Improving the neural network-based machine transliteration for low-resourced language pair

Improving the neural network-based machine transliteration for low-resourced language pair Ngoc Tan Le Universite du Quebec a Montreal / 201, avenue du President-Kennedy, Montreal, Canada le.ngoc_tan@uqam.ca Fatiha Sadat Universite du Quebec a Montreal / 201, avenue du President-Kennedy, Montreal, Canada sadat.fatiha@uqam.ca Abstract Grapheme-to-phoneme models are key com- ponents in automatic speech recognition and text-to-speech systems. With low-resourced language pairs that do not have available and well-developed pronunciation lexicons, grapheme-to-phoneme models are particularly useful. The current work presents an approach that applies an alignment representation for in- put sequences and pre-trained source and tar- get embeddings to overcome the translitera- tion challenge for a low-resourced languages pair. The proposed method is tested with French and Vietnamese low-resourced lan- guage pair. The results showed promising improvement compared to the state-of-the-art approaches, with a large increase of +7.30 BLEU and a reduction in translation error rate (TER) of −8.16 and phoneme error rate (PER) of −14.17. 1 Introduction In many domains, including machine translation, end-to-end deep learning models have become a valuable alternative to more traditional statistical ap- proaches (Wu et al., 2016; Koehn, 2017). This is our motivation for applying a similar approach to build a machine transliteration system. In the state-of-the-art, the grapheme-to-phoneme methods were based on the use of grapheme- phoneme mappings (Oh et al., 2006; Duan et al., 2016). However, recurrent neural networks ap- proaches do not require any alignment informa- tion. In this study, we propose a method to build a low-resourced machine transliteration system, using RNN-based models and alignment information for input sequences. We are interested in solving out-of- vocabulary words for machine translation systems, such as proper nouns or technical terms, for a low- resourced language pair. The structure of the article is as follows: Section 2 presents the state of the art on machine transliter- ation. In section 3, we describe some background and our proposed approach. Then, in section 4, we present several experiments, the evaluations and an error analysis. Finally, section 5 concludes with some perspectives. 2 Related work Transliteration can be considered as a subtask of machine translation, when we need to translate source graphemes into target phonemes. In other words, an alignment model needs to be constructed first, and the translation model is built on the ba- sis of the alignments. Transliterating a word from the language of its origin to a foreign language is called Forward Transliteration, while transliterating a loan-word written in a foreign language back to the language of its origin is called Backward Transliter- ation (Karimi et al., 2011). Statistical techniques based on large parallel transliteration corpora work well for rich-resource languages but low-resource languages do not have the luxury of such resources. For such languages, rule-based transliteration is the only viable option. From 2009 to 2016, various transliteration sys- tems were proposed during the Named Entities PACLIC 32 331 32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018 Copyright 2018 by the authors Workshop evaluation campaigns1 (Duan et al., 2016). These …

Content truncated to 500 words