Towards knowledge-based, robust question answering

Abstract

Current question answering systems face two major challenges; the ability to employ external knowledge and to robustly generalize to unseen expressions of questions need to be improved. In this thesis, I introduce two works that can together help advance question answering. First, I introduce TransINT, a novel and interpretable knowledge graph embedding method that isomorphically preserves the implication ordering among relations in the embedding space. Second, I present methods to train sequence-to-sequence semantic parsing models robust to unseen paraphrases. These two works could together serve as steps to create human-like question answering systems that can understand unseen paraphrases and link existing and external facts for logical inference.