A Comprehensive Survey on Transformer-Based Machine Translation: Identifying Research Gaps and Solutions for Large Language Models

This paper handles the key challenges in transformer-based architectures for machine translation, proposing solutions to specific issues and highlighting areas where researchers can focus on bridging existing gaps, thereby reducing the effort needed to identify research opportunities. Notably, our findings show that the BERT-based model performs better as compared to the transformer-based model in accuracy. Additionally, we discuss various applications of LLMs. Furthermore, we conducted a statistical test (t-value, p-value, Mann-Whitney U, and Cliff’s Delta) on top of the WMT14 dataset for English, German, and French translations. Our results confirm that the BERT-based model consistently performs better in machine translation. As far as we know, this type of analysis has not been performed in any MT survey paper. Simulations conducted with Google Translator, LLAMA-3, and ChatGPT on the UMCorpus and EnIndic datasets reveal that Google Translator performs better than ChatGPT, especially in translating low-resource languages, where ChatGPT’s performance noticeably declines. We also observed that Gemini, LLAMA-3, and ChatGPT must improve translation quality for technical, dense sentences.