[developers] Valid MRS? Bug in ERG?

[goodman.m.w at gmail.com]

Thanks, Emily, for the correction!

On Fri, Sep 25, 2020 at 11:22 AM Alexandre Rademaker <arademaker at gmail.com>
wrote:

>
> Ops! I came to hasty conclusions! LKB FOS parsed the sentence with ERG
> trunk. Two analysis, two MRS and two EDS with no error!! In the MRSs from
> LKB, ARG2 of `_communicate_v_to` and `_express_v_to` are both qeq to the
> label of the `unknown` predicate.
>
> I don’t know what I can conclude now… ACE error?
>

Regarding the MRS you reported: broken HCONS are a somewhat common issue,
and they are the symptom of a bug. The MRS is not valid.

Regarding conversion to EDS, the LKB goes to greater lengths to give an EDS
even for ill-formed MRSs, while PyDelphin tries to avoid grammar-specific
solutions. The result is that PyDelphin will, I think, apply
predicate-modification more broadly than the LKB, but it is a bit more
brittle.

While there may be other reasons to use DMRS instead, in this case the
different behavior in PyDelphin is just because it issues a warning instead
of an error, then prints out the partial DMRS, dropping the disconnected
nodes.



> > On 25 Sep 2020, at 00:05, Alexandre Rademaker <arademaker at gmail.com>
> wrote:
> >
> >
> > Thank you Emily, I found one mention in
> https://plato.stanford.edu/entries/situations-semantics/. But in the
> context of the Lisp code, I am curious why the authors used this in the
> suffix of the function name…
> >
> > BTW, I reported two issues on pydelphin and ERG repositories:
> >
> > https://github.com/delph-in/erg/issues/25
> >
> > The trunk version of ERG gave me this MRS..
> >
> > [ TOP: h0
> >  INDEX: e2 [ e SF: prop TENSE: tensed MOOD: indicative PROG: - PERF: - ]
> >  RELS: < [ _communicate_v_to<0:11> LBL: h1 ARG0: e4 [ e SF: prop TENSE:
> tensed MOOD: indicative PROG: - PERF: - ] ARG1: i3 ARG2: h5 ARG3: i6 ]
> >          [ _or_c<12:14> LBL: h1 ARG0: e2 ARG1: e4 ARG2: e7 [ e SF: prop
> TENSE: pres MOOD: indicative PROG: - PERF: - ] ]
> >          [ _express_v_to<15:22> LBL: h1 ARG0: e7 ARG1: i3 ARG2: h8 ARG3:
> i9 ]
> >          [ unknown<23:33> LBL: h10 ARG: u12 ARG0: e11 [ e SF: prop
> TENSE: untensed MOOD: indicative ] ]
> >          [ _by_p_means<23:25> LBL: h10 ARG0: e11 ARG1: u13 ARG2: x14 ]
> >          [ udef_q<26:33> LBL: h15 ARG0: x14 RSTR: h16 BODY: h17 ]
> >          [ nominalization<26:33> LBL: h18 ARG0: x14 ARG1: h19 ]
> >          [ _write_v_to<26:33> LBL: h19 ARG0: e20 [ e SF: prop TENSE:
> untensed MOOD: indicative PROG: + PERF: - ] ARG1: i21 ARG2: i22 ] >
> >  HCONS: < h0 qeq h1 h5 qeq h23 h8 qeq h23 h16 qeq h18 > ]
> >
> > Handle h23 does not appear in the predicates. The h5 and h8 only in the
> arguments. Is it valid?
> >
> > Pydelphin transformation to DMRS works with a warning "broken handle
> constraint”. Can it be transformed to EDS? Is this an evidence that MRS to
> EDS is much less robust than MRS to DMRS?
> >
> > BTW, since I am having so many trouble with MRS to EDS, and my goal is
> to compare a golden version of a profile with its 1-best parsed version to
> evaluate the parse selection model, I wonder if I could be doing that with
> DMRS instead of EDS… any idea? Any alternative to
> https://github.com/delph-in/delphin.edm using DMRS?
> >
> > Best,
> > Alexandre
>
>
>

-- 
-Michael Wayne Goodman
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20200925/d92946b3/attachment.html>