[developers] Valid MRS? Bug in ERG?

goodman.m.w at gmail.com goodman.m.w at gmail.com
Sun Sep 27 17:37:32 CEST 2020


On Sun, Sep 27, 2020 at 3:16 PM Stephan Oepen <oe at ifi.uio.no> wrote:

> [...]
> conceptually, i think all MRSs should be converted to EDSs, and no
> information that can be expressed in the EDS graph should be lost.
> more practically: each EP (and each ICONS) in the MRS should introduce
> a node into the EDS, and each semantic role whose value is associated
> with a node should yield an edge (additional edges should be
> introduced for instances of predicate modification).  additional or
> illformed information in the MRS (e.g. invalid handle constraints or
> roles whose value does not correspond to the label or intrinsic
> variable of an EP) should be ignored.
>

I agree with this, except the part about ICONS introducing nodes surprised
me.. I thought that EDS, like DMRS, is yet to provide a treatment for
ICONS. Care to explain?


>
> [...]
>
> mike. why would you think any of the above might call for
> grammar-specific solutions?  i would like to encourage pyDelphin to
> embrace the same robustness goals as the LKB-based converter.  EDSs
> were originally invented for practical utility, to make it easier for
> downstream applications to work with ERG analyses; for that goal, any
> MRS that comes out of the parser should also yield an EDS, no matter
> its structure or contents.


First, some clarifications/corrections: The EDS conversion error in
PyDelphin is a bug, not expected behavior. Also, the dropped nodes for the
disconnected DMRS was not completely described (sorry, Alexandre). The
conversion actually creates nodes even for the disconnected EPs, but I was
viewing the output of the dmrs-penman codec which is not capable of
representing disconnected graphs (the same goes for eds-penman). They are
present in other serialization formats.

Regarding the grammar-specific solutions, as I understand the LKB's EDS
code still maintains lists of ERG (1214?) predicates, roles, etc. for
various uses, such as predicate modification. In the case of predicate
modification, it and PyDelphin suture the unlinked nodes with the ARG1
role, if it was otherwise unused. I was under the impression that the LKB's
converter did a bit more surgery in order to normalize some anticipated
ill-formed structures. But if I was mistaken and the only other
"value-added" part of conversion is the aforementioned top-selection with
the other MRS-maladies ignored, then I think we share the same goal and
view of robustness.

-- 
-Michael Wayne Goodman
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20200927/8f83bedf/attachment.html>


More information about the developers mailing list