In my opinion, it depends on the application. For some reason, in Computer Science we are always trying to find "the right number", but we need to learn that there is not a one-size-fits-all solution for everything. Before setting a number at what "feels right", you have to ask yourself why are you doing this for? who will read the summaries? What is the size of the original texts? How much information can you afford to lose? For example, if the person that will read the summary is Donald Trump, and you have to summarize a 1,000 sentences long report, then you will need like a 2% compression rate. But if your target audience is a college student that doesn't have enough time to read a whole novel, maybe 30% is reasonable. My advice here (and with any CS problem) is: always know your application, know your audience and know your data, and then start making decisions.