Standard clustering algorithms (found in any off-the-shelf natural text processing library) and text summation with libots should suffice for most of the heavy lifting.
http://tldr.it/ http://libots.sourceforge.net/
Further, most news articles' first paragraph is a practical (although you may have not noticed) summary.
Coming from NLP, unless you can influence the source and the source being Web, the story should be an 80%-20% in the best case -- and you'll work VERY hard to correct the remaining 20%, and you WILL remain with a percentage of content you just can't summarize properly.
What would make a difference is a real people-driven summation, not machines (see what voicebunny did for text-to-speech, for example). And yes, it would have been fun to combine the two as well.