Abstract:

Open-domain question answering (QA) requires a system to automatically find the answer for any natural language question. It is not only a core element of a range of real-world applications, but also a fundamental benchmark for natural language understanding. In this tak, I will first describe recent advances in open-domain QA systems that retrieve evidence text from a large collection of text. We introduce two orthogonal methods that overcome the limitation of previous, text-matching approach: an approach that constructs a graph of passages and models cross-passage interaction, and a dual encoder-based approach that builds expressive dense representations of passages. Second, I will describe a new class of the QA task – answering ambiguous questions. Previous work only focused on questions with a definite answer, ignoring intrinsic ambiguity that arises frequently in the questions written during information gathering. We propose a task that requires a system to answer a potentially ambiguous question by finding a set of all answers and disambiguating them. Together with new tasks and new models, we are making progress on significantly broadening the scope of questions that can automatically be answered.