43 ms·

The Languages That Defy Auto-Translate

• pseingatl 6y ago
  It's amazing how far machine translation has come. I'd say we are now in the same place chess engines were in the 80's before Stockfish, Big Blue and others. For many languages the translations are astonishingly good--not native speaker good, but good enough for government work good. Mandarin approaches 90% accuracy for many kinds of texts; Arabic is hit or miss--either 80-85% or it's missing words so you're down to 30% or worse. German and French are excellent; take a look at deepl.com.
• yosito 6y ago
  > Humans don't need to read years' worth of parliamentary records to learn a language Anyone who has ever successfully learned a language knows that you need at least this many resources, and probably far more, to train yourself to become fluent in a language. The main difference is while machines may be trained on parliamentary records, humans are usually trained on textbooks, dictionaries, trips to the grocery store, news articles, text messages, music and conversations with neighbors.
  • Valgrim 6y ago
    Of course you need a variety of data sources in different contexts, but I don't think you need to put a microphone on people during their daily life to build an accurate model. Call center exchanges might be a good source of data, or recorded student online conversations on a variety of subjects.