A Novel Statistical Pre-Processing Model for Rule-Based Machine Translation System

A Novel Statistical Pre-Processing Model for Rule-Based Machine Translation System Yanli Sun SALIS, Dublin City University yanli.sun2@m ail.dcu.ie Sharon O’Brien SALIS, Dublin City University Sharon.obrien@ dcu.ie Minako O’Hagan SALIS, Dublin City University Minako.ohagan@d cu.ie Fred Hollowood Research & Deployment (SES), Symantec Corporation, Ireland FHollowood@symantec .com Abstract This paper introduces a new statistical pre- processing model for Rule-Based Machine Translation (RBMT) systems. We train a Statistical Machine Translation (SMT) system using monolingual corpora. This model can transform a source input to an RBMT system into a more target-language friendly or RBMT- system friendly “pivot” language. We apply this proposed model to translation from English to Chinese in a pilot project. Automatic evaluation scores (BLEU, TER and GTM) show that this pre-processing model can increase the quality of the output of the RBMT system, especially with an increase in the size of the training corpus. This model is applicable to language pairs which differ in grammar and language structures. 1 Introduction Recently, a new pre-processing approach that suggests changing the source language to be closer to the structure of the target language has been reported. Wang et al. (2007) reported that transforming Chinese sentences by using hand- coded linguistic rules to be closer to English in terms of syntactic structure, could increase the scores of the final translation by an MT System. Xu and Seneff (2008) transformed English texts into "Zhonglish" (English words in Chinese structure) before translating them by an MT system and found that human evaluations prefer the translation of “Zhonglish” to the translation of the original English texts. A number of other researchers have also described their pre- processing methods on other language pairs. Xia and MacCord (2004) reported the effect of automatically learnt rewrite patterns in improving English and French translation. Crego and Marino (2007) proposed an approach to coupling reordering and decoding in statistical machine translation and reported significant improvements in translation quality between English and Spanish. These papers focused on incorporating syntactic information into Statistical MT (SMT) systems with rules either hand-crafted or automatically extracted. Babych et al. (2009) conducted a similar study for an RBMT system. They applied "construction-level human evaluation” to discover systematically mistranslated contexts and then to “create automatic pre-editing rules to make the constructions more tractable for an RBMT system" (p36). Their study concentrated on some of the most frequently occurring light verb constructions (“verb phrases with a semantically depleted verb and its objects, such as take part, put pressure” etc.) In addition, they still needed to compose the pre-editing rules manually. Another common pre-processing approach to improve the performance of RBMT systems is Controlled Language (CL) authoring. CL rules can reduce post-editing effort and improve comprehensibility of the final translation (O’Brien, 2003; O’Brien and Roturier, 2007). However, again CL rules have to be manually crafted. Technical writers usually have to manually correct the sentences that violate certain CL rules. Besides, it is difficult to define specific rules and some general rules such as “discourage the use of dangling prepositions” (O’Brien, 2003), …

Content truncated to 500 words