As most of human knowledge is recorded in text data, building machines that can effectively extract and reason with the knowledge buried in human language is an essential and long-standing goal of Artificial Intelligence. This line of research also has lots of practical use cases as enormous text data are emerging on the Internet at any given moment. While traditional information extraction and query systems usually depend on human-annotated knowledge bases, recent breakthroughs in natural language processing (e.g., advanced attention mechanism and large-scale pre-trained language models) has shown the possibility of building this kind of systems directly over text data.
In this talk, I will first discuss the techniques used for querying and reasoning over large-scale knowledge graphs. This will include semantic parsing, graph attention networks, and reinforcement learning. I will then cover the state-of-the-art reading comprehension techniques (e.g., Bidirectional Attention Flow, BERT) that can be used to extract knowledge from natural language articles and answer human-written questions, and introduce more recent research on multi-hop question answering. I will finally discuss the possibility of getting the best of these two directions and developing more efficient and effective knowledge-aware natural language understanding models.