“Generalized Transition Systems for Natural Language Analysis“
Senior Staff Research Scientist at Google
Date & Time: Monday, 17th October, 15:00
“Athena” Research & Innovation Centre, Institute for Language & Speech Processing,
Transition-based parsing has become a dominant paradigm in natural language processing (Nivre 2008). Primarily for morphosyntactic analysis, but also for a variety of other tasks from language generation to information extraction. Their popularity stems from their simplicity, efficiency, efficacy and reduction to multi-class classification that makes extensions to deep learning natural. Because of this, a number of transition systems have been proposed, covering different language phenomena, but also trading-off speed versus accuracy. In this talk, I will show how the majority of these systems can be described by a generalized transition system where a small set of control parameters dictates unique instantiations. Critically, this generalization allows us to explore previously unstudied systems. For example, we define a bounded capacity easy-first system that produces state-of-the-art results, is more efficient than the unconstrained variant and can be motivated by how humans process sentences.
Joint work with Bernd Bohnet, Ji Ma and Emily Pitler
Ryan McDonald is a Senior Staff Research Scientist at Google. Currently, he is the head of the London Natural Language Understanding team and more broadly the London Research & Machine Intelligence group. Previously he was the head of the New York Natural Language Understanding team at Google. His work focuses on multilingual morphosyntactic analysis and its applications to technologies like search, dialogue understanding and machine translation. Prior to Google he gained his PhD from the University of Pennsylvania on graph-based dependency parsing.