The Roots of Machine Translation
Joss Moorkens, Andy Way, Séamus Lankford · 2024
This chapter explains the provenance of machine translation (MT) going back to Warren Weaver’s famous memorandum in 1949 and bringing us up to date with an overview of what has happened in the interim. From the early ‘toy’ experiments in the 1950s through to neural MT and approaches to translation using large language models (LLMs), each paradigm is explained. Starting with rule-based MT, readers are shown how this worked, but also what the problems were with this approach, leading to the statistical MT models of the late 1980s which became the dominant paradigm for 35 years. Statistical MT quality was surpassed in the last ten years by approaches based on neural networks, and neural MT is now the state of the art, although it is currently being challenged by multilingual LLMs. During this high-level presentation of the various methods that have been used to try to tackle translation, we meet some of the important historical figures, and see how automatic MT evaluation metrics and shared tasks helped contribute to improvements in MT quality. Finally, we contend that overhyping the capabilities of MT helps no-one and demonstrate clearly that MT is far from being a ‘solved problem’.