Learning to Fuse Disparate Sentences

摘要

We present a system for fusing sentences which are drawn from the same source document but have different content. Unlike previous work, our approach is supervised, training on real-world examples of sentences fused by professional journalists in the process of editing news articles. Like Filippova and Strube (2008), our system merges dependency graphs using Integer Linear Programming. However, instead of aligning the inputs as a preprocess, we integrate the tasks of finding an alignment and selecting a merged sentence into a joint optimization problem, and learn parameters for this optimization using a structured online algorithm. Evaluation by human judges shows that our technique produces fused sentences that are both informative and readable.