Expanding machine translation training data with an out-of-domain corpus using language modeling based vocabulary saturation

Expanding Machine Translation Training Data with an Out-of-Domain Corpus using Language Modeling based Vocabulary Saturation Burak Aydın burak.aydin@tubitak.gov.tr T ¨UB˙ITAK-B˙ILGEM, Gebze 41470, KOCAEL˙I, TURKEY Department of Computer Engineering, Bo˘gazic¸i University, Bebek 34342, ˙ISTANBUL, TURKEY Arzucan ¨Ozg¨ur arzucan.ozgur@boun.edu.tr Department of Computer Engineering, Bo˘gazic¸i University, Bebek 34342, ˙ISTANBUL, TURKEY Abstract The training data size is of utmost importance for statistical machine translation (SMT), since it affects the training time, model size, decoding speed, as well as the system’s overall success. One of the challenges for developing SMT systems for languages with less resources is the lim- ited sizes of the available training data. In this paper, we propose an approach for expanding the training data by including parallel texts from an out-of-domain corpus. Selecting the best out- of-domain sentences for inclusion in the training set is important for the overall performance of the system. Our method is based on first ranking the out-of-domain sentences using a language modeling approach, and then, including the sentences to the training set by using the vocabu- lary saturation filter technique. We evaluated our approach for the English-Turkish language pair and obtained promising results. Performance improvements of up to +0.8 BLEU points for the English-Turkish translation system are achieved. We compared our results with the translation model combination approaches as well and reported the improvements. Moreover, we implemented our system with dependency parse tree based language modeling in addition to the n-gram based language modeling and reported comparable results. 1 Introduction Most of the statistical methods that attempt to solve natural language processing problems achieve better results with increasing training data sizes. In statistical machine translation, the amount of data directly affects a system’s overall success. Increasing the size of the training data by effectively utilizing the data available in other domains (e.g. web, news, medical) using domain adaptation and data selection techniques is a promising research direction for improving the performance of an SMT system. This is especially important for low-resource languages and domains, for which there are only limited amounts of training data available. The English- Turkish language pair is an example low-resource language pair for machine translation. Most of the publicly available corpora for this language pair contain only thousands of sentences, whereas the training sets of language pairs with more resources (e.g. English-French) usually contain millions of sentences. The number of parallel sentences in the training set for English- Turkish hardly reaches to millions, even when all available corpora from different domains are combined. Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 180 In this paper, we introduce an approach that effectively combines different data selection methods for expanding in-domain training data with the available out-of-domain data in statis- tical machine translation. The method first scores the sentences in the out-of-domain corpus based on their similarities to the in-domain corpus using a language modeling approach. Then, it adapts the vocabulary saturation filter technique, which has recently been proposed in (Lewis and Eetemadi, 2013) for …

Content truncated to 500 words