TRANSFORMING MACHINE TRANSLATION FOR ISOLATING LANGUAGES WITH MULTI-SOURCE NEURAL MODEL

Recent advancements in Artificial intelligence (AI) and Deep learning have facilitated the rapid development of machine translation technologies, among them, Neural machine translation (NMT) models have demonstrated impressive performance, especially in handling multiple language pairs. However, due to their complexity and lack of appropriate data, contemporary NMT models still have a lot of challenges when applied to isolated languages, despite their great accomplishments. This paper proposes a multi-source neural model that employs two different encoders to process both the source word sequence and the linguistic feature sequences of isolating languages. Unlike traditional NMT models, this approach improves the encoders’ input embeddings by incorporating a second encoder that integrates the linguistic elements, including part-of-speech (POS) tags and lemma. To enhance the source sentence's context representation, this article combines the encoders' conditional data with the outputs of the decoders using a serial combination technique. In this way, different metrics such as METEOR and BLEU are examined to assess the suggested model's precision of translation. Experimental results indicate that our methodology works efficiently for isolating language translation, as evidenced by the improvement of +3.9 BLEU and +3.2 METEOR scores on translation tasks conventional NMT models perform. This highlights a significant advancement in integrating linguistic features to enhance translation accuracy for isolating languages.