You need to parse the text into units that make sense to humans - phrases, sentences and paragraphs. Next score those units according to the presence of keyword(s), now select the best of the units that are "hits" according to typical writing principles and the size of the summary you are aming for.
What do I mean about writing principles? Think about how you yourself scan text.
For example you expect the first sentence of a paragraph to be meaningful in terms of the content of that paragraph. You expect a good chance that the last paragraph of an article to summarize the article.
In the prehistoric era of computers (showing my age now) there was an indexing technique called KWIC - Key
Word In Context. It created a listing with the n words preceeding a key word plus the n words following. This put a burden on the reader to recognize a significant context versus a trival one.
This is a topic of continued interest to me, let us know what you come up with.
Bill
