Training Data Generation for Libyan Dialect–MSA Translation Using Word Replacement and Fine-tuning Strategies

Abstract

Translating Arabic dialects into modern standard Arabic is a difficult task due to the significant differences in grammar and vocabulary between them. Using neural machine translation approaches to translate Arabic dialects faces a significant challenge, which is the lack of parallel data. Unavailability of parallel data is a common problem in low-resource languages such as the Libyan dialect. To solve this problem, this paper presents a novel method to address this problem. This method uses both word replacement and fine-tuning techniques to build Libyan dialect-modern standard Arabic parallel data for neural machine translation. To build parallel data from monolingual modern standard Arabic sentences, it uses a word replacement to exchange modern standard Arabic words in the modern standard Arabic text with Libyan dialect equivalents. It also uses fine-tuning techniques to enhance the overall translation of the neural machine translation model, which is already pretrained on parallel data that is generated by word replacement. The method is evaluated using a test dataset of 1,000 parallel sentences. The results show that the method proposed in this paper achieves promising results compared to the baseline technique. It obtained a score of 1.37 for BLEU, a score of 12.87 for chrF, and 87.34 for TER. In contrast, the baseline model achieved a BLEU score of 0.12, a chrF score of 6.89, and a higher TER of 100.02.