Video Recordings

The linguistics researcher or language technologist often wonders “what corpus should I use, or should I build one of my own? If I build one of my own, how will I know if I have done a good job?” Currently there is very little help available for them. They are in need of a framework for evaluating corpora. We develop such a framework, in relation to corpora which aim for good coverage of ‘general language’. The task we set is automatic creation of a publication-quality collocations dictionary. For a sample of 100 headwords of  Czech and 100 of English, we identify a gold standard dataset of (ideally) all the collocations that should appear for these headwords in a collocations dictionary. We then use them to determine precision and recall for a range of  corpora, with a range of parameters.