I built a similar service a while back, with a small modification to the common algorithm.
You can improve contextual summarization by splitting the x sentences into x/n buckets. Then based on the percent of article to be summarized (eg return 60% of the article), pick the sentences ranked in the top 60% of each bucket. Then do this for all the x sentences, ie top 60% across buckets, and combine them together.
This prevents the bias rising from picking a sentence with a lot of critical words.
You can improve contextual summarization by splitting the x sentences into x/n buckets. Then based on the percent of article to be summarized (eg return 60% of the article), pick the sentences ranked in the top 60% of each bucket. Then do this for all the x sentences, ie top 60% across buckets, and combine them together.
This prevents the bias rising from picking a sentence with a lot of critical words.