Nutri-bullets: Summarizing Health Studies by Composing Segments

We introduce Nutri-bullets, a multi-document summarization task for health and nutrition. First, we present two datasets of food summaries from multiple scientific studies. Furthermore, propose novel extract-compose model to solve the problem in regime limited parallel data. explicitly select key spans several abstracts using policy network, followed by composing selected summary via specific language model. Compared state-of-the-art methods, our approach leads more faithful, relevant diverse -- properties imperative this application. For instance, on BreastCancer dataset gets than 50% improvement relevance faithfulness.