Using Linguistic Knowledge in Statistical Machine Translation

In this thesis, we present methods for using linguistically motivated information to enhance the performance of statistical machine translation (SMT). One of the advantages of the statistical approach to machine translation is that it is largely languageagnostic. Machine learning models are used to automatically learn translation patterns from data. SMT can, however, be improved by using linguistic knowledge to address specific areas of the translation process, where translations would be hard to learn fully automatically. We present methods that use linguistic knowledge at various levels to improve statistical machine translation, focusing on Arabic-English translation as a case study. In the first part, morphological information is used to preprocess the Arabic text for Arabic-to-English and English-to-Arabic translation, which reduces the gap in the complexity of the morphology between Arabic and English. The second method addresses the issue of long-distance reordering in translation to account for the difference in the syntax of the two languages. In the third part, we show how additional local context information on the source side is incorporated, which helps reduce lexical ambiguity. Two methods are proposed for using binary decision trees to control the amount of context information introduced. These methods are successfully applied to the use of diacritized Arabic source in Arabic-to-English translation. The final method combines the outputs of an SMT system and a Rule-based MT (RBMT) system, taking advantage of the flexibility of the statistical approach and the rich linguistic knowledge embedded in the rule-based MT system. Thesis Supervisor: James R. Glass Title: Principal Research Scientist of the Computer Science and Artificial Intelligence Laboratory Thesis Supervisor: Steven R. Lerman Title: Professor of Civil and Environmental Engineering