[developers] [sdp-organizers] From EDS/RMS to DM

Stephan Oepen oe at ifi.uio.no
Thu Jul 30 13:15:47 CEST 2020


hi again, alexandre,

in general, i used to recommend that most users work with actual ERG
releases rather than with whatever state you find in the trunk on a
given day (which, after all, is an internal work in progress).  from
your observations, it sounds as if dan (possibly around his joint work
with colleagues at NTU) is experimenting with a mal-configuration of
the ERG, and just now at least the default parameterization of the
grammar in ACE differs from the defaults in the LKB and PET; that
would likely not be the case in a release.  from what you describe, i
doubt you want the mal-extensions in your parses?

for a grammar to be compatible with a treebank means that it can
re-build all derivation trees recorded in the profile.  the 'same'
grammar will always be compatible, but sometimes it can be desirable
to actually improve (or revise) the grammar in ways that do not
inhibit re-unification of derivation trees but change the contents of
the feature structures and, thus, derived representations like the
MRS, EDS, DM, etc.  this is one aspect in which we refer to the
Redwoods treebanking approach as 'dynamic': the gold-standard HPSG
derivation can be output in various derived views.

exporting from a treebank is an interpretative process, i.e. there is
no way to make it succeed (in how i designed things in [incr tsdb()]
at least) without re-building all recorded derivations.  arguably,
MRSs should not be recorded in the treebanked profiles (they are there
in recent ERG releases for convenience).  the LOGON 'redwoods' scripts
can be forced to always re-compute them, using the '/blind' modifier
on its '--export' option.

the LOGON environment provides the 'terg' (for trunk or test or trial)
target so that users can put a grammar version of their choice there;
please see the 'LogonExtras' page on the wiki for details; i expect it
should work to 'switch' to the 2018 release of the ERG roughly as
follows

cd $LOGONROOT/terg
svn switch $LOGONSVN/erg/tags/2018

once you are in a universe with a grammar (when loaded into the LKB)
that matches your treebanked derivations, i would hope that exporting
to DM will also become functional?  as you note, there is a
non-trivial amount  of grammar-specific configuration in the DM
converter (categorizing different predicates into the various classes
distinguished by ivanova et al., 2012), which could lead to
sub-optimal results here and there.  however, from what i know about
the ERG evolution between 1214 and 2018, i believe the MRSs have been
comparatively stable, so DM exports from a 2018 treebank should still
be decent, i would hope!

best wishes, oe

On Tue, Jul 28, 2020 at 9:07 PM Alexandre Rademaker
<arademaker at gmail.com> wrote:
>
>
> > On 28 Jul 2020, at 13:08, Stephan Oepen <oe at ifi.uio.no> wrote:
> >
> > the export code will want to rebuild the derivation, i.e. the version of the grammar loaded needs to be fully compatible with the treebank (or parsed profile).
>
> Do you mean that `redwoods` reads the derivation just to check if the grammar passed as parameter to it was compatible with the grammar used to process the profile? So can I bypass this check and simply copy the derivation tree to the .gz file?
>
> What does it means a grammar be full compatible with a profile? Does it means that the grammar is the same used to process the profile?
>
> > i wonder whether ‘a_det_rbst’ is available at the time of exporting?  it sounds like a mal-configuration of the grammar, maybe?
> > which you would have to match on the LKB side then, e.g. push the right feature or load the right ‘script’ file?
>
> Yes, you are right. I found this entry in the lexicon-rbst.tdl:
>
> a_det_rbst := d_-_sg-a_le_mal &
>  [ ORTH < "a" >,
>    SYNSEM [ LKEYS.KEYREL.PRED _a_q_rel,
>             PHON.ONSET voc ] ].
>
>
> This file is included in the english.tdl file and ACE loads to the ace/config.tdl that declares english.tdl as the grammar-top. But LKB loads the lkb/script and it doesn’t mentioned the english.tdl… So you are probably right. Unfortunately, I don’t know how to make LKB load the same grammar files that ACE is loading.
>
> I suspect this situation is what Michael would like to avoid when he proposed the http://moin.delph-in.net/VirtualSharedConfigs discussion. So far, I was considering that making logon and ACE pointing to the terg trunk would be enough, now I am realising that I wasn’t paying attention to the configurations.
>
> I hope Dan is reading this thread!! ;-)
>
> Maybe a easier solution would be to use the last stable release of ERG where lkb/script and ace/config.tdl should be compatible. But my LOGON/lingo/erg/Version.lsp has `(defparameter *grammar-version* "ERG (1214)”)`. The LOGON/lingo/terg/Version.lisp has `(defparameter *grammar-version* "ERG (trunk)”)`. How to make LOGON use ERG 2018 instead of 1214?
>
> > greetings from the road (metaphorically), oe
>
> Thank you.
>
> Best,
> Alexandre
>



More information about the developers mailing list