Crowdsourcing Translation: Professional Quality from Non-Professionals

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1220–1229, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Crowdsourcing Translation: Professional Quality from Non-Professionals Omar F. Zaidan and Chris Callison-Burch Dept. of Computer Science, Johns Hopkins University Baltimore, MD 21218, USA {ozaidan,ccb}@cs.jhu.edu Abstract Naively collecting translations by crowd- sourcing the task to non-professional trans- lators yields disfluent, low-quality results if no quality control is exercised. We demon- strate a variety of mechanisms that increase the translation quality to near professional lev- els. Specifically, we solicit redundant transla- tions and edits to them, and automatically se- lect the best output among them. We propose a set of features that model both the translations and the translators, such as country of resi- dence, LM perplexity of the translation, edit rate from the other translations, and (option- ally) calibration against professional transla- tors. Using these features to score the col- lected translations, we are able to discriminate between acceptable and unacceptable transla- tions. We recreate the NIST 2009 Urdu-to- English evaluation set with Mechanical Turk, and quantitatively show that our models are able to select translations within the range of quality that we expect from professional trans- lators. The total cost is more than an order of magnitude lower than professional translation. 1 Introduction In natural language processing research, translations are most often used in statistical machine translation (SMT), where systems are trained using bilingual sentence-aligned parallel corpora. SMT owes its ex- istence to data like the Canadian Hansards (which by law must be published in both French and English). SMT can be applied to any language pair for which there is sufficient data, and it has been shown to pro- duce state-of-the-art results for language pairs like Arabic–English, where there is ample data. How- ever, large bilingual parallel corpora exist for rela- tively few languages pairs. There are various options for creating new train- ing resources for new language pairs. These include harvesting the web for translations or comparable corpora (Resnik and Smith, 2003; Munteanu and Marcu, 2005; Smith et al., 2010; Uszkoreit et al., 2010), improving SMT models so that they are bet- ter suited to the low resource setting (Al-Onaizan et al., 2002; Probst et al., 2002; Oard et al., 2003; Niessen and Ney, 2004), or designing models that are capable of learning translations from monolin- gual corpora (Rapp, 1995; Fung and Yee, 1998; Schafer and Yarowsky, 2002; Haghighi et al., 2008). Relatively little consideration is given to the idea of simply hiring translators to create parallel data, be- cause it would seem to be prohibitively expensive. For example, Germann (2001) estimated the cost of hiring professional translators to create a Tamil- English corpus at $0.36/word. At that rate, translat- ing enough data to build even a small parallel corpus like the LDC’s 1.5 million word Urdu–English cor- pus would exceed half a million dollars. In this paper we examine the idea of creating low cost translations via crowdscouring. We use Ama- zon’s Mechanical …

Content truncated to 500 words