don't make sense in some of the situation semantics
>     literature.
>
>     Decomposed events have been proposed in a number of contexts where
>     the adverbial seems to refer to a preparatory state or whatever. 
>     Higginbotham and various Generative Lexicon people (Pustejovsky et
>     al), for instance.  e.g.,
>
>     (13) Mary came in an hour for an hour.
>
>     from a paper that talks about the event decomposition idea (which
>     I just found with an extremely cursory search, so don't take it as
>     a proper citation)
>     https://www.ucl.ac.uk/pals/research/linguistics/publications/wpl/96papers/evans
>     <https://www.ucl.ac.uk/pals/research/linguistics/publications/wpl/96papers/evans>
>
>     So the idea that one can say that there's a preparatory state of
>     not talking in:
>
>     (2a) Kim didn't speak for a long time.
>
>     is perhaps sort of plausible.  i.e., one could claim that the
>     single event allows an underspecification of the two readings.
>     But then
>
>     (2b) For a long time, Kim didn't speak. ;;; Kim was silent for a
>     long time
>
>     is problematic in that it only has the one reading.  One could
>     stipulate that, of course, but it's not pretty.
>
>     Maybe I'm wrong to be so worried and someone has seriously
>     proposed not events.  Ask Alex? What one's looking for (in terms
>     of the object language) is a literature where the denotation is
>     discussed - not simply an argument from ambiguity / readings.
>
>     Cheers,
>
>     Ann
>
>
>
>     On 19/05/2017 04:17, Emily M. Bender wrote:
>>     Right---I'm trying to understand why it is that we give different
>>     representations
>>     for not v. other modal operators wrt which event variable is
>>     exposed, with the
>>     longer range goal of getting to tests that could in principle be
>>     applied in other languages
>>     too, so we could find out if the representation we pick for
>>     sentential negation
>>     works across languages.
>>
>>     Collecting the data that has come up so far in this thread:
>>
>>     (1a) We could unexpectedly close the window. ;;;
>>      could(unexpectedly(close)) / unexpectedly(could(close))
>>     (1b) We did not unexpectedly close the window. ;;;
>>     not(unexpectedly(close))
>>     (1c) Unexpectedly we did not close the window. ;;;
>>     unexpectedly(not(close))
>>
>>     [Aside: The reason I was asking about extraction is that we do
>>     have a construction that
>>     allows an adverb to attach low in the semantics but appear at the
>>     left edge of the clause.
>>     That would predict not(unexpectedly(close)) for (1c), which I
>>     think isn't available.]
>>
>>     (2a) Kim didn't speak for a long time. ;;; Kim spoke for only a
>>     short time / Kim was silent for a long time
>>     (2b) For a long time, Kim didn't speak. ;;; Kim was silent for a
>>     long time
>>
>>     If (1a) is really ambiguous, is that meant to be an argument that
>>     'could' has its
>>     own event that can be modified?  Why is it less problematic for a
>>     modal operator like
>>     'could' to introduce an event (in terms of the underlying
>>     semantics) than something
>>     like 'probably' or 'not'?  Do the readings of (1b) and (1c)
>>     correspond to the two readings of (1a)?
>>
>>     Just now it seems to me that the two readings of (1a) and the
>>     pair (1b)/(1c) aren't really
>>     relevant to the question of which INDEX is propagated, because in
>>     any case the ARG1 of
>>     unexpectedly or not is handle-valued.  But, we'd consider 'for a
>>     long time' to be a non-scopal
>>     modifier in (2), right?   So what do we want 'for' to take as its
>>     ARG in (2b)/the second
>>     reading of (2a)?
>>
>>     Emily
>>
>>
>>
>>
>>     On Thu, May 18, 2017 at 6:55 PM, Guy Emerson <gete2 at cam.ac.uk
>>     <mailto:gete2 at cam.ac.uk>> wrote:
>>
>>         I think Emily's goal was to figure out what representation we
>>         should use, and whether we need to have different
>>         representations cross-linguistically. (Emily, is that a fair
>>         summary?)  I can see that a negated event could be
>>         problematic, but I was going off the ERG semantics, where
>>         neg_rel has two arguments, so it looks like we do have
>>         not(e,P).  In DMRS, we can avoid saying whether there is an
>>         event, but it's there in the MRS.
>>
>>
>>
>>         2017-05-18 8:01 GMT-07:00 Ann Copestake <aac10 at cl.cam.ac.uk
>>         <mailto:aac10 at cl.cam.ac.uk>>:
>>
>>             I do think it's really important to be clear what the
>>             goals are.  Are you trying to figure out what the
>>             representation should be in terms of the underlying
>>             semantics?  Because then talking about negation events
>>             could well be problematic.  There are moves one can make
>>             which might work - e.g., situations in Barwise and Perry
>>             terms (but then that doesn't necessarily fit with other
>>             things we're doing) - but one can't simply write e.g.,
>>             not(e,P) and assume it's meaningful.  I mean, maybe you
>>             want e to refer to the period of time when not(P) holds. 
>>             But I guess you can see that this is not something that
>>             is obviously OK.
>>
>>             Alternatively, you're essentially leaving the object
>>             language up to someone else and trying to come up with a
>>             representation which captures the right things about the
>>             syntax/semantics interface.  But I still think you have
>>             to know something about plausible target object languages.
>>
>>             All best,
>>
>>             Ann
>>
>>
>>             On 17/05/17 21:14, Guy Emerson wrote:
>>>             To bring this back to Emily's question, I can think of
>>>             two ways that we might represent the "silent for a long
>>>             time" reading:
>>>
>>>             Option 1. "for a long time" takes the neg_rel's variable
>>>             as an argument. This could be constructed
>>>             compositionally using the negation-as-a-modal analysis
>>>             that Emily mentioned.  This would then allow neg_rel to
>>>             have a consistent semantics in the Grammar Matrix.
>>>
>>>             On the downside, if we push the INDEX up to the neg_rel,
>>>             we can't get hold of _speak_v_rel any more - which we
>>>             need if we're going to model adverbs attaching after
>>>             negation but scoping underneath negation.  With DMRS
>>>             composition, we can construct it compositionally even if
>>>             we stick with the scopal modifier approach (so the INDEX
>>>             is still "speak"), and then connect an ARG/EQ link to
>>>             the LTOP.  This would, however, mean relaxing the
>>>             constraints in the proposed DMRS algebra, since we have
>>>             an /EQ link selecting the LTOP, not the INDEX.
>>>
>>>             Option 2. "for a long time" shares a label with the
>>>             neg_rel, but still takes _speak_v_rel as an argument. 
>>>             So then "for a long time" is outside the scope of
>>>             negation.  To construct this compositionally, we want
>>>             _speak_v_rel to be the INDEX (for both MRS and DMRS
>>>             composition).
>>>
>>>             If we take this approach, then we can treat modals as
>>>             scopal modifiers and still get two readings.  So this
>>>             doesn't directly answer Emily's question, because now
>>>             there are two different ways of getting two readings.
>>>             But it would at least suggest that we can treat modals
>>>             as scopal modifiers, which would allow a more consistent
>>>             semantics of negation in the Grammar Matrix.
>>>
>>>
>>>
>>>             That's the main thing I wanted to say - but Re: Robin Hood:
>>>
>>>             I've found Ivan Sag's discussion of the jailing Robin
>>>             Hood examples
>>>             (https://www.academia.edu/2798317/Adjunct_scope
>>>             <https://www.academia.edu/2798317/Adjunct_scope>),
>>>             apparently discussed by Dowty (1979).  I can see the
>>>             relevance, in that "for three years" could refer to the
>>>             time in jail, or the time spent putting him in jail. But
>>>             I'm not convinced by the argument that we should
>>>             decompose this as a causative - otherwise, the verb
>>>             "sentence" also seems like it could be decomposed into
>>>             something like cause(be-in-jail), but it doesn't pattern
>>>             like "jail":
>>>
>>>             The Sheriff of Nottingham jailed Robin Hood for three years.
>>>             *The Sheriff of Nottingham jailed Robin Hood to three years.
>>>             The Sheriff of Nottingham sentenced Robin Hood for three
>>>             years. (repeated jailing reading)
>>>             The Sheriff of Nottingham sentenced Robin Hood to three
>>>             years. (single jailing reading)
>>>
>>>             In any case, we can get different readings for verbs
>>>             without an obvious lexical decomposition:
>>>
>>>             I ate meat for a year (but then became vegetarian)
>>>             I ate meat for an hour (and then I was very full)
>>>
>>>             Bouma&Malouf&Sag also discuss "open again", but
>>>             similarly, "Kim bought X and sold it again" has a
>>>             reading where this is the first time Kim sold it.  And
>>>             explicitly representing that reading by decomposing
>>>             "sell" would require something like cause(be-sold). 
>>>             This seems dubious to me. I'm much more tempted to say
>>>             that "again" has a fuzzier meaning than Dowty assumes.
>>>
>>>             I couldn't find any examples which convinced me that
>>>             there's an interaction with the morphosyntax, so I feel
>>>             like this is all something that we can safely leave out
>>>             of the MRS.
>>>
>>>
>>>             2017-05-17 3:57 GMT-07:00 Ann Copestake
>>>             <aac10 at cl.cam.ac.uk <mailto:aac10 at cl.cam.ac.uk>>:
>>>
>>>                 I get those readings but note:
>>>
>>>                 3. For a long time, Kim didn't speak.
>>>
>>>                 only has your reading 2.
>>>
>>>                 so although I'd want to try and give an
>>>                 underspecified semantics for your sentence, one
>>>                 would have to do that in a way that recognised this
>>>                 has a different semantics.
>>>
>>>                 for negation there's an extensive literature - I'd
>>>                 recommend Horn's book.
>>>
>>>                 For some of these type of examples, I've played
>>>                 around with an account that decomposes the event
>>>                 variable so that one might claim that the negation
>>>                 was operating over different parts of a complex
>>>                 event structure in standard MRS.  But that only
>>>                 allows for 3 in a very stipulative way, if it works
>>>                 at all. Negated events are complicated.
>>>
>>>                 Incidentally, Ivan Sag (somewhere) had a discussion
>>>                 of examples like:
>>>
>>>                   The Sheriff of Nottingham jailed Robin Hood for
>>>                 three years.
>>>
>>>                 which may be relevant - I honestly can't remember.
>>>
>>>                 Anyway - I was trying to answer a slightly different
>>>                 type of question, which was what the semantics of
>>>                 unexpected_rel might be.  I was just trying to
>>>                 convey the modal flavour - not talking about the
>>>                 different readings the English sentence might have. 
>>>                 It may be that with some sort of account that did
>>>                 the negation examples, one could also get a
>>>                 non-scopal `unexpectedly' to give two structurally
>>>                 different readings, but that's a somewhat different
>>>                 issue.
>>>
>>>                 All best,
>>>
>>>                 Ann
>>>
>>>
>>>                 On 17/05/17 02:08, Guy Emerson wrote:
>>>>                 So, if I've understood correctly:
>>>>
>>>>                 - using a scopal modifier for negation only leaves
>>>>                 one variable for non-scopal modifiers
>>>>                 - using a modal for negation would allow non-scopal
>>>>                 modifiers to take either the main verb's variable,
>>>>                 or the modal's variable
>>>>
>>>>                 But then, what about "Kim didn't speak for a long
>>>>                 time", which I think can have two readings:
>>>>
>>>>                 1. Kim spoke for only a short time
>>>>                 2. Kim was silent for a long time
>>>>
>>>>                 It looks like the ERG just gets the first reading.
>>>>
>>>>
>>>>
>>>>                 2017-05-11 13:55 GMT-07:00 Ann Copestake
>>>>                 <aac10 at cl.cam.ac.uk <mailto:aac10 at cl.cam.ac.uk>>:
>>>>
>>>>                     I think /unexpectedly/ is scopal in at least
>>>>                     some circumstances.  Specifically I would say
>>>>                     the semantics of /unexpectedly/ is modal (in a
>>>>                     broad sense) - e.g., I could treat it in terms
>>>>                     of possible worlds that I'm considering at some
>>>>                     timepoint t - if in only 1% of possible worlds
>>>>                     does P happen, and P actually happens by t'
>>>>                     (where t' > t) then unexpected(P). This is very
>>>>                     crude and incomplete, but all I'm trying to do
>>>>                     here is convey the modal intuition.
>>>>
>>>>                     Under this interpretation:
>>>>
>>>>                     unexpected(not(win(Kim)))
>>>>
>>>>                     means that at time t I thought not(win(Kim))
>>>>                     had 1% chance, but at t' not(win(Kim)) has come
>>>>                     to pass
>>>>
>>>>                     this isn't the same as:
>>>>
>>>>                     not(unexpected(win(Kim)))
>>>>
>>>>                     which means it-is-not-the-case that [ at time t
>>>>                     I thought win(Kim) had 1% chance and at t'
>>>>                     win(Kim) has come to pass ]  i.e., either I
>>>>                     expected Kim to win all along or Kim actually
>>>>                     didn't win
>>>>
>>>>>                     Also, in (3), unexpectedly could be a
>>>>>                     sentence-initial discourse
>>>>>                     adverb (scopal?) or an adverb extracted from
>>>>>                     lower in the clause...
>>>>
>>>>                     As I remember it, the discussion about possible
>>>>                     sentence situation meaning is a semantic one
>>>>                     rather than depending on whether there's
>>>>                     extraction or not.
>>>>
>>>>                     All best,
>>>>
>>>>                     Ann
>>>>
>>>>
>>>>                     On 11/05/2017 21:13, Emily M. Bender wrote:
>>>>>                     Thanks, Ann, for the quick reply! This
>>>>>                     connects to other things I've been
>>>>>                     curious about recently, including how we
>>>>>                     decide if something like "unexpectedly"
>>>>>                     is scopal or not. Also, in (3), unexpectedly
>>>>>                     could be a sentence-initial discourse
>>>>>                     adverb (scopal?) or an adverb extracted from
>>>>>                     lower in the clause...
>>>>>
>>>>>                     Emily
>>>>>
>>>>>                     On Wed, May 10, 2017 at 2:11 AM, Ann Copestake
>>>>>                     <aac10 at cl.cam.ac.uk
>>>>>                     <mailto:aac10 at cl.cam.ac.uk>> wrote:
>>>>>
>>>>>                         I think the idea is to represent the
>>>>>                         contrast between:
>>>>>
>>>>>                         1   We could unexpectedly close the window.
>>>>>
>>>>>                         either ability to close or actual closure
>>>>>                         is unexpected
>>>>>
>>>>>                         2   We did not unexpectedly close the window.
>>>>>
>>>>>                         only the closure (if it had happened)
>>>>>                         would be unexpected.
>>>>>
>>>>>                         I don't think this is actually the best
>>>>>                         analysis.  For instance, for me,
>>>>>
>>>>>                         3 Unexpectedly we did not close the window.
>>>>>
>>>>>                         has another reading, which we are not
>>>>>                         capturing in MRS.  Claudia Maiernborn
>>>>>                         would (perhaps) treat this as a sentential
>>>>>                         situation rather than an event
>>>>>                         modification and it may be that analysis
>>>>>                         is also available for 1 instead of the
>>>>>                         modal modification analysis.
>>>>>
>>>>>                         I'm afraid I don't have time to discuss
>>>>>                         this properly at the moment, though.  I
>>>>>                         feel such a discussion has taken place,
>>>>>                         but don't remember the venue.
>>>>>
>>>>>                         All best,
>>>>>
>>>>>                         Ann
>>>>>
>>>>>
>>>>>
>>>>>                         On 10/05/2017 01:13, Emily M. Bender wrote:
>>>>>>                         Dear all,
>>>>>>
>>>>>>                         I'm curious about the different in
>>>>>>                         analysis between neg_rel and (other)
>>>>>>                         scopal adverbial
>>>>>>                         modifiers on the one hand and modals on
>>>>>>                         the other in the treatment of the INDEX:
>>>>>>
>>>>>>                         In (1) and (2), the INDEX of the whole
>>>>>>                         MRS points to the ARG0 of _sleep_v_rel:
>>>>>>
>>>>>>                         (1) Kim doesn't sleep.
>>>>>>                         (2) Kim probably sleeps.
>>>>>>
>>>>>>                         ... where in (3) and (4) it points to the
>>>>>>                         ARG0 of _can_v_rel and _would_v_rel
>>>>>>                         respectively:
>>>>>>
>>>>>>                         (3) Kim can sleep.
>>>>>>                         (4) Kim would sleep.
>>>>>>
>>>>>>                         I'm wondering what difference we intend
>>>>>>                         to model here.  (This question comes up now
>>>>>>                         because we're looking at negation in my
>>>>>>                         grammar engineering class, and the
>>>>>>                         out-of-the-box
>>>>>>                         analysis for languages which express
>>>>>>                         negation with an auxiliary has neg_rel
>>>>>>                         falling
>>>>>>                         in the latter class.)
>>>>>>
>>>>>>                         Thanks,
>>>>>>                         Emily
>>>>>>
>>>>>>
>>>>>>                         -- 
>>>>>>                         Emily M. Bender
>>>>>>                         Professor, Department of Linguistics
>>>>>>                         Check out CLMS on facebook!
>>>>>>                         http://www.facebook.com/uwclma
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>                     -- 
>>>>>                     Emily M. Bender
>>>>>                     Professor, Department of Linguistics
>>>>>                     Check out CLMS on facebook!
>>>>>                     http://www.facebook.com/uwclma
>>>>
>>>>
>>>
>>>
>>
>>
>>
>>
>>
>>     -- 
>>     Emily M. Bender
>>     Professor, Department of Linguistics
>>     Check out CLMS on facebook! http://www.facebook.com/uwclma
>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20170520/6fcd062a/attachment-0001.html>


More information about the developers mailing list