[developers] A question about *gen-filtering-p*
oe at ifi.uio.no
Tue Jan 6 01:06:20 CET 2009
hi francis, my apologies for a tardy reply!
> Can anyone explain what *gen-filtering-p* is meant to do?
that global is a boolean flag, activating what Carroll & Oepen (2005)
discuss under the name index accessibility filtering. in a nutshell,
it (potentially greatly) improves generator efficiency by pruning from
the chart edges that are judged semantically incomplete. in principle,
that test will judge edges incomplete that (a) do not contain a SKOLEM
constant corresponding to some MRS variable v (anywhere in the FS), for
which (b) there exists an EP (in the input semantics) referencing v but
not included in the EP coverage of that edge, and (c) are derived from
an edge that still contained (the SKOLEM constant for) v.
i believe berthold observed in the past that there can be surprises in
the interaction with packing (and FS restriction), particularly if the
grammar is (more) creative (than the ERG) in how it percolates indexes
(i.e. FS sub-structures corresponding to MRS variables).
turning this flag off should restore correctness in generation, at the
cost of some efficiency. how bad is it? alternatively, if you disable
packing, does the problem go away?
it is possible to make the generator verbose (*gen-filtering-debug*),
to first determine which of the edges required for the correct result
get filtered. next, you might have to find-gen-edge-given-id(), so as
to inspect the feature structure of the `lost' edge (and maybe the one
it took as input), and work out what happens with SKOLEM constants.
> With the very latest jacy in svn, [...]
is that SVN as in the LOGON tree? if so, with a stellar recipe like
the one you provided, i could maybe give the debugging a shot another
night. though, with in-depth knowledge of the grammar, it might be
easier for you actually to do the last part, the `working out' step.
all best - oe
+++ Universitetet i Oslo (IFI); Boks 1080 Blindern; 0316 Oslo; (+47) 2284 0125
+++ CSLI Stanford; Ventura Hall; Stanford, CA 94305; (+1 650) 723 0515
+++ --- oe at ifi.uio.no; oe at csli.stanford.edu; stephan at oepen.net ---
More information about the developers