... by presenting a novel data collectionframework that produces highly parallel text data relatively inexpensively and on a largescale. The highly parallel nature of this data allows us to use ... significantamount of translation data, unique in its multilingualparallelism. While included in our data release, weleave aside a full discussion of this multilingual data for future work.192To ... bilingual data to sup-port paraphrase extraction. In contrast, our approachonly requires monolingual data, and evaluation canbe performed using arbitrarily small, highly -parallel datasets....