Learning from the past: Improving news summarization with past news articles

One common approach to single-document news summarization involves scoring and ranking individual sentences within an input story. We demonstrate that the accuracy of this scoring process can be improved by looking beyond the text found within each input news story. Leveraging on an external corpus of past news articles, we show that summarization performance can be greatly enhanced if we also consider signals and cues from other related news stories. Working on top of a basic keyword-based summarization system, we expanded the set of keywords we have from the original news stories with related stories retrieved from the external corpus. With this enhancement, we are able to get significant improvements of at least 10% and 16% in ROUGE-1 and ROUGE-2 respectively.