search

propbank / propbank-frames

Lexicon of frame files used by Propbank annotation. A searchable, readable version of the latest release is here: http://propbank.github.io/v3.4.0/frames/
Creative Commons Attribution Share Alike 4.0 International
95 stars 27 forks source link

make.02: missing sense or arg? #9

Closed arademaker closed 2 years ago

arademaker commented 2 years ago

One of the example is:

Financial problems make them unattractive takeover targets.

The UD analysis, based on EWT corpus, take the token them the iobj of make. I would annotate as obj, but anyway, this is a dependent of the verb make, a different constituent from the rest of the sentence unattractive takeover targets.

image

In the sense documentation, the example is analysed as them as part of the constituent them unattractive takeover targets. This decision is probably based on the constituent analysis of a similar sentence in OntoNotes:

(TOP (S (S-NOM-SBJ (NP-SBJ (-NONE- *PRO*))
                   (VP (VBG Living)
                       (PP-LOC-CLR (IN in)
                                   (NP (NNP Hong)
                                       (NNP Kong)))))
        (PRN (, ,)
             (S (NP-SBJ (PRP he))
                (VP (VBZ says)
                    (SBAR (-NONE- 0)
                          (S (-NONE- *?*)))))
             (, ,))
        (VP (VBD made)
            (S (NP-SBJ (PRP him))
               (ADJP-PRD (JJ sensitive)
                         (PP (IN to)
                             (NP (NP (DT the)
                                     (NNS limits))
                                 (PP (IN on)
                                     (NP (NML (NN food)
                                              (, ,)
                                              (NN power)
                                              (CC and)
                                              (NN water))
                                         (NNS supplies))))))))
        (. .)))
  1. Arg0: Financial problems
  2. Rel: make
  3. Arg1: them unattractive takeover targets

What about the UD analysis? What about this example be the case of sense make.01 if we take the syntatic analysis of UD?

  1. Arg0: Financial problems
  2. Rel: made
  3. Arg1: them
  4. Arg2: unattractive takeover targets
MarthaSPalmer commented 2 years ago

This is tricky.

Make.01 is really supposed to be restricted to creation events and make.02 is intended to to fit this example, as well as similar ones that get a small clause analysis. It matches the VN act.114 class.

But the story is more complicated. The original WSJ treebank had a small clause analysis for a LOT of things that PB gave a ditransitive analysis to, in concert with Levin classes. We had a long negotiation over those kinds of verbs and ending up doing a major retrofit of both the TB and the PB to better synchronize them, where we changed both TB and PB. It is a gray area where there are strong arguments for both of the competing analyses. I suspect this was a place where PB lost the negotiation, and we changed to match TB. Part of my evidence for that is the Render.29.90 VerbNet class which predates Act.114 by several years and which matches the UD analysis. On the whole we changed TB significantly more than PB, but not here.

Martha

On Nov 17, 2021, at 9:07 AM, Alexandre Rademaker @.***> wrote:



One of the example is:

Financial problems make them unattractive takeover targets.

The UD analysis, based on EWT corpus, take the token them the iobj of make. I would annotate as obj, but anyway, this is a dependent of the verb make, a different constituent from the rest of the sentence unattractive takeover targets.

[image]https://user-images.githubusercontent.com/86403/142230715-41444aa5-5d8c-49b7-8feb-3155256a6992.png

In the sense documentation, the example is analysed as them as part of the constituent them unattractive takeover targets. This follows the constituent analysis of a similar sentence in OntoNotes:

(TOP (S (S-NOM-SBJ (NP-SBJ (-NONE- PRO)) (VP (VBG Living) (PP-LOC-CLR (IN in) (NP (NNP Hong) (NNP Kong))))) (PRN (, ,) (S (NP-SBJ (PRP he)) (VP (VBZ says) (SBAR (-NONE- 0) (S (-NONE- ?))))) (, ,)) (VP (VBD made) (S (NP-SBJ (PRP him)) (ADJP-PRD (JJ sensitive) (PP (IN to) (NP (NP (DT the) (NNS limits)) (PP (IN on) (NP (NML (NN food) (, ,) (NN power) (CC and) (NN water)) (NNS supplies)))))))) (. .)))

  1. Arg0: Financial problems
  2. Rel: make
  3. Arg1: them unattractive takeover targets

What about the UD analysis? What about this example be the case of sense make.01 if we take the syntatic analysis of UD?

  1. Arg0: Financial problems
  2. Rel: made
  3. Arg1: them
  4. Arg2: unattractive takeover targets

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/propbank/propbank-frames/issues/9, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ABB327WOFQL4LTZERKZJIWDUMPG2FANCNFSM5IHLTIWA. Triage notifications on the go with GitHub Mobile for iOShttps://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Androidhttps://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

nschneid commented 2 years ago

@arademaker TBC, the UD parse you showed is not from EWT, right? EWT would specify obj, not iobj. examples

arademaker commented 2 years ago

UDPipe trained with EWT 2.6:

image
nschneid commented 2 years ago

OK. Could be a parser error, or the data could have changed between the 2.6 and 2.8 releases.

leoalenc commented 2 years ago

@arademaker , from a syntactical point of view, this does not make any sense to me:

  1. Arg0: Financial problems
  2. Rel: make
  3. Arg1: them unattractive takeover targets

This analysis does not take into account the fact that them is an object of make and, at the same time, the ("logical" or better external) subject of the relation unattractive takeover targets.

What about the UD analysis? What about this example be the case of sense make.01 if we take the syntatic analysis of UD?

I agree with this analysis:

  1. Arg0: Financial problems
  2. Rel: made
  3. Arg1: them
  4. Arg2: unattractive takeover targets

In Deep UD, Arg1 is the external subject of Arg2, which is an xcomp.

Abstracting away from syntax, a modified version of the first analysis would make sense, according to the analysis below:

dmrs (14)

The verb make in this case seems to be an object raising verb, i.e., its direct object them is not one of its semantic arguments. Instead, it is a semantic argument (ARG1) of its second argument (ARG2), as shown in the dependency graph above.

leoalenc commented 2 years ago

@arademaker TBC, the UD parse you showed is not from EWT, right? EWT would specify obj, not iobj. examples

@arademaker , iobj is definitely wrong, it should be obj.

arademaker commented 2 years ago

Hi @leoalenc, yes definitely obj not iobj if we take the UD ditransitive analysis.

Since you mentioned, investigating the ERG HPSG Grammar analyses:

  1. ontonotes sentence: Living in Hong Kong, he says, made him sensitive to the limits on food, power and water supplies.

    % echo Living in Hong Kong, he says, made him sensitive to the limits on food, power and water supplies. | ace -g hpsg/sick/erg.dat -Tf -n 60000 | grep make_v | awk '{print $2}' | sort | uniq -c
NOTE: hit RAM limit while unpacking
NOTE: 55290 readings, added 20576 / 14857 edges to chart (3520 fully instantiated, 794 actives used, 4557 passives used)    RAM: 1536019k
NOTE: parsed 1 / 1 sentences, avg 1536019k, time 15.37637s
32196 _make_v_1<30:34>
23094 _make_v_cause<30:34>

  2. propbank example for make.02: Financial problems make them unattractive takeover targets.

    $ echo Financial problems make them unattractive takeover targets. | ace -g hpsg/sick/erg.dat -Tf -n 100 | grep make_v | awk '{print $2}' | sort | uniq -c
NOTE: 8 readings, added 3010 / 1053 edges to chart (329 fully instantiated, 219 actives used, 236 passives used)    RAM: 11626k
NOTE: parsed 1 / 1 sentences, avg 11626k, time 0.49658s
8 _make_v_1<19:23>

  3. @leoalenc example: The money made her attractive.

    % echo The money made her attractive. | ace -g hpsg/sick/erg.dat -Tf -n 100 | grep make_v | awk '{print $2}' | sort | uniq -c
NOTE: 4 readings, added 2122 / 707 edges to chart (224 fully instantiated, 149 actives used, 194 passives used) RAM: 9780k
NOTE: parsed 1 / 1 sentences, avg 9780k, time 0.53433s
3 _make_v_1<10:14>
1 _make_v_cause<10:14>

In [1] among the 55.290 possible analyses we have a good distribution between the ditransitive and small clause analyses. In [2] we have only the ditransitive analysis (pointed by @MarthaSPalmer). In your example [3] we have also both analyses with the preference for the ditransitive analysis. This is ERG trunk version that has parsing selection model traned in the WSJ corpus (sentence 1 is part of it).

In my conversion from the TB to UD, the small clause analysis from TB was converted to them [2] as subj of the ADJ unattractive which does not make sense given the accusative case of the preposition. So my questions are:

  1. If we keep the UD ditransitive analysis, what sense of make should/can I use?
  2. how to make a small clause analysis in UD for sentence [2] and [1]? The accusative pronous can't be subject of the adjective, right?
arademaker commented 2 years ago

Considering that in the UD analysis the semantic argument of make is discontinuous. That is, part is the OBJ of want and the rest is xcomp, if we follow the propbank guidelines (Sec 1.9), we can use the C-ARG1 for annotating them?

MarthaSPalmer commented 2 years ago

Yes

Martha

On Nov 17, 2021, at 12:08 PM, Alexandre Rademaker @.**@.>> wrote:

Considering that in the UD analysis the semantic argument of make is discontinuous. That is, part is the OBJ of want and the rest is xcomp, if we follow the propbank guidelines, we can use the C-ARG1https://github.com/propbank/propbank-release/blob/master/docs/conll-conversion-notes.md#c--discontinuous-arguments for annotate them?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/propbank/propbank-frames/issues/9#issuecomment-971888361, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ABB327Q7IIBACCORB3HTC6TUMP4SVANCNFSM5IHLTIWA. Triage notifications on the go with GitHub Mobile for iOShttps://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Androidhttps://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

leoalenc commented 2 years ago

@arademaker , this use of the verb make exemplifies one of the many known mismatches between semantics and syntax. Semantically, make is a two-place predicate: a causer brings about a (parametrized) state-of-affairs (psoa), see Davis (2001) on the term psoa. Syntactically, it is a raising-to-object verb: the first argument of the embedded predicate (i.e., the psoa or small-clause in the Chomskyan generative syntax tradition) "raises" to the object position of the higher verb (the causative verb in this case). There's no dative case in languages such as English, German or Brazilian Portuguese. In French and European Portuguese (EP), causative verbs build complex predicates, so that when the embedded predicate is transitive, the CAUSEE is assigned dative case:

Eu mandei escrever uma carta aos alunos. (EP) (See Mateus et al. 1989, Falk 2001 on French causatives) I made the students [accusative case!] write a letter. Ich habe die Studenten [accusative case!] einen Brief schreiben lassen. (German)

In Deep UD, there should be the addtional annotation on the obj of make, specifying that it is the external subject of the small clause/psoa/embedded predicate.