"Learning To Search" Methods In Automatic Language Processing: A State Of The Art

Structured prediction lies at the heart of modern Natural language Processing (NLP). In this paper, we study a specific family of structured learning algorithms, loosely refered to as "Learning-to-search" algorithms. They differ in several important ways from more studied methods such as Conditional Random Fields, and their study highlights several important tradeoffs of structured learning for NLP. We also present an overview of existing applications of these techniques to NLP problems and discuss their potential benefits.