Predicting Structured Outputs in NLP
Speaker: Hal Daumé III
Institution: University of Utah
Abstract:
Machine learning technologies have historically focused on learning the relationships between complex inputs and simple outputs, for instance images to binary labels. There is comparatively less work on learning the relationships between complex inputs and complex outputs, for instance Arabic sentences to English sentences. This talk will discuss methods for learning to predict complex outputs. The introduction will focus on exact methods, for which the output space is reasonably simple (eg., sequences or trees), and quickly move on to the more common case in NLP where the output space is highly structured. Some basic knowledge of machine learning (for instance, binary classification) will be helpful, but is not necessary.
