Abstract

This work showcases the Reference-based Weak Supervision (RWS), a fully automatic large-scale data pipeline that harvests high-quality weakly- supervised answer sentences from Web data, only requiring a question-reference pair as input.