search

delph-in / erg

English Resource Grammar
MIT License
17 stars 3 forks source link

Swapped arguments with subject raising? #11

Closed goodmami closed 1 year ago

goodmami commented 5 years ago

It looks like the semantic arguments of subject raising swap depending on the construction used. Consider the following (I've indicated the questionable parts of the MRS in the images below):

The _appear_v_to nodes occupy the same place in the graph (both are the scopal top and link only to _sleep_v_1), but the role (or argument position) is different. In the first sentence the ARG2 links to _sleep_v_1, while in the it-clefted expletive-it version the ARG1 does.

The SEM-I lists two predicate synopses:

  _appear_v_to : ARG0 e, ARG1 u, ARG2 h.
  _appear_v_to : ARG0 e, ARG1 h, ARG2 i.

Although it is unclear why one has 'u' and the other 'i' for the optional argument. But this part of the SEM-I is generated automatically, so it is just reflecting the variation that is encoded in the grammar.

Note that filling that argument ("Kim appears to me to sleep." and "It appears to me Kim sleeps") does not affect change the argument order (i.e., they are also swapped as before).

This appears to me to be a bug.

goodmami commented 5 years ago

I edited the issue because the example wasn't an it-cleft but an expletive it. There is another issue with it-clefts (#8). Although "It is appearing that Kim sleeps." also exhibits the same argument swapping. I think the original sentence ("Kim appears to sleep.") has the proper semantics.

Also, this does not happen with some other kinds of raising: "Kim expected Pat to sleep" and "It was expected that Pat sleep." both have _sleep_v_1 on the ARG2. Similarly, both "That Kim sleeps is bad" and "It is bad that Kim sleeps." have _sleep_v_1 on _bad_a_for's ARG1.

danflick commented 5 years ago

I agree that the ARGn assignments should be the same for "Kim appears to sleep" and "It appears that Kim sleeps", so the reported behavior is indeed a bug. But since there are other verbs such as "follow" that take an expletive-it subject and only a sentential complement (with no optional PP, and hence no second semantic argument), it seems better to treat the obligatory sentential complement consistently as the ARG1, and the optional PP as the ARG2. So once I have checked in the corrections, you should see only the one SEM-I entry: _appear_v_to : ARG0 e, ARG1 h, ARG2 i. and this should be used for both variants: subj-raising and expl-it subject.

goodmami commented 5 years ago

Ok thanks. My statement that "Kim appears to sleep" has the proper semantics was based mainly on my comparisons to expect, but on further reflection those are rather different semantically so such a comparison is not motivated. I agree with your choice of argument order on _appear_v_to (and related predicates).

It would be nice to have some established consistency in the argument order for categories of words. I could probably put together some kind of regression-test-like setup to ensure things are as we expect across grammar versions. These lists of words/predicates could probably also be documented somewhere (the ErgSemantics wikis?). E.g.:

Although I guess if you know the lexical type you can look them up in the LTDB, but frankly I have a hard time decoding lexical type names.

fcbond commented 5 years ago

I think some classes we have also linked to wikipedia (control verbs?). David, were those mappings released anywhere?

On Sat, Apr 13, 2019 at 9:37 AM Michael Wayne Goodman < notifications@github.com> wrote:

Ok thanks. My statement that "Kim appears to sleep" has the proper semantics was based mainly on my comparisons to expect, but on further reflection those are rather different semantically so such a comparison is not motivated. I agree with your choice of argument order on _appear_v_to (and related predicates).

It would be nice to have some established consistency in the argument order for categories of words. I could probably put together some kind of regression-test-like setup to ensure things are as we expect across grammar versions. These lists of words/predicates could probably also be documented somewhere (the ErgSemantics wikis?). E.g.:

  • follow, help, ...
  • appear, seem, ...
  • expect, ask
  • promise, ...
  • try, ...

Although I guess if you know the lexical type you can look them up in the LTDB, but frankly I have a hard time decoding lexical type names.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/delph-in/erg/issues/11#issuecomment-482765559, or mute the thread https://github.com/notifications/unsubscribe-auth/ABD8xi_q1preleST_q6Yu9PUzCEPoHNxks5vgTTugaJpZM4comPM .

-- Francis Bond http://www3.ntu.edu.sg/home/fcbond/ Division of Linguistics and Multilingual Studies Nanyang Technological University

danflick commented 1 year ago

I will aim to add more explicit documentation to lexical types to assist in their decoding, and add some better descriptions of more general lexical types such as those for control. Closing this particular issue.