Title: Information Status in Generation Ranking

In this talk, we will present work on surface realisation ranking for German, 
i.e. the task of choosing the most natural string, given a set of possible 
realisations for a given input.

It is well established that information structure plays an important role in the 
choice of surface realisation uttered by a speaker. We present our work on ways
to integrate information status (an aspect of information structure that 
classifies nominal constructions in a semantically motivated way according to
whether they are given or new) into the process.

We first present our findings on the asymmetries that hold between the different
types of information status (IS) and show how they can be incorporated into the
realisation ranking process by means of morpho-syntactic approximations. We then
present current research on automatically labelling nominal structures with IS
labels. We will present encouraging results, and compare the performance of our
statistical model with human annotators. We will then outline a number of
experiments to show what effect automatically labelled text has on the
realisation ranking process.