[lkb] Generator question
Emily M. Bender
ebender at u.washington.edu
Thu Mar 29 08:15:27 CEST 2007
Yes, thanks. We will reorganize our feature structures then.
Emily
On Wed, Mar 28, 2007 at 09:11:16PM -0800, Stephan Oepen wrote:
> hi emily,
>
> > This seems to be a problem for the generator: So long as cog-st is a
> > sort (no features), the cog-st value is respected in generation. As
> > soon as we add the feature SPECI, however, the generator will ignore
> > the type of the COG-ST value.
>
> i think your problem is strictly speaking not about the generator. it
> is important to acknowledge the difference between the grammar-internal
> universe of typed feature structures and the grammar-independent world
> of MRSs. MRSs are not typed in a very interesting sense, and they have
> very little accomodation for embedding of structures (i.e. recursion).
>
> a typed feature structure
>
> ref-index & [ COG-ST foo ]
>
> may correspond to an MRS variable, say,
>
> x42 { COG-ST: foo }
>
> for convenience, MRS variables can have internal properties, but their
> values are constrained to be atomic. hence,
>
> ref-index & [ COG-ST foo & [ SPECI + ] ]
>
> converts into
>
> x42 { COG-ST.SPECI + }
>
> the process of constructing an MRS from its FS description, helpfully,
> flattens out embedded structures and creates property names from paths
> of feature names. hence, even though `COG-ST.SPECI' looks a lot like
> the TDL path notation, in the MRS universe it is a plain symbol, viz.
> the name of that property. there is no room for the sortal constraint
> `foo' from the midst of the underlying path in that representation.
>
> i hope this helps clarify the issue? 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 csli.stanford.edu; oe at ifi.uio.no; oepen at idi.ntnu.no ---
> +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
More information about the lkb
mailing list