An Algorithm for Unsupervised Transliteration Mining with an Application to Word Alignment

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 430–439, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics An Algorithm for Unsupervised Transliteration Mining with an Application to Word Alignment Hassan Sajjad Alexander Fraser Helmut Schmid Institute for Natural Language Processing University of Stuttgart {sajjad,fraser,schmid}@ims.uni-stuttgart.de Abstract We propose a language-independent method for the automatic extraction of transliteration pairs from parallel corpora. In contrast to previous work, our method uses no form of supervision, and does not require linguisti- cally informed preprocessing. We conduct experiments on data sets from the NEWS 2010 shared task on transliteration mining and achieve an F-measure of up to 92%, out- performing most of the semi-supervised sys- tems that were submitted. We also apply our method to English/Hindi and English/Arabic parallel corpora and compare the results with manually built gold standards which mark transliterated word pairs. Finally, we integrate the transliteration module into the GIZA++ word aligner and evaluate it on two word alignment tasks achieving improvements in both precision and recall measured against gold standard word alignments. 1 Introduction Most previous methods for building transliteration systems were supervised, requiring either hand- crafted rules or a clean list of transliteration pairs, both of which are expensive to create. Such re- sources are also not applicable to other language pairs. In this paper, we show that it is possible to ex- tract transliteration pairs from a parallel corpus us- ing an unsupervised method. We first align a bilin- gual corpus at the word level using GIZA++ and create a list of word pairs containing a mix of non- transliterations and transliterations. We train a sta- tistical transliterator on the list of word pairs. We then filter out a few word pairs (those which have the lowest transliteration probabilities according to the trained transliteration system) which are likely to be non-transliterations. We retrain the translitera- tor on the filtered data set. This process is iterated, filtering out more and more non-transliteration pairs until a nearly clean list of transliteration word pairs is left. The optimal number of iterations is automat- ically determined by a novel stopping criterion. We compare our unsupervised transliteration min- ing method with the semi-supervised systems pre- sented at the NEWS 2010 shared task on translit- eration mining (Kumaran et al., 2010) using four language pairs. We refer to this task as NEWS10. These systems used a manually labelled set of data for initial supervised training, which means that they are semi-supervised systems. In contrast, our system is fully unsupervised. We achieve an F- measure of up to 92% outperforming most of the semi-supervised systems. The NEWS10 data sets are extracted Wikipedia InterLanguage Links (WIL) which consist of par- allel phrases, whereas a parallel corpus consists of parallel sentences. Transliteration mining on the WIL data sets is easier due to a higher percentage of transliterations than in parallel corpora. We also do experiments on parallel corpora for two language pairs. To this end, we created gold standards in which …

Content truncated to 500 words