Learning Translation Rules from Bilingual English - Filipino Corpus

Proceedings of PACLIC 19, the 19th Asia-Pacific Conference on Language, Information and Computation. Learning Translation Rules from Bilingual English – Filipino Corpus Michelle Wendy Tan, Raymond Joseph Ang, Natasja Gail Bautista, Ya Rong Cai, Bianca Tanlo De La Salle University - Manila 2401 Taft Avenue 1004 Manila, Philippines tanmw@dlsu.edu.ph, 10137661@dlsu.edu.ph, 10598677@dlsu.edu.ph, 10592709@dlsu.edu.ph, 10127003@dlsu.edu.ph Abstract Most machine translators are implemented using example based, rule based, and statistical approaches. However, each of these paradigms has its drawbacks. Example based and statistical based approaches are domain specific and requires a large database of examples to produce accurate translation results. Although rule based approach is known to produce high quality translations, a linguist is necessary in deriving the set of rules to be used. To address these problems, we present an approach that uses the rule based approach in translating from English to Filipino text. It incorporates learning of rules based on the analysis of a bilingual corpus in an attempt to eliminate the need for a linguist. The learning algorithm is based on seeded version space learning algorithm as presented by Probst (2002). Implementation of the algorithm has been modified to allow learning of non-lexically aligned languages and to adapt to the complex free word order of the Filipino language. 1. Introduction The demand for language translation has greatly increased as globalization takes place. Some businesses have turned to machine translators (MT) that usually provide fast and consistent translation results as compared to human translators. However, the quality of such translation services is usually poor as compared to text translated by humans. Several MT paradigms have attempted to improve the quality of translation such as rule based, example based, and statistical- based MT. Rule based MT systems usually have impressive results for a given domain. However, creating the translation rules is tedious and time consuming, requiring a linguist who thoroughly knows the construct of both languages. In addition, since language is constantly changing, these rules should be updated regularly in order to maintain the system. Example based MT systems presents a different approach. It uses a bilingual corpus as the basis for translation. Although this approach is more flexible, the quality of the translation greatly depends on the quality of the examples. A new approach in Statistical-based MT is by using probabilities to determine if a phrase or word occurs in the current context. Using a corpus, statistical-based MT allows flexibilty to accomodate different domains. However, it still has its drawbacks. Usually the output is ungrammatical and usually unacceptable to human linguist. It is still important to incorporate abstract syntax rules to improve the quality of translation. The paper presents an approach that combines these paradigms. Rather than building the rules by hand, the system incorporates a training phase where it learns transfer rules by examples found in a bilingual corpus. As such, since the system learns from examples, there is no need for a human linguist to generate the rules for translation. It also allows the MT system to translate in different …

Content truncated to 500 words