System Combination Architectures

Hans van Halteren, Radboud University Nijmegen, The Netherlands

When building a system that processes natural language, except in the most trivial manner, we are faced with the problem that we do not fully understand natural language. Of course, we can deal with the problem in various ways, and this means we tend to use very diverse approaches, e.g. using rules created by human experts or using machine learning from data sets. Experience shows that, for practically any NLP task, every approach manages to do part of the job well (sometimes really well), and manages to fail at other parts (sometimes really spectacularly). So, in practice, we are faced with a performance ceiling.

However, a simple trick lets us go beyond the ceiling of even the best approach. If we do not restrict ourselves to a single approach, but rather combine the opinions of a number of approaches, we can push the ceiling higher. How much higher depends mostly on the dissimilarity of the component approaches and on the architecture used for the combination.

In this course, you will learn the basic ideas behind system combination, component selection and combination architectures. Each day, a new subject is discussed, after which you yourself will run some experiments to gain hands-on experience. In order to make the course even more directly useful for your own research, you are encouraged to bring your own data sets for these experiments.