Deep learning will change nothing; yet it changes everything

machinelearning

Deep learning, or machine learning, is part of recent development in pursuing real artificial intelligence. Unlike algorithms defined by humans, deep learning frameworks are geared to find patterns that humans did not know that existed and for which an algorithm could not have been written.

For example, deep learning promises great results in detecting cancer from tissue samples. A human might focus on what looks alarming; however, the computer has no preferences – it just finds patterns, if there are any to find.

There are plenty of great use cases for the current capabilities of machine learning: object labeling from photos, synthesize images and so much more – it can be even used to compose music.

Now even though this is incredible and things progress fast, too many people are expecting too much for machine learning based translation.

Machine translation has been a hot topic well over 50 years and a near future solution has been anticipated but yet not delivered. Machine learning will change nothing within the next few years. But eventually it will change everything.

Currently deep learning is not capable of reasoning. It is unable to produce a of chain of decisions that could have effects on future and past decisions. Computers are bad at reasoning, especially given that the decision must often be made on partial facts and vague hints. For example, if a text says that a car is flying, human processes that as a science fiction, a recent technology showcase or that a horrible accident is about to unravel. A machine interpreting that will have a hard time parsing the information. Even if we provide a context, e.g. a science fiction novel, the interpreted meaning could be incorrect.

Machine translation based on statistics will improve a lot when combined to deep learning; however, it will not make professional translators unnecessary. The end-result might be good enough for a professional translator to work from but it will not be good enough to publish without review.

On a sentence level, it will get many things right – especially when there is no ambiguity. However, when uncertainty in the process increases, the quality of the translation drops below acceptable. This is why we will need human translators for quite some time.

There are other areas that a machine is likely to make interpretation mistakes such as errors in the source text itself. Humans are prone to errors but often very good judging if there is a logical flaw in the source. Computers in contrast spot easily errors that fit into patterns such as spelling errors but on the other hand struggle dearly with errors that are revealed by the context and require human logic to be spotted.

Last but not least, machine learning could produce language that is grammatically correct but not idiomatic. Correctly translating idioms is always a challenge, more so when a computer cannot extrapolate the context well enough to make a suitable replacement.

The world evolves rapidly and so does language. Only 10 years ago we started to talk about tablets as touch-screen devices. The meaning of the word tablet has evolved from a stone with etched symbols to a hand-held computer.

A machine’s method of learning is based in categorized samples. Without the ability to reason, the current technology is not ready to revolutionize the translation industry – yet. It certainly will, and Transfluent will be one of the first companies to use it when the time comes. Whether it’s another five, ten or 30 years until then, our professional translators are here to produce accurate human-powered translation for you!

Image source: Deep learning applied to Transfluent logo (by deepdreamgenerator.com)

Author

Tomi Heiskanen

My main mission is to superpower you! Technology boosts productivity when it is easy to use and all manual tasks are automated.