[developers] Characterisation in cheap

Berthold Crysmann crysmann at ifk.uni-bonn.de
Wed Aug 19 23:22:30 CEST 2009


On Wed, 2009-08-19 at 19:23 +0200, Stephan Oepen wrote:
> hi again,
> 
> > trying to add characterisation to HaG, I experienced some surprising
> > behaviour with cheap -t. 
> > 
> > Currently, there is no characterisation on GPREDs, regardless of whether
> > they are set within the grammar or not. LKB by contrast appears to
> > honour GPREDs set in the grammar.
> 
> what exactly do you mean by: set in the grammar or not?


> > To test with the ERG, try 
> > 
> > echo "John, Paul, and George  barked." | \
> > cheap -t -cm -nsolutions=1 -mrs  english
> > 
> > You'll find there is no characterisation on the implicit_conj_rel.

> once more, i think you are pointing out a genuine bug.  but this is a
> problem in the ERG, not in PET. 

Ok. With that information in hand I managed to hunt down the bug in HaG
as well. Still I am quite surprised that it worked as expected in the
LKB. 

BTW: why does the LKB count characters, but Pet tokens? ANd how do I
changed the behaviour in any of these platforms?

B

>  in the new universe, the assumption is
> that characterization is controlled by the grammar, i.e. there is FROM
> and TO information on the token FSs; lexical items copy these into an
> appropriate place in the signs (`ORTH.FROM' and `ORTH.TO' in the ERG);
> all rules make sure to percolate correct values along the periphery of
> phrases; and any sign introducing a new EP co-indexes CFROM and CTO on
> the PRED with its own FROM and TO value.  dan felt that it was not very
> difficult adding co-references to implement this strategy, as existing
> abstractions in the type hierarchy were suitable.  however, it appears
> there is a missing co-reference on the sign introducing `implicit_conj'
> here.
>  
> > Setting or unsetting MRS-CFROM-PATH etc. seems to have no effect.
> 
> the ERG deliberately does not set `mrs-cfrom-path' and `mrs-cto-path'.
> these activate the procedural characterization mechanism in PET, and i
> sincerely hope all code that destructively modifies feature structures
> `behind the scenes' in the parser is obsoleted with chart mapping.  in
> the not too distant future, we should try spring cleaning the PET code
> base for removal of such legacy mechanisms ...
> 
>                                                      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 mailing list