Machine Translation is moving fast, so which one will be left in the past?
Long long ago, translations were done the hard way, by a person sitting in a cave translating Ugg to Uck so one tribe could understand another. These days, caves have now become nice airy offices with ergonomic keyboards, and as translation’s evolved, and so has machine translation (MT). They’ve developed from basic dictionary software capable of one word at a time, and now we’re in an age that allows MT to translate over six million words an hour. But as impressive as that may be, does it really work?
Like all tough decisions in business, it always comes down to quality vs quantity. In this case, machine translation is the quantity due to its high turnout rate. MT can provide translations in bulk very fast, but it still requires proofreading with a fine-tooth comb to ensure all the right tenses and syntaxes are used.
MT is good, but it’s quality is definitely something that can never be compared to a human. Scientists in Korea even tested it out and humans won hands down. So, if you are looking for quality, human is the way to go.
Both processes may sound very similar, but here’s the key difference. MT has been evolving more and more, with the aim that it will require less and less corrections. From an online one word at a time dictionary, to “Rule based MT” (being able to form sentences) to phrasal MT – MT has been on a massive journey and it doesn’t seem to be coming to a halt any time soon. Are we now we’re seeing MT in what might be its Final Evolution?
Neural Machine Translation (NMT) was big news for google, and even made headline news. NMT is the pinnacle of machine translation. When MT was being developed, they started small and coded in the rules needed to make sentences.
As you may have guessed with the word Neural everywhere, NMT has been coded to a near Artificial Intelligence level. So, it learns from its mistakes and picks up on writing styles being used. It basically attempts to use recurrent neural network (RNN) to increase translation quality. Unlike standard MT, NMT builds a single neural network that can be jointly turned to maximize the translation quality. NMT uses what’s called deep learning to build an artificial neural network, whereas MT uses glossaries etc. to train the MT engine. Unlike MT which runs on CPUs, NMT runs on GPUs. Any translator or translation agency who has had the chance to use NMT will tell you that the quality is incomparable.