From the SVM output we collect per-topic as well as
per-sentence precision and recall to compute macro- (pertopic)
and micro-averaged (per-sentence) precision, recall
and F1 measures. We also compute the ROUGE measure
for the coverage of the automatic summary with respect to
the human model extracts. ROUGE is recall-oriented, based
on n-gram overlap, and correlates well with human evaluations
[6]. The DUC data set provides human model extracts
for summaries of length 200 and 400 words.