In automatic summarization, interest is shifting ever more from bigram-based extractive techniques towards text-to-text and fully abstractive methods. A major challenge in the field is to determine what kind of semantic representations is rich enough to support useful inferences yet easily learnable from data. In this talk, I discuss how distributional semantics (DS) can help to determine the content and form of summaries in support of these goals. First, I present a method that integrates phrasal DS representations into a probabilistic content model in order to learn about the important events and slots in a domain. This method aims to use automatically learned domain knowledge to inform content selection, and is validated by an extractive summarization experiment. Then, I discuss the use of DS within a text-to-text sentence fusion setting, in which parts of multiple source text sentences are fused together to form a novel output sentence. This work points to the potential of DS to support broad-coverage natural language generation applications.