Learning Paraphrase Models from Google New Headlines

Data sources like the clusters of news headlines at Google News present an exciting opportunity to learn paraphrase models from data automatically. We present both a novel dataset and a novel approach to automatic, unsupervised learning of paraphrase models from that datset. Leveraging existing NLP tools such as the Stanford Parser and lexical resources such as WordNet and Infomap, we constructed a system that first aligns the typed dependency graphs of large numbers of parallel headlines (on the order of hundreds) and then uses aligned paths between corresponding nodes as candidates to a paraphrase extraction system. We present some preliminary results in the form of actual learned paraphrase models. This project serves as a proof of concept for this approach and sheds some light on likely next steps.