Projects in Previous Years

2nd Semester 2010/11: Machine Learning of Compositional Semantics

Jelle Zuidema
If you are interested in this project, please contact Jelle by email.
For both formal semanticists and computational linguists a long standing question is: How can we induce -- from corpora of plain text, sentence-meaning pairs or question-answer pairs -- the formal system needed to interpret sentences never seen before? In the last few years some important advances have been made on this issue in the computational linguistics community. In this project, targeted at students with some background in formal semantics, we introduce those developments and try to evaluate them from different perspectives. The project starts with one or more lectures (depending on your backgrounds) on the essentials of statistical inference (e.g., EM) and probabilistic grammars (e.g., PCFG, STSG, DMV). We continue with a computational exercise (running EM on DMV; no programming required), and then proceed with discussing the various papers on the topic by Klein & Manning, Zettlemoyer & Collins, Zelle & Mooney and Liang, Jordan & Klein. The most recent paper in this series contains the best results so-far anand many of the relevant citations: