Translation quality evaluation of Croatian-to-English machine-translated administrative texts
Translation quality evaluation of Croatian-to-English machine-translated administrative texts

Summary/Abstract: A once heavily flawed method of translation, machine translation (MT) has improved and continues to improve every day. One of the major issues during its development was its quality and how to measure and assess it objectively. This paper presents an attempt to apply a triangulation of translation quality assessment (TQA) methods: automatic and human assessment, corpus-based analysis and error analysis, on the translation of a sample of administrative texts from Croatian into English, to provide a comprehensive view of the said translation, compare the results of different methods and identify areas of greatest discrepancy in machine-generated translation. The texts were translated using Google Translate (GT), which relies on the currently dominant neural model that significantly reduces translation errors when compared to the phrase-based model, and it is expected to deal with issues such as congruence (i.e. agreement) and inflection better than other systems. This is of special importance for morphologically rich languages such as Croatian. Even though literature on MT and the evaluation of MT is abundant, this paper aims to contribute to research of an under-resourced Slavic language, Croatian.

• Details
• Contents