Substring-based unsupervised transliteration with phonetic and contextual knowledge

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 270–279, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Substring-based unsupervised transliteration with phonetic and contextual knowledge Anoop Kunchukuttan1, Pushpak Bhattacharyya1, Mitesh M. Khapra2 1 Department of Computer Science & Engineering, Indian Institute of Technology Bombay. {anoopk,pb}@cse.iitb.ac.in 2 IBM Research India. mikhapra@in.ibm.com Abstract We propose an unsupervised approach for substring-based transliteration which in- corporates two new sources of knowledge in the learning process: (i) context by learning substring mappings, as opposed to single character mappings, and (ii) pho- netic features which capture cross-lingual character similarity via prior distributions. Our approach is a two-stage iterative, boot-strapping solution, which vastly outperforms Ravi and Knight (2009)’s state-of-the-art unsupervised translitera- tion method and outperforms a rule-based baseline by up to 50% for top-1 accuracy on multiple language pairs. We show that substring-based models are superior to character-based models, and observe that their top-10 accuracy is comparable to the top-1 accuracy of supervised systems. Our method only requires a phonemic rep- resentation of the words. This is possi- ble for many language-script combinations which have a high grapheme-to-phoneme correspondence e.g. scripts of Indian lan- guages derived from the Brahmi script. Hence, Indian languages were the focus of our experiments. For other languages, a grapheme-to-phoneme converter would be required. 1 Introduction Transliteration is a key building block for multi- lingual and cross-lingual NLP since it is useful for user-friendly input methods and applications like machine translation and cross-lingual information retrieval. The best performing solutions are su- pervised, discriminative learning methods which learn transliteration models from parallel translit- eration corpora. However, such corpora are avail- able only for some language pairs. It is also expen- sive and time-consuming to build a parallel corpus. This limitation can be addressed in three ways: (i) train a transliteration model on mined parallel transliterations. The transliterations can be mined from monolingual comparable corpora (Jagarla- mudi and Daumé III, 2012) or parallel translation corpora (Sajjad et al., 2012). However, it may not be possible to mine enough transliteration pairs to train a system for most languages (Irvine et al., 2010). (ii) transliterate via a bridge language (Khapra et al., 2010) when transliteration corpora involving bridge languages is available. (iii) learn transliteration models in an unsupervised setting using only monolingual word lists. Unsupervised transliteration can be defined as: Learn a translit- eration model (T X) from the source language (F) to the target (E) language given their respective monolingual word lists, WF and WE respectively. We explore this direction in the present work, ad- dressing shortcomings in the previous work (Ravi and Knight, 2009; Chinnakotla et al., 2010). Our work addresses two major limitations in existing unsupervised transliteration approaches: (i) lack of linguistic signals to drive the learning, and (ii) limited use of context since their model is character-based. Due to this knowledge-lite ap- proach, these model performs poorly. Our pri- mary contributions are novel methods to incor- porate two knowledge sources, phonetic and con- textual, in the …

Content truncated to 500 words