Learning Translation Model to Translate Croatian Dialects to Modern Croatian Language

The task of translating dialects into modern language is a challenging task since it requires enormous parallel data. Such data is hard to find especially when it comes to low-resource languages. Among them is the Croatian language which has very few datasets in the standardized version, let alone enough resources for its dialects. In the Croatian language, there are three main groups depending on the geographical position Shtokavian, Kajkavian, and Chakavian which also include other more specific local dialects. For solving these kinds of problems, unsupervised neural machine translation (UNMT) models are considered a good solution since they can be trained on monolingual data. In this paper, we propose an application of a modified version of the state-of-the-art UNMT model for dialect translation on monolingual data of the standardized Croatian language and its dialects. We experimented with several types of cross-lingual embeddings of the input data to determine the best approach that can leverage the similarities and differences between the language and the dialect. All techniques are evaluated on a small parallel dataset using the BLEU metric. Translating these dialects to the modern Croatian language helps in improving communication and access to information for all speakers.