In this article we present a novel corpus-based method for inducing romanization systems for languages through a bilingual alignment of transliteration word pairs. First, the word pairs are aligned using a non-parametric Bayesian approach, and then for each grapheme sequence to be romanized, a particular romanization is selected according to a user-specified criterion. As far as we are aware, this paper is the only one to describe a method for automatically deriving complete romanization systems. Unlike existing human-derived romanization systems, the proposed method is able to discover induced romanization systems tailored for specific purposes, for example, for use in data mining, or efficient user input methods. Our experiments study the romanization of four totally different languages: Russian, Japanese, Hindi and Myanmar. The first two languages already have standard romanization systems in regular use, Hindi has a large number of diverse systems, and Myanmar has no standard system for romanization. We compare our induced romanization system to existing systems for Russian and Japanese. We find that the systems so induced are almost identical to Russian, and 69% identical to Japanese. We applied our approach to the task of transliteration mining, and used Levenshtein distance as the romanization selection criterion. Our experiments show that our induced romanization system was able to match the performance of the human created system for Russian, and offer substantially improved mining performance for Japanese. We provide an analysis of the mechanism our approach uses to improve mining performance, and also analyse the differences in characteristics between the induced system for Japanese and the official Japanese Nihon-shiki system. In order to investigate the limits of our approach, we studied the romanization of Myanmar, a low-resource language with a large vocabulary of graphemes. We estimate the approximate corpus size required to effectively romanize the most frequency k graphemes in the language for all values of k up to 1800.

The contribution of this paper is two-fold. Firstly, we conduct a large-scale real-world evaluation of the effectiveness of integrating an automatic transliteration system with a machine translation system. A human evaluation is usually preferable to an automatic evaluation, and in the case of this evaluation especially so, since the common machine translation evaluation methods are affected by the length of the translations they are evaluating, often being biassed towards translations in terms of their length rather than the information they convey. We evaluate our transliteration system on data collected in field experiments conducted all over Japan. Our results conclusively show that using a transliteration system can improve machine translation quality when translating unknown words. Our second contribution is to propose a novel Bayesian model for unsupervised bilingual character sequence segmentation of corpora for transliteration. The system is based on a Dirichlet process model trained using Bayesian inference through blocked Gibbs sampling implemented using an efficient forward filtering/backward sampling dynamic programming algorithm. The Bayesian approach is able to overcome the overfitting problem inherent in maximum likelihood training. We demonstrate the effectiveness of our Bayesian segmentation by using it to build a translation model for a phrase-based statistical machine translation (SMT) system trained to perform transliteration by monotonic transduction from character sequence to character sequence. The Bayesian segmentation was used to construct a phrase-table and we compared the quality of this phrase-table to one generated in the usual manner by the state-of-the-art GIZA++ word alignment process used in combination with phrase extraction heuristics from the MOSES statistical machine translation system, by using both to perform transliteration generation within an identical framework. In our experiments on English-Japanese data from the NEWS2010 transliteration generation shared task, we used our technique to bilingually co-segment the training corpus. We then derived a phrase-table from the segmentation from the sample at the final iteration of the training procedure, and the resulting phrase-table was used to directly substitute for the phrase-table extracted by using GIZA++/MOSES. The phrase-table resulting from our Bayesian segmentation model was approximately 30% smaller than that produced by the SMT system's training procedure, and gave an increase in transliteration quality measured in terms of both word accuracy and F-score.

This paper proposes an unsupervised word segmentation algorithm that identifies word boundaries in continuous source language text in order to improve the translation quality of statistical machine translation (SMT) approaches. The method can be applied to any language pair in which the source language is unsegmented and the target language segmentation is known. In the first step, an iterative bootstrap method is applied to learn multiple segmentation schemes that are consistent with the phrasal segmentations of an SMT system trained on the resegmented bitext. In the second step, multiple segmentation schemes are integrated into a single SMT system by characterizing the source language side and merging identical translation pairs of differently segmented SMT models. Experimental results translating five Asian languages into English revealed that the proposed method of integrating multiple segmentation schemes outperforms SMT models trained on any of the learned word segmentations and performs comparably to available monolingually built segmentation tools.

This paper presents a technique for class-dependent decoding for statistical machine translation (SMT). The approach differs from previous methods of class-dependent translation in that the class-dependent forms of all models are integrated directly into the decoding process. We employ probabilistic mixture weights between models that can change dynamically on a sentence-by-sentence basis depending on the characteristics of the source sentence. The effectiveness of this approach is demonstrated by evaluating its performance on travel conversation data. We used this approach to tackle the translation of questions and declarative sentences using class-dependent models. To achieve this, our system integrated two sets of models specifically built to deal with sentences that fall into one of two classes of dialog sentence: questions and declarations, with a third set of models built with all of the data to handle the general case. The technique was thoroughly evaluated on data from 16 language pairs using 6 machine translation evaluation metrics. We found the results were corpus-dependent, but in most cases our system was able to improve translation performance, and for some languages the improvements were substantial.