Improving Ellipsis Resolution with Transformation-Based Learning

We describe an experiment in using Transformation-Based Learning to improve the output of a VP ellipsis resolution system. In this preliminary study, we address system errors in which the correct VP antecedent was chosen, but in which right-peripheral material should either be added or removed. The system was trained on 256 examples of VP ellipsis from the Brown Corpus, and was tested on 128 examples. While the small size of the data sets did not lead to conclusive empirical results, many of the patterns learned appear to represent correct generalizations, and we are currently engaged in experiments with larger data sets.