English-latvian SMT: the challenge of translating into a free word order language

This paper presents a comparative study of two approaches to statistical machine translation (SMT) and their application to a task of English-to-Latvian translation, which is still an open research line in the field of automatic translation. We consider a state-of-the-art phrase-based SMT and an alternative N-gram-based SMT systems. The major differences between these two approaches lie in the distinct representations of bilingual units, which are the components of the bilingual model driving translation process and in the statistical modeling of the translation context. Latvian being a rather free word order language implies additional difficulties to the translation process. We contrast different reordering models and investigate how well they deal with the word ordering issue. Moving beyond automatic scores of translation quality that are classically presented in MT research papers, we contribute presenting a manual error analysis of MT systems output that helps to shed light on advantages and disadvantages of the SMT systems under consideration and identify the most prominent source of errors typical for both SMT systems.