Query-focused summarization by supervised sentence ranking and skewed word distributions

We present a supervised sentence ranking approach for use in extractive summarization. The supervised approach achieves domain independence by making use of a range of word distribution statistics as features, of the sort typically used for unsupervised domain-independent ranking. We present empirical trials on the DUC 2006 query-directed multi-document summarization task, and demonstrate that the very general machine learning approaches taken can provide competitive results for this task. The general approach provides great flexibility for incorporating many more features.