Statistical Machine Translation of Croatian Weather Forecasts: How Much Data Do We Need?

Abstract

This research is the first step towards developing a system for translating Croatian weather forecasts into multiple languages. This step deals with the Croatian-English language pair. The parallel corpus consists of a one-year sample of the weather forecasts for the Adriatic, consisting of 7,893 sentence pairs. Evaluation is performed by the automatic evaluation measures BLUE, NIST and METEOR, as well as by manually evaluating a sample of 200 translations. We have shown that with a small-sized training set and the state-of-the-art Moses system, decoding can be done with 96% accuracy concerning adequacy and fluency. Additional improvement is expected by increasing the training set size. Finally, the correlation of the recorded evaluation measures is explored.