Evaluation of Centering Theory with the GNOME Corpus

Centering theory (Grosz, Joshi and Weinstein, 1986, 1995; Walker et al, 1998) is a theory of local discourse coherence and salience that has proven useful to study language phenomena such as pronominalization, and as the basis for algorithms for anaphora interpretation and natural language generation for languages including English, German, Italian, Japanese, and Turkish. The theory's claims are formulated in terms of notions such as utterance, ranking, and realization, that have been argued to be defined differently for different languages - i.e., to be parameters of the theory. 

This page allows you to examine the effect of different instantiations of these parameters, using the methods discussed  in (Poesio, Stevenson, Di Eugenio, and Hitzeman, 2004), using the GNOME corpus.  The  extended version of the Computational Linguistics paper available here discusses methodology and experiments  in more detail,   and  explains how to set the parameters in order to replicate the results. 

The form next page displays all the parameters of Centering theory that you can vary from the website. After choosing the instantiation  you are interested in, you can start the script that will compute a number of statistics concerning that configuration; the results are displayed in a new page. The link at the bottom of the results page allow you to see the results for each file of the corpus.

Massimo Poesio

click here to proceed to the form