A Comparative Study of Extremely Low-Resource Transliteration of the World’s Languages

A Comparative Study of Extremely Low-Resource Transliteration of the World’s Languages Winston Wu, David Yarowsky Department of Computer Science, Center for Language and Speech Processing Johns Hopkins University {wswu,yarowsky}@jhu.edu Abstract Transliteration from low-resource languages is difficult, in large part due to the small amounts of data available for training transliteration systems. In this paper, we evaluate the effectiveness of several translation methods in the task of transliterating around 1000 Bible names from 591 languages into English. In this extremely low-resource task, we found that a phrase-based MT system performs much better than other methods, including a g2p system and a neural MT system. However, by combining the data and training a single neural system, we discovered significant gains over single-language systems. We release the output from each system for comparative analysis. Keywords: Bible, alignment, named entities, translation, transliteration 1. Introduction Transliteration is the process of converting text from one script to another. For example, Росси́ я is transliterated as Rossiya but would be translated as Russia. Transliter- ation is important in the process of borrowing names be- tween languages. In the case of low-resource languages, there is often little training data with which to train translit- eration models. Thus, one major obstacle for low-resource languages is the problem of out of vocabulary words (in the transliteration setting, these would be unknown characters). Machine translation systems might use a transliteration step to resolve unknown words, especially names; in transliter- ation, one can perform a similar process by attempting to resolve unknown characters. In this paper, we compare and contrast multiple methods of transliterating Bible names in over 500 languages into English and evaluate the effective- ness of a pre- and post-transliteration step in transliteration in resolving unknown characters. Our results indicate that a phrase-based machine translation system is the most effec- tive when training on data in the order of hundreds of words, while by simply concatenating multiple languages’ training data with some preprocessing, a single neural MT system trained to transliterate multiple languages to a single target language significantly outperforms the single-language sys- tems. We release each system’s transliteration output as a dataset for comparative analysis. To our knowledge, this is the first study of such scale that compares such a variety of methods on such small corpora. 2. Related Work Machine transliteration has been tackled using a variety of methods. For a comprehensive survey, we refer the reader to Karimi et al. (2011). In this paper, we focus on translit- eration of named entities across multiple language, espe- cially in a low-resource setting, using the paradigm of mono- tonic machine translation (Virga and Khudanpur, 2003). Our work is most similar to Irvine et al. (2010), who built character-based machine translation systems using names mined from Wikipedia. Their setting is higher resource, as they use data acquired from the web, and they experiment on a smaller set of languages. Other recent approaches to transliteration of low resource languages include Mayhew et al. (2016), who explore using surrogate languages in place …

Content truncated to 500 words