Over the past several years there has been a resurgence of interest in robust semantic analysis in computational linguistics. This resurgence has been made possible, in large part, by the availability of large collections of semantically annotated texts such as FrameNet and Propbank which can serve as training materials for supervised machine learning techniques. I will first present the results of our efforts to build semantic analysis systems along these lines for English, Chinese and Arabic, and then discuss some of the fundamental challenges of applying supervised machine learning techniques in this domain.