Computational Lessons From and For Language

Barely any task is more challenging and more effortlessly carried out by humans than the efficient use of language. Within a couple of years, children figure out a learning problem that even computers with large, extensively annotated training sets fail at. On a daily basis, we hear thousands of sentences and within split seconds assign them highly elaborate structures whose complexity matches and even exceeds what we find in protein folding or the specification of programming languages.  The conundrum of language is that all research and all our theorems point towards it being a very hard computational problem, yet for humans it is no more difficult than jumping on one leg.

In this talk, I present recent research that suggests ways of resolving this contradiction. The central lesson is that complex structures are simple structures in disguise. I illustrate this idea with concrete linguistic examples, and I show how the formal insights can be generalized beyond language. In the other direction, I also highlight open complexity issues in linguistics where supercomputation may prove indispensable for finding simpler solutions. These issues can be tackled with the tools and techniques that already exist today, we just have to step up and actually do it.