Wouldn't it be great if your computer could do all of that for you: gather the right sources (e.g. paragraphs from relevant Wikipedia pages), synthetize the information, and write up an easy-to-read, original summary of the relevant points? Such a system isn't quite available yet, at least not one that can provide reliable information in its summary. Even though current systems excel at finding an extractive span that answers a factoid question in a given document, they still find open-domain settings where a model needs to find its own sources of information and long answer generation challenging.

Thankfully, a number of recent advances in natural language understanding and generation have made working toward solving this problem much easier! These advances include progress in the pre-training (e.g. BART, T5) and evaluation (e.g. for factuality) of sequence-to-sequence models for conditional text generation, new ways to use language understanding models to find information in Wikipedia (e.g. REALM, DPR), and a new training dataset introduced in the paper ELI5: Long Form Question Answering.