|
Grammar-based techniques for efficient surface generation with an HPSG for Germanold_uid | 10717 |
---|
title | Grammar-based techniques for efficient surface generation with an HPSG for German |
---|
start_date | 2012/01/20 |
---|
schedule | 11h-13h |
---|
online | no |
---|
summary | In this talk I will present recent results on increasing generation
efficiency with GG, a large-scale HPSG for German.
Chart parsing with HPSG grammars such as the ERG has proven to be highly
efficient in practice, owing to a combination of different performance
measures, such as ambiguity factoring (packing), index filtering, and
selective unpacking (Oepen & Carroll 2005).
Generation efficiency with GG, however, has remained unsatisfactory
until very recently, arguably because of an increased level of local
ambiguity introduced by the non-configurational nature of German syntax:
in particular, variable verb placement, discontinuous verb clusters
(with local underspecification of valence information), free permutation
of arguments, free interspersal of modifiers, as well as case morphology
all contribute to a massively increased search space.
I shall present an overview of the current efficiency measures in HPSG
generation, followed by an in-depth analysis of the constructional
factors impeding efficient generation in German. In order to contain
some of the added local ambiguity, I suggest use a transfer grammar in
order to augment the input semantics with syntactic information that
will provide for a more informed search during the chart generation
process, while at the same time keeping the external semantic interface
to the generator free from syntactic information. I shall discuss the
individual syntactic features used and present some preliminary results
on generation efficiency. |
---|
responsibles | Candito |
---|
| |
|