[developers] bi-directional subsumption test
oe at csli.Stanford.EDU
Mon Aug 8 11:54:04 CEST 2005
as i am obviously approaching senility, let me use email for permanent
storage. working with woodley last night, we re-discovered a problem
in the bi-directional subsumption test that ulrich and i had addressed
in PET only a few years ago:
Change 601 by oe at oe on 2003/06/01 02:19:17
Finally! Fix bug in subsumption test related to subgraph sharing: since
we may end up having the same dag in both input structures (which cannot
happen for unification during parsing, as the chart adjacency condition
rules out combinations of edges with joint daughters). Using the copy
slot to record the `corresponding' node for re-entrant nodes during
subsumption in both directions, means we may hallucinate coreferences
when we effectively visit the same dag as part of both input dags.
The fix was identified on the rooftop of 601 4th Street: use separate
slots for forward and backward correspondents; assume that when we call
the subsumption test it is save to use the forward slot.
Affected files ...
... //pet/main/cheap/dag-tomabechi.cpp#15 edit
==== //pet/main/cheap/dag-tomabechi.cpp#15 (text) ====
< dag_node *c1 = dag_get_copy(dag1),
> dag_node *c1 = dag_get_forward(dag1),
< dag_set_copy(dag1, dag2);
> dag_set_forward(dag1, dag2);
to my dismay i see no records of emailing people about this earlier. i
would suggest making the same change in the LKB now.
what do you think? - oe
+++ Universitetet i Oslo (ILN); 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 ---
More information about the developers