Extending centering theory for the measure of entity coherence

This thesis extends Centering Theory by tracking all entities in an utterance to improve the measure of a text's coherence. This also eliminates the compromises in choosing between a sentence or a clause as the unit of analysis. Experiments show that tracking all entities improves the evaluation of coherence compared to Centering Theory's measures of coherence. The proposed coherence model does not require training, alleviating the need for costly training data. The model can be used to identify coherence gaps and propose candidate sentences to bridge such gaps. Our results approach the state-of-the-art in a sentence ordering experiment which identifies the original text amongst a collection of its permutations. We perform worse than the state-of-the-art in a summary coherence experiment that ranks output from automatic summarization systems such that the rankings correlated with human rankings. These results illustrate the need to incorporate lexical cohesion for a more complete coherence model.