[erg] semantics of "so" as in "do so"

Ann Copestake Ann.Copestake at cl.cam.ac.uk
Thu Oct 31 01:20:36 CET 2013


> Debatable, but I get your point, thanks.
> 
> Does the fact that the pizzas are merely instrumental without performing 
> the satisfying explain the difference?  (Suggesting features for the 
> lexical entry indicating whether the subject performs or the object is 
> affected by the verb?)
> 

I don't know, I'm afraid.  I think intuitions about these examples are
generally rather unstable.

> Distinguishing lexical entries by semantics that rarely (if at all) 
> manifests themselves in surface realization seems nonetheless useful, 
> such as for ranking semantic interpretations for otherwise equivalent 
> parses.
> 
> It would be nice to have a representation that could vary from as 
> underspecified as MRS to more fully specified (even if an adequate "full 
> representation" may not be achievable).
> 
> That way, subsequent levels of analysis (including manual 
> disambiguation) could reduce underspecification within a standard or 
> uniform representation (which could be used in training, for example, 
> such as in improving the currently woeful state of text to logical form 
> technology).
> 
> This (text to less underspecified, if not fully specified FOL-ish) is 
> what we have to do to translate English with enough precision that we 
> can provide reliable, deep question answering systems, for example.
> 
> So, even if we don't distinguish the distributive/referential semantics 
> of reference in lexical entries, it would be nice to extend MRS or some 
> other USR to express such constraints.  (Same for the case of reference 
> previously discussed wrt "do so".)


there's a distinction between constraints which we need to be able to
build as part of compositional semantics and general constraints we
might want to be able to incrementally specialise an MRS - the latter
are much easier to work out without messing other things up.  As far
as distributivity facts go, one can go some way by specifying
constraints to apply to the verb's event/subevents, for example, and
that can just be done by adding a predication.  There are adverbial
phrases like "in pairs" (in "the girls danced in pairs") which one
might deal with that way, for instance (via a meaning postulate
operating on the compositionally constructed phrase, or whatever,
perhaps).  If there are such constraints that have to be worked out in
the grammar, because they are contributed by a construction, for
instance, we are more restricted in what we can do.

Cheers,

Ann

> 
> On 10/26/2013 4:33 PM, Ann Copestake wrote:
> > I'm sorry I have no time at the moment to properly contribute to this
> > interesting discussion but I thought I should mention that the
> > argument for making `die' distribute lexically is not clear, given
> > examples such as:
> >
> > `the entire team died in the plane crash' (distributes to the members
> > of the team)
> >
> > `the team died in 1993' - talking about the team as an entity
> >
> > Similarly, for `meet', which has been argued to be lexically collective:
> >
> > The unions and the management met (in separate rooms).
> >
> > and so on.
> >
> > And, of course, most verbs are much less clearcut than these.
> >
> > i.e., it's unclear that there are lexical constraints we can capture
> > in the compositional semantics and if we can't be sure we can do that
> > systematically, we leave everything underspecified.
> >
> > There's also the argument that if the additional constraints are
> > purely lexical, there's nothing to stop some later processing applying
> > those constraints to the MRS.
> >
> > The best candidate that I know of for a case where the MRS fails to
> > capture information with respect to distributivity is the argument
> > that there is a fully distributive, 12 pizza reading for:
> >
> > Three boys ate four pizzas.
> >
> > but no reading such that there are 12 boys.  If that distinction
> > exists systematically (i.e., subject distributes over object but not
> > vice versa), we're missing it in the MRS and that would be a problem.
> > However, I believe there's a 12 pizza reading for:
> >
> > Four pizzas satisfied three boys.
> >
> > (Think of contexts like `Four pizzas satisfied three boys but the
> > other boy wanted five pizzas'.)  If there is such a reading then
> > there's no uniform asymmetry due to the syntax, and the MRS approach
> > can be defended.
> >
> > Ann
> >
> >
> >> Also, perhaps MRS is too underspecified with regard to quantification
> >> and reference, as in "each man lifted each crate".   In this case, it is
> >> implicit that there are set of men and a set of crates which is not
> >> reflected in the ERG / MRS.
> >>
> >> Furthermore, the semantics of quantification may vary with lexical
> >> entry.  In "the men died", it occurs to each of them individually, no
> >> doubt, unlike in "the men lifted the crates".  Thus, the lexical entry
> >> for "die" could include semantics that indicate the event "occurs to"
> >> its subject(s) individually, while "lift" could be underspecified with
> >> regard to such argument feature.  Also consider the case of the unary
> >> "different" which requires a plurality as its single argument.  I'm not
> >> sure there is provision for such semantics in the ERG.
> >>
> >> Such semantic features might arguably improve the adverbial "so" in
> >> which the event lacks a quantifier, which if it existed would be
> >> explicit / definite rather than completely underspecified.
> >>
> >> Just saying... this can also be relegated to a further stage of
> >> processing, as you suggest, but the input lacks valuable information in
> >> such cases.
> >>
> >>
> >> On 10/26/2013 9:14 AM, Emily M. Bender wrote:
> >>> I believe this is intentional underspecification from the point of
> >>> view of the ERG.  If there is no syntactic configuration that can
> >>> disambiguate among the possibilities, then there is nothing to be
> >>> gained by enumerating them in different MRSs output by the grammar.
> >>>   That does, of course, leave the problem of enumerating and
> >>> disambiguating to a further stage of processing, though.
> >>>
> >>> Emily
> >>>
> >>>
> >>> On Sat, Oct 26, 2013 at 4:59 AM, Paul Haley <paul at haleyai.com
> >>> <mailto:paul at haleyai.com>> wrote:
> >>>
> >>>      With regard to "the men lifted the crates", and the general
> >>>      representation of resolved references in more fully specified
> >>>      semantics:
> >>>
> >>>      The classic problem is that any of the following interpretations
> >>>      is valid:
> >>>
> >>>      1. all the men together lifted all the crates at once.
> >>>      2. each crate was lifted by some of the men
> >>>      3. each crate was lifted by one of the men
> >>>      4. ...
> >>>
> >>>      The type of reference is more than just collective versus
> >>>      distributive reference since some of the crates may have been
> >>>      lifted by several of the men lifting together.
> >>>
> >>>      I'll table  "substances cross the plasma membrane at different
> >>>      rates", because there are multiple issues with the intended meaning
> >>>
> >>>        * for each pair of distinct substances that cross the plasma
> >>>          membrane the pair does so at different rates
> >>>
> >>>        * for each type of substance that crosses the plasma membrane
> >>>          for every other type of substance that crosses the plasma
> >>>          membrane the rate at which the first substance crosses the
> >>>          plasma membrane is different than the rate at which the second
> >>>          substance crosses the plasma membrane
> >>>
> >>>      vs. the ERG representation, as below:
> >>>
> >>>
> >>>
> >>>
> >>>      Paul
> >>>      On 10/25/2013 4:23 PM, Emily M. Bender wrote:
> >>>>      Yes, I would expect the do_so relation to show up for "and so did
> >>>>      Sandy", at least as one alternative parse.
> >>>>
> >>>>      I'm not sure what you mean by "the classic "the men lifted the
> >>>>      crates" or "substances cross the plasma membrane at different
> >>>>      rates"."
> >>>>
> >>>>      Emily
> >>>>
> >>>>
> >>>>      On Fri, Oct 25, 2013 at 10:19 AM, Paul Haley <paul at haleyai.com
> >>>>      <mailto:paul at haleyai.com>> wrote:
> >>>>
> >>>>          Perhaps I'm beginning to follow your perspective... thanks
> >>>>          for asking.
> >>>>
> >>>>          If, instead of resolving that ARG0 to the "leaving", which
> >>>>          Dan and I agree would be inappropriate for some potentially
> >>>>          intended semantics, there was a form of co-reference other
> >>>>          than (in)equality, such as the "same type" I suggested below,
> >>>>          I guess you would not need the additional argument for "so/it"!
> >>>>
> >>>>          Until now, I had no need to introduce additional predicates
> >>>>          into the semantics, but to do so seems appropriate after the
> >>>>          discussion , so thanks again for your patient coaching, Dan
> >>>>          and Emily.
> >>>>
> >>>>          Actually, any references on representing the forms of
> >>>>          reference as additional predicates in less underspecified
> >>>>          logical semantics would be sincerely appreciated. For
> >>>>          example, the classic "the men lifted the crates" or
> >>>>          "substances cross the plasma membrane at different rates".
> >>>>
> >>>>          I still think an issue lurks here, however, as shown in the
> >>>>          following examples.
> >>>>
> >>>>
> >>>>
> >>>>
> >>>>
> >>>>          Paul
> >>>>
> >>>>
> >>>>
> >>>>          On 10/24/2013 7:31 PM, Emily M. Bender wrote:
> >>>>>          Dear Paul,
> >>>>>
> >>>>>          Why doesn't the ARG0 of the do_so relation suffice for the
> >>>>>          variable/argument that you are looking for?
> >>>>>
> >>>>>          Emily
> >>>>>
> >>>>>
> >>>>>          On Tue, Oct 22, 2013 at 2:16 PM, Paul Haley
> >>>>>          <paul at haleyai.com <mailto:paul at haleyai.com>> wrote:
> >>>>>
> >>>>>              Thanks Dan.
> >>>>>
> >>>>>              No problem agreeing with the first clause of your second
> >>>>>              sentence, but I don't think the second clause involves
> >>>>>              coercion as much as proper logical semantic structure.
> >>>>>               I'm not suggesting anything more than a semantic
> >>>>>              argument is missing.
> >>>>>
> >>>>>              I'm not expecting sentence-oriented processing to do
> >>>>>              anything at all in terms of anaphora resolution.  That's
> >>>>>              how it works now (i.e., without a resolution mechanism,
> >>>>>              which is just fine, imo), but the variable/argument
> >>>>>              seems critical in any case.  I don't see how to approach
> >>>>>              it otherwise.  Inter- vs. intra-sentential resolution
> >>>>>              doesn't seem pertinent here.
> >>>>>
> >>>>>              I didn't know we had ICONSs!  Sounds interesting...  and
> >>>>>              potentially combinatoric.  Should be fun.  I'm not sure
> >>>>>              additional types of constraints for verbal anaphora are
> >>>>>              needed (i.e., we've lived long enough without ICONSs)
> >>>>>              but if there is a linguistic distinction between such
> >>>>>              reference to an instance versus a type of event, that
> >>>>>              would be appropriate too.  Without them, the semantics
> >>>>>              is just more underspecified, which is fine as long as it
> >>>>>              covers the intended meaning.  It seems clear that the
> >>>>>              representation we have now does not in the cases at
> >>>>>              issue here.
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>              On 10/22/2013 1:57 PM, Dan Flickinger wrote:
> >>>>>
> >>>>>                  Hi Paul -
> >>>>>
> >>>>>                  I agree that we don't want to identify the two
> >>>>>                  events as the only possible interpretation.  But we
> >>>>>                  have to allow anaphora resolution to perform its
> >>>>>                  magic quite generally, and it is I think misguided
> >>>>>                  to try to coerce the sentence-level semantic
> >>>>>                  composition to do too much.  If the text to be
> >>>>>                  parsed were two separate sentences, I hope you would
> >>>>>                  agree that our sentence-oriented processing could
> >>>>>                  not be expected to constrain the elided event via
> >>>>>                  unification:
> >>>>>                  "Kim bought a car.  Mary did so, too."
> >>>>>                  So we have to be content in the grammar to set the
> >>>>>                  stage for a currently unimplemented resolution
> >>>>>                  engine separate from the current grammar that will
> >>>>>                  bind these anaphoric elements both
> >>>>>                  sentence-internally and across discourse.  This
> >>>>>                  underspecification is very much the same as the
> >>>>>                  approach we take to ordinary pronoun binding, though
> >>>>>                  we do expect to enrich the grammar's MRSs for
> >>>>>                  sentence with pronouns a little more, now that we
> >>>>>                  have a place to assert structurally derived
> >>>>>                  constraints on equality and inequality of
> >>>>>                  individuals, as ICONSs.  But I don't know of
> >>>>>                  analogous structural constraints (such as c-command)
> >>>>>                  for verbal anaphora, and in any case these would
> >>>>>                  again be only sentence-internal.
> >>>>>
> >>>>>                    Dan
> >>>>>
> >>>>>                  ----- Original Message -----
> >>>>>                  From: "Paul Haley" <paul at haleyai.com
> >>>>>                  <mailto:paul at haleyai.com>>
> >>>>>                  To: "Emily M. Bender" <ebender at uw.edu
> >>>>>                  <mailto:ebender at uw.edu>>, "erg" <erg at delph-in.net
> >>>>>                  <mailto:erg at delph-in.net>>
> >>>>>                  Sent: Monday, October 21, 2013 7:39:03 AM
> >>>>>                  Subject: Re: [erg] semantics of "so" as in "do so"
> >>>>>
> >>>>>                  What logical semantics is appropriate for "kim left
> >>>>>                  and sandy did so, too"?
> >>>>>
> >>>>>                  They may have left together or at different times or
> >>>>>                  independently at
> >>>>>                  the same time.
> >>>>>
> >>>>>                  In theory, all of these logical/semantic
> >>>>>                  interpretations should be
> >>>>>                  consistent with the resulting underspecified semantics.
> >>>>>
> >>>>>                  The MRS below corresponds, roughly to:
> >>>>>
> >>>>>                  exists(e10,e14,x6,x17){leave(e10,x6),named(x6,Kim),do_so(e14,x17),named(x17,Sandy)}
> >>>>>
> >>>>>                  If e14 co-references e10, this implies there is one
> >>>>>                  leaving event
> >>>>>                  "performed by" both Kim and Sandy, which may or may
> >>>>>                  not be the
> >>>>>                  appropriate logical interpretation.
> >>>>>
> >>>>>                  If not, how is e14 to understood as a leaving?
> >>>>>
> >>>>>                  One resolution of this would be to have an argument,
> >>>>>                  such as follows:
> >>>>>
> >>>>>                  exists(e10,e14,x6,x17){leave(e10,x6),named(x6,Kim),do(e14,e10,x17),named(x17,Sandy)}
> >>>>>
> >>>>>                  This representation would allow either semantics to
> >>>>>                  result from further
> >>>>>                  (logical) disambiguation.
> >>>>>
> >>>>>                  I submit that the MRS resulting now is insufficient
> >>>>>                  to represent the
> >>>>>                  underspecified semantics.
> >>>>>
> >>>>>                  Alternatively, I suppose, one could introduce an
> >>>>>                  underspecified form of
> >>>>>                  co-reference in which e14  references e10 other than
> >>>>>                  as logically
> >>>>>                  equivalent, but that raises issues not previously
> >>>>>                  addressed (in any
> >>>>>                  literature that I have seen) with regard to the
> >>>>>                  relationship between
> >>>>>                  underspecified representation and logical axioms.
> >>>>>
> >>>>>
> >>>>>                  On 10/15/2013 11:29 AM, Emily M. Bender wrote:
> >>>>>
> >>>>>                      [Keeping this on-list]
> >>>>>
> >>>>>                      Hi Paul,
> >>>>>
> >>>>>                      The analysis in the ERG is that do+so is a
> >>>>>                      pro-verb, the whole thing
> >>>>>                      stands in for the event.  The point of my
> >>>>>                      examples was that that event
> >>>>>                      might have any number of participants, and so
> >>>>>                      looking for an ARG2
> >>>>>                      specifically seems misguided.
> >>>>>
> >>>>>                      Emily
> >>>>>
> >>>>>
> >>>>>
> >>>>>                      On Sat, Oct 12, 2013 at 2:56 AM, Paul Haley
> >>>>>                      <paul at haleyai.com <mailto:paul at haleyai.com>
> >>>>>                      <mailto:paul at haleyai.com
> >>>>>                      <mailto:paul at haleyai.com>>> wrote:
> >>>>>
> >>>>>                           But in the MRS there is nothing that
> >>>>>                      relates the doing to the
> >>>>>                           leaving or betting!?  That's the problem.
> >>>>>                      The semantics is wrong.
> >>>>>
> >>>>>                           Worse, "it" is frequently interchangeable
> >>>>>                      with "so" in such
> >>>>>                           constructions, as shown below (as in the
> >>>>>                      case of my first example
> >>>>>                           further below).  The pronoun refers to the
> >>>>>                      event, of course.  That
> >>>>>                           reference is missing in the semantics for "so".
> >>>>>
> >>>>>                           Seems to me that "so" in this construction
> >>>>>                      is an 'e' pronoun
> >>>>>                           (where "it" is a 'x' pronoun below, which
> >>>>>                      could also be a bug, imo.)
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>                           On 10/10/2013 7:28 PM, Emily M. Bender wrote:
> >>>>>
> >>>>>                               But "so" in "do so" doesn't actually
> >>>>>                          stand in for the ARG2:
> >>>>>
> >>>>>                               Kim left, and Sandy did so too.
> >>>>>                               Kim bet Pat $500 that the Giants would
> >>>>>                          win, and Sandy did so too.
> >>>>>
> >>>>>                               Emily
> >>>>>
> >>>>>
> >>>>>
> >>>>>                               On Wed, Oct 9, 2013 at 5:42 AM, Paul
> >>>>>                          Haley <paul at haleyai.com
> >>>>>                          <mailto:paul at haleyai.com>
> >>>>>                               <mailto:paul at haleyai.com
> >>>>>                          <mailto:paul at haleyai.com>>> wrote:
> >>>>>
> >>>>>                                   Apologies for a couple of typos
> >>>>>                          below, and one clarification.
> >>>>>
> >>>>>                                   It's not really important whether
> >>>>>                          "so" is treated as a
> >>>>>                                   pronoun or do-so as a proto-verb
> >>>>>                          but by "direct object" I
> >>>>>                                   meant an ARG2 in the predication
> >>>>>                          for do_v_so corresponding to
> >>>>>                                   whatever "so" references or
> >>>>>                          introduces or substitutes for.
> >>>>>
> >>>>>
> >>>>>
> >>>>>                                   On 10/9/2013 8:34 AM, Paul Haley wrote:
> >>>>>
> >>>>>                                       Hi Emily!
> >>>>>
> >>>>>                                       Yes, but I'm suggest that
> >>>>>                              "pro-" is "pronomial" not
> >>>>>                                       "proto"!-) Generally, don't we
> >>>>>                              want elipsis to be reflected
> >>>>>                                       in the semantics?  That is, in
> >>>>>                              the "u" and "i" type
> >>>>>                                       variables in the MRS (or
> >>>>>                              unresolved pronouns)?
> >>>>>
> >>>>>                                       Shouldn't the MRS for for that
> >>>>>                              doing have an argument to be
> >>>>>                                       resolved against the
> >>>>>                              situational argument for the moving?
> >>>>>                                       That argument would be "so"
> >>>>>                              treated as a pronoun, which
> >>>>>                                       seems the proper semantics
> >>>>>                              since the "so"
> >>>>>                               actually/semantically references some
> >>>>>                              event/situation, no?
> >>>>>                                       That is, if pronomial "so" was
> >>>>>                              the direct object of "do"
> >>>>>                                       here, I think all would be well.
> >>>>>
> >>>>>                                       Paul
> >>>>>
> >>>>>
> >>>>>
> >>>>>                                       On 10/8/2013 9:14 PM, Emily M.
> >>>>>                              Bender wrote:
> >>>>>
> >>>>>                                           Hello Paul,
> >>>>>
> >>>>>                                           It looks like "do so" is
> >>>>>                                  being treated as a "pro-verb", and
> >>>>>                                           that seems appropriate to
> >>>>>                                  me.  Proverbs (like ellipsis)
> >>>>>                                           take their interpretation
> >>>>>                                  from context.  So this says
> >>>>>                                           basically that
> >>>>>                                           x6 is doing something, but
> >>>>>                                  what that something is needs to
> >>>>>                                           be resolved.
> >>>>>
> >>>>>                                           Emily
> >>>>>
> >>>>>
> >>>>>                                           On Tue, Oct 8, 2013 at 9:33
> >>>>>                                  AM, Paul Haley
> >>>>>                                           <paul at haleyai.com
> >>>>>                                  <mailto:paul at haleyai.com>
> >>>>>                                  <mailto:paul at haleyai.com
> >>>>>                                  <mailto:paul at haleyai.com>>> wrote:
> >>>>>
> >>>>>                                               Hi All,
> >>>>>
> >>>>>                                               In the following, it
> >>>>>                                  seems that "so" is more of a
> >>>>>                                   pronoun than a preposition (at
> >>>>>                                  least it seems "so" to me!).
> >>>>>
> >>>>>                                               I would appreciate your
> >>>>>                                  thoughts on getting reasonable
> >>>>>                                               logic from the ERG for
> >>>>>                                  this sentence, which is quite
> >>>>>                                   interesting when you also consider
> >>>>>                                  quantification...
> >>>>>
> >>>>>
> >>>>>
> >>>>>                                               Thank you and best regards,
> >>>>>                                               Paul
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>                                           --
> >>>>>                                           Emily M. Bender
> >>>>>                                           Associate Professor
> >>>>>                                   Department of Linguistics
> >>>>>                                           Check out CLMS on facebook!
> >>>>>                                  http://www.facebook.com/uwclma
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>                               --
> >>>>>                               Emily M. Bender
> >>>>>                               Associate Professor
> >>>>>                               Department of Linguistics
> >>>>>                               Check out CLMS on facebook!
> >>>>>                          http://www.facebook.com/uwclma
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>                      --
> >>>>>                      Emily M. Bender
> >>>>>                      Associate Professor
> >>>>>                      Department of Linguistics
> >>>>>                      Check out CLMS on facebook!
> >>>>>                      http://www.facebook.com/uwclma
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>          --
> >>>>>          Emily M. Bender
> >>>>>          Associate Professor
> >>>>>          Department of Linguistics
> >>>>>          Check out CLMS on facebook! http://www.facebook.com/uwclma
> >>>>
> >>>>
> >>>>
> >>>>      --
> >>>>      Emily M. Bender
> >>>>      Associate Professor
> >>>>      Department of Linguistics
> >>>>      Check out CLMS on facebook! http://www.facebook.com/uwclma
> >>>
> >>>
> >>>
> >>> -- 
> >>> Emily M. Bender
> >>> Associate Professor
> >>> Department of Linguistics
> >>> Check out CLMS on facebook! http://www.facebook.com/uwclma
> 




More information about the erg mailing list