Linguistically Motivated Complementizer Choice in Surface Realization

This paper shows that using linguistically motivated features for English that-complementizer choice in an averaged perceptron model for classification can improve upon the prediction accuracy of a state-of-the-art realization ranking model. We report results on a binary classification task for predicting the presence/absence of a that-complementizer using features adapted from Jaeger's (2010) investigation of the uniform information density principle in the context of that-mentioning. Our experiments confirm the efficacy of the features based on Jaeger's work, including information density--based features. The experiments also show that the improvements in prediction accuracy apply to cases in which the presence of a that-complementizer arguably makes a substantial difference to fluency or intelligiblity. Our ultimate goal is to improve the performance of a ranking model for surface realization, and to this end we conclude with a discussion of how we plan to combine the local complementizer-choice features with those in the global ranking model.