Automatic acquisition of lfg resources for german - as good as it gets

Traditionally, deep, wide-coverage linguistic resources are hand-crafted and their creation is time-consuming and costly. Considerable effort has been put into overcoming this problem by automatically inducing linguistic resources such as rich, deep grammars, lexicons or subcategorisation frames from corpora. Most work so far has concentrated on English, like the one of Hockenmaier and Steedman [7], Nakanishi et al. [8] or Cahill et al. [3]. They present approaches for the acquisition of deep linguistic resources from the Penn-II treebank, using different grammar frame-works like CCG, HPSG and LFG. English, however, is a configurational language, where strict word-order constraints help to disambiguate predicate-argument structure. Porting these approaches to a semi-free word order language, the question arises: how good can it get? Can we expect similar results when dealing with (semi-)free word order? Can data-driven methods cope when dealing with ambiguous data structures and sparse data? And, furthermore, what impact has treebank design on the automatic acquisition of linguistic resources such as deep grammars? This paper describes approaches to treebank-based acquisition of LFG resources for a semi-free word order lan-guage, based on and substantially extending the method of Cahill et al., O’Donovan et al. and Burke et al. [2, 9, 1], who presented large-scale acquisition of LFG grammars and lexical resources from the English Penn-II and Penn-III treebanks. They also presented work on data-driven multilingual unification grammar development for Spanish, Chi-nese and German. While treebank-based deep grammar acquisition can be applied to other languages, results are lowe