Recently, Google announced that its translation support has reached 103 languages. The addition of artificial intelligence technology similar to neural network learning not only makes the electronic synthesis pronunciation more natural, but also enables more fluent Chinese-English and English-Chinese translation. In an earlier announcement, Google also announced that the Google Translate system can now translate between different languages, thereby improving the translation performance of languages that were previously unable to directly translate each other.
Based on the current foundation of Google Translate, it mostly compares translation results and adjusts the results based on the opinions of many people regarding different semantics to make the translation content more accurate. However, it still basically translates different languages into English. When translating from English as the intended result, for example, if you want to translate from Chinese to Korean, you will first translate from Chinese to English, then from English to Korean, and vice versa. However, if neither the original translation language nor the intended translation language supports English translation, the system may have to perform language translation two or three times or more before it can successfully translate into the intended language result, which may make the translation result more unnatural.
By introducing the previous neural network-like learning technology, Google Translate will pre-translate all languages into an intermediate language that the system can understand, and then convert it into the expected translation language. Therefore, multiple translations can be avoided in the translation process between different languages. If all goes well, it can even achieve direct translation of various languages without the need for translation into other languages, avoiding deviation from the original meaning.
As for the operating mode behind Google's neural network machine translation system, the most important thing is to compare all language-related information, compare the sentences and words one by one, and "understand" their relevant relationships, and then understand the meaning of the entire sentence. It will not produce the strange sentence results produced by word-for-word translation in the past. Through this kind of comparison mode, the system will be able to compare sentences expressing the same meaning in different languages, thereby achieving a literal translation effect that is closer to natural sentences.
According to the Google research team, the introduction of 12 languages into the translation model will significantly reduce the space required for language translation data in the past, while only slightly reducing the impact on translation quality results. However, it is expected to increase the efficiency of translation between more different languages, and at the same time achieve different language translation effects with less data storage.
