Optimising SQL Generation for Afrikaans Through an English-to-Afrikaans Double Translation Method

Neural Machine Translation (NMT) has achieved remarkable progress in translating languages with varying resource levels. However, the performance can be hindered by limited data availability and poor-quality corpora for low-resource languages (LRL). The aim of this study is to optimise SQL generation for Afrikaans through a “query double translation” method. Here, high-resource language (HRL) prompts (e.g., English) are translated to a LRL (e.g., Afrikaans) to improve SQL query generation. First, we evaluate three NMT models for English-to-Afrikaans translation, selecting the best-performing model to translate an English-to-SQL dataset. This dataset is then fine-tuned for Afrikaans-to-SQL conversion. The resulting model's performance is benchmarked against the original English SQL data and compared with two large language models (LLM). This study addresses NMT challenges in LRLs and highlights potential SQL generation improvements for underrepresented languages.