Predicting corpus example quality via supervised machine learning

In this paper we present a supervised-learning approach to extracting good dictionary examples from corpora.We train our predictor of quality on a dataset of corpus examples annotated with a four-level ordinal variable, ranging from a very bad to a very good example. Each of the examples is formally described through 23 variables; the dependence of the quality of which is modelled using a regression model. The evaluation of the ranked results for each of the collocations in the annotated dataset shows that we obtain precision on 10 top-ranked examples of ~80% and a precision of ~90% on the three top-ranked examples. Our approach is highly language independent as well, suffering almost no loss on the 10 top-ranked examples and a loss of ~4% on the three highest-ranked examples once the language-dependent and knowledge-source-dependent features are removed.