Learning to Fuse Disparate Sentences

The work presented here is based on a paper under submission with Micha Elsner. This thesis presents a system for fusing sentences which are drawn from the same souce document but differ significantly in content. The system described here is supervised, trained from real world examples of sentences fused by professional journalists in the process of editing news articles. The system is similar to previous work in that, it merges dependency graphs using Integer Linear Programming. Instead of aligning the inputs as a preprocess, it integrates the task of finding an alignment along with the selecting a merged sentence into a single joint optimization problem for which the parameters are learned using a structured online learning algorithm. Human evaluation shows that the technique produces fused sentences which are both informative and highly readable.