When the Google Translate service inserted neural networks into its programming – a type of artificial intelligence that learns with experience – the translations improved surprisingly. Now this technology works in the platform’s mobile application without the need for the Internet, offering a better interpretation of the texts in other languages, but it is possible that it is also finding “hidden messages” that have attracted attention.
The site of “Vice” “Motherboard” informs that if you try to translate “dog” in Google Translator repeating the word 19 times from Maori to English, the meaningless query becomes: “Doomsday Clock is three minutes at We are experiencing characters and a dramatic development in the world, which indicate that we are getting closer and closer to the end of time and the return of Jesus. ”
The Reddit site was filled with examples like this. Translations from the Somali, for example, make mythological references to the Bible and quote the “name of the Lord”, the Deuteronomies and more. While on Twitter, people blame “ghosts” and “demons”, the Reddit channel TranslateGate speculates that they would be texts collected from emails or private messages.
“Google Translate learns from examples of translations on the web and does not use ‘private messages’ to carry out translations, nor would the system even have access to that content,” said Justin Burr, Google spokesperson, in an email to “Motherboard ” “This is simply a function of introducing nonsense into the system, which generates nonsense.”
When the spokesperson received an example of these strange translations, he disappeared from the platform.
According to the cited means, there are many possible explanations of why this type of messages. It may be disgruntled employees of Google, users who play to “suggest an edit”. But Harvard assistant professor Andrew Rush, who studies natural language processing and computer translation, explained that internal quality filters would likely detect manipulations of this kind.
For Rush, these strange translations may be more related to the change applied by Google Translator some years ago, when he began to use the technique of “neuronal machine translation”. This is a system that trains with large amounts of texts in one language and corresponding translations in another to create a model that allows you to move between the two.
When the mentioned translation method is fed with nonsense inputs, it can “hallucinate” strange outputs, such as the way in which Google DeepDream identifies and accentuates the patterns in the images. “The models are black boxes, which are learned from so many training instances that you can find,” said Rush.
“The vast majority of these will look like human language, and when you give it a new one it is trained to produce something, at all costs, that also looks like human language, but if you give it something very different, the best translation it will be something fluid, but not connected to the entrance “.
BBN Technologies senior scientist Sean Colbath, who specializes in machine translation, agreed with Rush’s explanation. He added that the languages that yield the strangest results (Somali, Hawaiian and Maori) have smaller bodies of translated text than the most spoken languages, such as English or Spanish. That is why it uses texts from the Bible, which has been translated into many languages, to generate a translation model.
“If they tried to build a model with those things, it’s possible that the model simply throws a hail pass (a semi-intentional pun) and slashes a part of their training,” Colbath said in a personal way. Rush agrees with this explanation, although Burr refused to confirm that Google uses this material as part of Google Translator training.