Enhancing Domain-Specific Neural Machine Translation with Ensemble Attention Mechanisms

Neural machine translation (NMT) is a method for developing a translation model using neural networks. Various modifications to NMT have been created to enhance its translation performance. The incorporation of attention techniques in NMT has become the state-of-the-art approach for various language pairs. However, this method still faces challenges when translating terms from specific domains. This research presents an approach to improve a domain-specific NMT system using an ensemble attention technique. The main idea behind our domain adaptation strategy is to combine general out-domain pre-trained models with a specific indomain pre-trained model as an additional attention mechanism. The experimental result shows significant improvements using our proposed method, measured by BLEU metrics for translation quality. Our approach achieved BLEU scores of 29.17 for the IndonesianEnglish (ID-EN) task and 30.39 for the Indonesian-Arabic (ID-AR) task.