Why Not Grab a Free Lunch? Mining Large Corpora for Parallel Sentences to Improve Translation Modeling

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 626–630, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Why Not Grab a Free Lunch? Mining Large Corpora for Parallel Sentences to Improve Translation Modeling Ferhan Ture Dept. of Computer Science, University of Maryland fture@cs.umd.edu Jimmy Lin The iSchool University of Maryland jimmylin@umd.edu Abstract It is well known that the output quality of statistical machine translation (SMT) systems increases with more training data. To ob- tain more parallel text for translation mod- eling, researchers have turned to the web to mine parallel sentences, but most previous ap- proaches have avoided the difficult problem of pairwise similarity on cross-lingual docu- ments and instead rely on heuristics. In con- trast, we confront this challenge head on us- ing the MapReduce framework. On a mod- est cluster, our scalable end-to-end processing pipeline was able to automatically gather 5.8m parallel sentence pairs from English and Ger- man Wikipedia. Augmenting existing bitext with these data yielded significant improve- ments over a state-of-the-art baseline (2.39 BLEU points in the best case). 1 Introduction It has been repeatedly shown that “throwing more data at the problem” is effective in increasing SMT output quality, both for translation modeling (Dyer et al., 2008) and for language modeling (Brants et al., 2007). In this paper, we bring together two re- lated research threads to gather parallel sentences for improved translation modeling: cross-lingual pair- wise similarity to mine comparable documents and classification to identify sentence pairs that are mu- tual translations. Unlike most previous work, which sidesteps the computationally-intensive task of pairwise compar- isons to mine comparable documents and instead re- lies on heuristics, we tackle the challenge head on. This paper describes a fully open-source, scalable MapReduce-based processing pipeline that is able to automatically extract large quantities of parallel sen- tences. Experiments examine the impact data size has on a state-of-the-art SMT system. We acknowledge that different components of this work are not novel and the general principles behind “big data” MT are well known. However, when con- sidered together with our previous work (Ture et al., 2011), to our knowledge this is the first exposition in which all the pieces have been “put together” in an end-to-end pipeline that is accessible to academic research groups. The framework described in this paper is entirely open source, and the computational resources necessary to replicate our results are rela- tively modest. Starting from nothing more than two corpora in different languages (in German and English, in our case), we are able to extract bitext and improve translation quality by a significant margin (2.39 BLEU points), essentially “for free”. By varying both the quantity and quality of the bitext, we char- acterize the tradeoffs between the amount of data, computational costs, and translation quality. 2 Related Work The idea of mining parallel sentences, particularly from the web, is of course not new. Most adopt a two step process: 1. identify comparable documents and generate …

Content truncated to 500 words