Leaping Forward with Neural Nets
Ten years after introducing Translate, Google has made a major improvement. App users can now enjoy greater accuracy with several major foreign languages. They announced the good news mid-November: Google is now using Neural Machine Translation. A recent study testing the new method showed 60% greater accuracy in translation compared to the old.
In the past, the app has translated 103 languages using Statistical Machine Translation (STM). STM models use analysis of large data sets from texts in sets of two languages. They’re fairly accurate for stand-alone word and phrase translations between similar language pairs.
However, they aren’t rule-based and don’t play well with languages using different word-order patterns. For inflected languages similar to Latin, even working with very short sentences can be problematic.
Language Lessons for Machines
Neural Machine Translation (NMT), on the other hand, can tackle entire sentences, paragraphs, or web pages. NMT models are capable of making decisions about meaning based on context. They typically use two neural networks – one for input and another to create the output. In the past, NMT was less accurate than Statistical Translation for very large sets of data. It worked more slowly and tended to ignore unusual words.
However, the new study testing 500 Wikipedia sample sentences found Google’s new system to be fast and accurate in English translations to and from Chinese, French, and Spanish. The 8 languages now available cover about 35% of all current Translate requests: French, German, Spanish, Portuguese, Chinese, Japanese, Korean, and Turkish.
The company plans to expand to all 103.
It’s going to take a few more years for Google to master Latin and Greek. In the meantime … the next time you travel to Paris, Madrid, or Tokyo without fluency in the native language, you’ll be able to read and speak with greater confidence.