Although the translation provided by machine may not always be as good as the translation done by an experienced translator, machine translation has nowadays become a helpful and convenient tool for people to reduce language barriers. After decades of development, machine translation can provide translation with high accuracy based on the neural translation model. Below the systems which machine translation has been established on over the past decades will be introduced.
Rule-based translation is one of the earliest systems used in machine translation. It’s constructed based on the linguistics rules containing syntactic and semantic information of both the source language and the target language. In rule-based translation, the words and the sentence structures are processed and linked to the appropriate words organized according to the right grammatical manner in the target language. However, the rules are not very helpful when the words used have ambiguous meanings and the sentences have very complicated structures. Moreover, it’s costly to build up a large set of rules sufficient to do the conversion between the source and the target language, and many rules do have limitations and may not be accurate under certain circumstances.
The statistical translation model is then used to enhance the quality of machine translation. Machine is trained by analyzing the existing corpora of the source and target languages. It would then predict the way to translate a sentence appropriated based on the corpora previously studied. Compared to the rule-based approach, people no longer need to build a large set of rules for translation, and machine can learn to translate just based on the existing corpora on its own; moreover, it’s not limited by rules and can provide greater translation when it an exception to a rule arises. However, there may not be adequate corpora for certain language pairs, and as machine learns to translate on its own, some errors may arise and it’s hard for people to detect those errors. In addition, when the language structures of the source language and the target language are very different, the accuracy of the translation may be a problem.
The neural model is the model prevalently used nowadays for machine translation. The training method of the neural model is similar to that of the statistical model, but for neural model an integrated neural model is built to help machine predict the appropriate translation of the words in a sentence. Once the best way to translate is learned, the neural model can provide translation with high accuracy after reading text in the source language. The neural model has improved the accuracy of machine translation, and it’s not limited by some shortcomings of the models used earlier.