Open-Domain Question Answering from Large Text Collections