[developers] LKB vs. PET divergences

Stephan Oepen oe at csli.Stanford.EDU
Mon Feb 14 09:03:43 CET 2005

hi again,

francis is still visiting, cracking the whip.  so we sat down to debug
an unpleasant problem: the LKB and PET producing different results for
some inputs.  after many hours staring at feature structures, it seems
the problem originates in different `interpretations' of one rule from
the grammar, viz. the `vstem-vstem-rule'.  i attach the AVMs for just
the rule below (`dag.lkb' vs. `dag.pet'), so you will quickly see that
the LKB has

  --> [ ], [ SYNSEM.LOCAL.CONT.HOOK #0 ]

whereas PET lacks the coreference with the HOOK of the second daughter.
this would explain why PET sometimes finds additional derivations, i.e.
ones that later fail to re-build in the LKB when treebanking.

looking at the definition of this rule in `japgram.tdl' and walking up
the type hierarchy a little, we fail to see why the coreference should
be there (the one to the second daughter, that is).  but given general
exhaustion and the power of well-formed unification, we feel not quite
certain of our analysis (having to assume that either the LKB or PET
could be incorrect was hard enough :-).

hence, we are hoping experts in the grammar or formalism could give it
a critical look over-night?  i attach the full grammar and the two DAG
representations for the rule.  how many re-entrancies for HOOK should
we expect to see in a fully expanded, well-formed `vstem-vstem-rule'?

looking at the definition of the rule, i also noticed it uses multiple
supertypes in the instance definition.  this is generally illegal, but
happens to succeed in this case, because for:

  vstem-vstem-rule := head-final-type & orth-princ & [...]

there appears to be a glb in the type hierarchy.  both systems seem to
compile the above without complaint, but i believe the above situation
should be flagged or outright rejected as syntactically ill-formed.

                                                       cheers  -  oe

+++ Universitetet i Oslo (ILF); Boks 1102 Blindern; 0317 Oslo; (+47) 2285 7989
+++     CSLI Stanford; Ventura Hall; Stanford, CA 94305; (+1 650) 723 0515
+++       --- oe at csli.stanford.edu; oe at hf.uio.no; stephan at oepen.net ---

-------------- next part --------------
A non-text attachment was scrubbed...
Name: dag.lkb
Type: application/octet-stream
Size: 4934 bytes
Desc: not available
URL: <http://lists.delph-in.net/archives/developers/attachments/20050214/0249d7cb/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dag.pet
Type: application/octet-stream
Size: 5015 bytes
Desc: not available
URL: <http://lists.delph-in.net/archives/developers/attachments/20050214/0249d7cb/attachment-0001.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: jacy.feb-04.tgz
Type: application/x-gtar
Size: 1615134 bytes
Desc: not available
URL: <http://lists.delph-in.net/archives/developers/attachments/20050214/0249d7cb/attachment.gtar>

More information about the developers mailing list