Popular Science Article
The Past and Present of Machine Translation
From the Tower of Babel to Esperanto, and now to machine translation, humanity has been trying to eliminate the barriers of language communication. The Tower of Babel fell short, Esperanto failed, where will machine translation go? Let us first review the development history of machine translation.
The early machine translation was rule-based translation, where machines integrated word-by-word translations according to grammatical rules written by human experts. Although this method allows for deep language analysis, the manual rule writing is costly and often falls into the trap of repeatedly debugging rules.
In the early 1990s, IBM proposed the concept of statistical machine translation. Subsequently, statistical machine translation gradually evolved from word-based models to phrase-based and syntax-based models, gradually replacing rule-based translation. Statistical machine translation builds a corpus from existing bilingual texts, dividing the text into entries of a certain length, and selects the translation with the highest matching degree from the corpus to provide the “most likely” translation result. When the corpus data is sufficient, statistical machine translation outperforms rule-based translation. However, due to its adherence to the fidelity of the original text, the translations are often very stiff and “literal translations.”
The technology of neural machine translation was proposed in 2013 and has developed rapidly in just a few years. It takes a different path, using neural networks to directly achieve automatic translation between two sentences, allowing the machine to learn language features from the corpus, learning how to translate from the source language to the target language. Moreover, because there is no process of assembling small segments into larger segments, the translation results are closer to natural language, especially like what a native speaker would say.
0
Google’s neural machine translation uses Recurrent Neural Networks (RNN), which can take the previous state as input and bring it into the next calculation. This means that when translating a word, it considers the meanings of all previous words. When translating the second word, it considers the meaning of the first word, when translating the third word, it considers the meanings of the first two, and so on. Thus, it achieves translation at the sentence level, and the results are more logical. However, it still makes significant errors that human translators would never make, such as omissions and mistranslations of proper nouns or rare terms.
In May 2017, the Facebook AI research team released a machine translation technology based on Convolutional Neural Networks (CNN). Its translation accuracy surpassed that of Google’s neural machine translation, and the translation speed increased by 9 times. This is because RNN follows a strict “left-to-right” or “right-to-left” order, processing one word at a time. CNN, on the other hand, can process all elements simultaneously, and its information processing is hierarchical, making it easier to capture the complex relationships in the data.
Currently, neural machine translation has rapidly replaced statistical translation, becoming a mainstream research topic in academia and widely applied in industry. We have reason to believe that it will further develop in the future, providing high-quality machine translation services to the public.
(Contributed by editor Wen Yingying)