Using Repeated Patterns across Comparable Articles for Paraphrase Acquisition

We focus on paraphrases for information extraction: expressions which should produce the same extraction output. These expressions are acquired automatically from comparable news articles (articles from the same day, on the same topic). Candidate paraphrases are paths in predicate argument structure starting from matching anchors (typically, names) in the two sentences. By using such syntactically-regularized structures and limiting ourselves to single paths, we increased the likelihood of observing repeated patterns. We measured the frequency of such candidate patterns over a large corpus, and confirmed a correlation between frequency and their accuracy as paraphrases.