We investigate a novel framework for Automatic Text Summarization. In this framework underlying language-use features are learned from a minimal sample corpus. We argue the low complexity of this kind of features allows relying in generalization ability of a learning machine, rather than in diverse human-abstracted summaries. In this way, our method reliably estimates a relevance measure for predicting summary candidature scores, regardless topics in unseen documents. Our output summaries are comparable to the state-of-the-art. Thus we show that in order to extract meaning summaries, it is not crucial what is being said; but rather how it is being said.