This dissertation explores improvements to reasoning from natural language through three distinct lines of investigation. Our contributions include an empirical investigation of the problem of sample selection for self-training syntactic parsers, the development of the tree-to-tree transduction system TTT, and expanded knowledge representation for first-order probabilistic commonsense reasoning. We find that constraining the automatically labeled data by sentence length and vocabulary novelty has a strong but noisy effect on the overall accuracy of self-trained models. Applying support vector regression using these features led us to obtain a new state of the art F1 measure of 92.52 on the standard WSJ subset of the Penn Treebank. The TTT language developed herein has proved to be a useful tool for manipulating both treebank parses and logical forms. The language borrows ideas from regular expressions and formal tree transducers. It is Turing complete and therefore able to compactly express a wide variety of transformations. Finally, the extensions we develop for symbolic logic are built upon set theory, causality, and an algebraic probability framework with well-defined algorithms for lifted inference. We validate the core techniques of our approach by showing favorable performance with respect to previous work in a relational logic domain. By virtue of these contributions we have pushed the field of artificial intelligence toward better reasoning from natural language.
Co-advisors: Professor Lenhart Schubert and Professor Daniel Gildea
Committee: Professor Gregory Carlson, Professor Daniel Stefankovic