Learning Contrastive Connectives in Sentence Realization Ranking

Abstract

We look at the average frequency of contrastive connectives in the SPaRKy Restaurant Corpus with respect to realization ratings by human judges. We implement a discriminative n-gram ranker to model these ratings and analyze the resulting n-gram weights to determine if our ranker learns this distribution. Surprisingly, our ranker learns to avoid contrastive connectives. We look at possible explanations for this distribution, and recommend improvements to both the generator and ranker of the sentence plans/realizations.