For the sake of simplicity: Unsupervised extraction of lexical simplifications from Wikipedia

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 365–368, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics For the sake of simplicity: Unsupervised extraction of lexical simplifications from Wikipedia Mark Yatskar, Bo Pang, Cristian Danescu-Niculescu-Mizil and Lillian Lee my89@cornell.edu, bopang@yahoo-inc.com, cristian@cs.cornell.edu, llee@cs.cornell.edu Abstract We report on work in progress on extract- ing lexical simplifications (e.g., “collaborate” → “work together”), focusing on utilizing edit histories in Simple English Wikipedia for this task. We consider two main approaches: (1) deriving simplification probabilities via an edit model that accounts for a mixture of dif- ferent operations, and (2) using metadata to focus on edits that are more likely to be sim- plification operations. We find our methods to outperform a reasonable baseline and yield many high-quality lexical simplifications not included in an independently-created manu- ally prepared list. 1 Introduction Nothing is more simple than greatness; indeed, to be simple is to be great. —Emerson, Literary Ethics Style is an important aspect of information pre- sentation; indeed, different contexts call for differ- ent styles. Here, we consider an important dimen- sion of style, namely, simplicity. Systems that can rewrite text into simpler versions promise to make information available to a broader audience, such as non-native speakers, children, laypeople, and so on. One major effort to produce such text is the Simple English Wikipedia (henceforth SimpleEW)1, a sort of spin-off of the well-known English Wikipedia (henceforth ComplexEW) where hu- man editors enforce simplicity of language through rewriting. The crux of our proposal is to learn lexical simplifications from SimpleEW edit histories, thus leveraging the efforts of the 18K pseudonymous in- dividuals who work on SimpleEW. Importantly, not all the changes on SimpleEW are simplifications; we thus also make use of ComplexEW edits to filter out non-simplifications. Related work and related problems Previous work usually involves general syntactic-level trans- 1http://simple.wikipedia.org formation rules [1, 9, 10].2 In contrast, we explore data-driven methods to learn lexical simplifications (e.g., “collaborate” → “work together”), which are highly specific to the lexical items involved and thus cannot be captured by a few general rules. Simplification is strongly related to but distinct from paraphrasing and machine translation (MT). While it can be considered a directional form of the former, it differs in spirit because simplification must trade off meaning preservation (central to para- phrasing) against complexity reduction (not a con- sideration in paraphrasing). Simplification can also be considered to be a form of MT in which the two “languages” in question are highly related. How- ever, note that ComplexEW and SimpleEW do not together constitute a clean parallel corpus, but rather an extremely noisy comparable corpus. For ex- ample, Complex/Simple same-topic document pairs are often written completely independently of each other, and even when it is possible to get good sentence alignments between them, the sentence pairs may reflect operations other than simplifica- tion, such as corrections, additions, or edit spam. Our work joins others in using Wikipedia revi- sions to learn …

Content truncated to 500 words