A New Approach to the Corpus-based Statistical Investigation of Hungarian Multi-word Lexemes

We apply statistical methods to perform automatic extraction of Hungarian collocations from corpora. Due to the complexity of Hungarian morphology, a complex resource preparation tool chain has been developed. This tool chain implements a reusable and, in principle, language independent framework. In the first part, the paper describes the tool chain itself, then, in the second part, an experiment using this framework. The experiment deals with the extraction of patterns from the corpus as collocation candidates, in order to compare results to an experiment on Dutch V + PP patterns (Villada, 2004). Statistical processing on this dataset provided interesting observations, briefly explained in the evaluation section. We conclude by providing a summary of further steps required to improve the extraction process. This is not restricted to improvements in the resource preparation for statistical processing, but a proposal to use nonstatistical means as well, thus acquiring an efficient blend of different methods.