Abstract: Automated reasoning, the ability of computing systems to make new inferences from observed evidence, has been a long-standing goal of artificial intelligence. Knowledge bases (KBs), both automatically and manually constructed, are often incomplete. However, many valid unobserved facts can be inferred from observed facts by reasoning over the KB. We are interested in automated reasoning on large KBs with rich and diverse semantic types. An effective and user-friendly way of accessing the information stored in a KB is by issuing queries to it. Such queries can be structured (e.g. queries for booking flights) or unstructured (e.g. natural language queries). A challenge for question answering (QA) systems over KBs, is to handle queries whose answers are not directly stored (as a simple fact) in the KB. Instead, the QA model needs to reason in order to derive the answer from other observed facts. This thesis focuses on building QA systems over structured KBs that can perform such reasoning.
Recent advancements in representation learning by deep neural models have resulted in tremendous progress in the performance of QA systems. However, such deep models function as black-boxes with an opaque reasoning process, are brittle, and offer very limited control for debugging a wrong prediction. It is also unclear how to reliably add or update knowledge stored in their model parameters.
This thesis proposes nonparametric models for question answering that disentangle logic from knowledge. For a given query, such models are capable of deriving interpretable reasoning patterns "on-the-fly" from other contextually similar queries in the training set. The reasoning patterns can be chains (i.e. sequence of relations) or subgraphs with complex structures containing multiple relation types. We show that our models can seamlessly handle and reason with new knowledge as they are continuously added to the knowledge base. We also show that the same reasoning paradigm is effective for complex natural language queries where the logical form is derived from the logical forms of multiple natural language queries, achieving new state-of-the-art results in various knowledge base question answering and completion benchmarks. Lastly, leveraging our nonparametric approach, we demonstrate that it is possible to correct wrong predictions of deep QA models without any need for re-training, thus paving the way towards building practical production-ready QA systems.
Advisor: Andrew McCallum