Paraphrase fragment extraction from monolingual comparable corpora

We present a novel paraphrase fragment pair extraction method that uses a monolingual comparable corpus containing different arti-cles about the same topics or events. The pro-cedure consists of document pair extraction, sentence pair extraction, and fragment pair ex-traction. At each stage, we evaluate the in-termediate results manually, and tune the later stages accordingly. With this minimally su-pervised approach, we achieve 62 % of accu-racy on the paraphrase fragment pairs we col-lected and 67 % extracted from the MSR cor-pus. The results look promising, given the minimal supervision of the approach, which can be further scaled up.