Paraphrasing for Style

Proceedings of COLING 2012: Technical Papers, pages 2899–2914, COLING 2012, Mumbai, December 2012. Paraphrasing for Style Wei Xu1 Alan Ritter2 William B. Dolan3 Ralph Grishman1 Colin Cherr y4 (1) New York University (2) University of Washington (3) Microsoft Research (4) National Research Council, Canada xuwei@cims.nyu.edu, aritter@cs.washington.edu, billdol@microsoft.com, grishman@cs.nyu.edu, Colin.Cherry@nrc-cnrc.gc.ca ABSTRACT We present initial investigation into the task of paraphrasing language while targeting a particular writing style. The plays of William Shakespeare and their modern translations are used as a testbed for evaluating paraphrase systems targeting a specific style of writing. We show that even with a relatively small amount of parallel training data, it is possible to learn paraphrase models which capture stylistic phenomena, and these models outperform baselines based on dictionaries and out-of-domain parallel text. In addition we present an initial investigation into automatic evaluation metrics for paraphrasing writing style. To the best of our knowledge this is the first work to investigate the task of paraphrasing text with the goal of targeting a specific style of writing. KEYWORDS: Paraphrase, Writing Style. 2899 1 Introduction The same meaning can be expressed or paraphrased in many different ways; automatically detecting or generating different expressions with the same meaning is fundamental to many natural language understanding tasks(Giampiccolo et al., 2007), so much previous work has investigated methods for automatic paraphrasing(Madnani and Dorr, 2010). Paraphrases can differ along many dimensions, including utterance length, diction level, and speech register. There is a significant literature in sentence compression aimed at modeling the first of these, length: producing meaning-preserving alternations that reduce the length of the input string (Chandrasekar et al., 1996; Vanderwende et al., 2007; Clarke and Lapata, 2008; Cohn and Lapata, 2009; Yatskar et al., 2010). However, we know of no previous work aimed at modeling meaning-preserving transformations that systematically transform the register or style of an input string. Can we learn to reliably map from one form of language to another, transforming formal prose into a more colloquial form, or a casual email into a more formal equivalent? Systems capable of paraphrasing text targeting a specific writing style could be useful for a variety of applications. For example, they could: 1. Help authors of technical documents to adhere to appropriate stylistic guidelines. 2. Enable non-experts to better consume technical information, for example by translating legalese or medical jargon into nontechnical English. 3. Benefit educational applications, allowing students to: (a) Access modern English versions of works by authors they are studying. (b) Experiment with writing in the style of an author they are studying. In this paper, we investigate the task of automatic paraphrasing while targeting a particular writing style, focusing specifically on the style of Early Modern English employed by William Shakespeare. We explored several different methods, all of which rely on techniques from phrase-based MT, but which were trained on different types of parallel monolingual data. The first system was trained on the text of Shakespeare’s plays, along with parallel modern English “translations” that were written to help students better understand …

Content truncated to 500 words