Lexicon Acquisition with and for Symbolic NLP-Systems – a Bootstrapping Approach

We present a method of applying a broad-coverage LFG grammar of German in the process of semi-automatic lexicon acquisition from corpora. The identification of corpus instances that illustrate a certain subcategorization frame uniquely is done by a comparison of the numbers of analyses the grammar assigns to the corpus instances, under the assumption of different hypothetical lexicon entries for the candidate verb. Filtering conditions expressed on the feature representation output by the grammar further restrict the sentences that the automatic extraction step is based on. Experiments show that the grammar-based method produces better results than a method based on patterns in a corpus query language.