An Italian company has unveiled a new method of measuring AI progress: analyzing machine translation improvements.
Translateda translation service provider, used the approach to predict when we will reach singularity, a vague concept often defined as the point at which machines become smarter than humans.
The Rome-based company sets this milestone on the moment when AI offers “a perfect translation”. According to the new research, this comes when machine translation (MT) outperforms top human translation.
Buy your tickets for TNW Valencia in March!
The heart of technology comes to the heart of the Mediterranean
Translated’s analysis suggests this will happen before the end of the 2020s.
“[It will be] within this decade, at least for the top 10 languages in a context of medium complexity,” Marco Trombetti, the company’s CEO, tells TNW. “The reality is that this has already happened in some specific domains and in some languages. For some rare languages and domains, it may never come.”
Translated estimates are based on data taken from Mate cata computer aided translation tool (CAT).
The platform started in 2011 as an EU-funded research project. Three years later the system was released as open-source softwarethat professionals use improve their translations.
Translated offers Matecat as a freemium product. In return, users give the company data that is used to improve the models.
To chart the path to singularity, Translated tracked the time spent by users checking and correcting 2 billion MT suggestions. About 136,000 professionals around the world had performed these operations during Matecat’s 12 years of operation. The translations span diverse domains, from literature to technical subjects. They also include areas where MT still struggles, such as speech transcription.
“Singularity is very close.
The data suggests that AI is improving rapidly. In 2015, the average time it took the world’s leading translators to check and correct MT suggestions was about 3.5 seconds per word. Today, that number has dropped to 2 seconds per word.
At the current rate, the time will be 1 second in about five years. At that point, MT would provide the groundbreaking ‘perfect translation’. From a practical point of view, it is then more convenient to edit the translations of a machine than those of a top professional.
According to Trombetti, any task related to communication, understanding, listening and knowledge sharing will become multilingual with minimal investment.
“The exact date when we will reach the singularity point may vary, but the trend is clear: it is very close,” he says.
Advances in MT require increasing computing power, linguistic data, and algorithmic efficiency. Consequently, the researchers had assumed that progress would slow down as the singularity approached. To their surprise, the speed of development was very linear.
If this momentum continues as predicted, translated expect demand for MT to be at least 100 times higher. Employees may be concerned about their work being automated, but they can also benefit from it. Translated forecasts on at least a tenfold increase in requests for professional translations.
“All of our customers who deploy machine translation at scale are also spending more on human translation,” says Trombetti.
“Machine translation is a facilitator because it creates more interactions between markets and users that were previously disconnected. This generates business, and business generates higher quality content that requires professionals.”
Trombetti also expects new roles to emerge for top translators.
“To get the best quality out of machine translation, it needs to be trained by the best linguists. A significant number of translations are required to train and debug language models, so I think we will probably witness huge competition for the best translators in the coming years.”
“MT is a good predictor of the future of AI.
According to Translated, the new study is the first ever to quantify speed we are approaching singularity. The claim won’t convince every cynic, but MT is a compelling barometer of AI progress.
Human languages are notoriously tricky for machines to master. The subjectivity of linguistic meaning, the constantly evolving conventions, and the nuances of cultural references, puns, and tone can be elusive to computers.
In translation, these complexities must be modeled and linked in two languages. As result, algorithmic research, data collection and model sizes are often pioneers in the field. For example, the Transformer model was applied to MT for many years before being used in OpenAI’s GPT systems.
“MT is just a good predictor of what’s to come in AI,” says Trombetti.
If singularity is next, the Italian entrepreneur anticipates a new era for global communications.
He envisions universal translators, all content becomes available worldwide and everyone can speak their native language.
His definition of singularity may be questionable, but its appeal is undeniable.