Recently, I had
the opportunity to join a webinar on the topic of Microsoft’s Artificial Intelligence
(AI) machine translation. Machine translation is the use of programing and
machine to electronically translate language. Generally, this translation is
from one language into another. Until the advent of AI, translations were slow
and often mistranslated, lacking proper grammar and/or spelling. The biggest
advantage of using this technology is the precision of the translation.
According to Microsoft’s
own page on the topic, “new NMT [neural machine translation] packs produce
higher quality translations, which are up to 23 percent better, and about 50
percent smaller than the previous non-neural offline language packs”. This is a
quote referring to the use of offline apps available for mobile OS, such as
Android and Apple iOS.
During the
webinar, a product was shown that was a standalone translator, physically
smaller than a mobile phone. The ability to instantly, and precisely, translate
language will bring unprecedented accessibility across the globe. Looking
specifically at professional communication, this means far less
language-barriers to overcome when discussing business matters or any other type
of professional communication. Tasks such as subtitle localization and
instructional/informational documents will be completed like never before.
Another exciting
avenue of this technology is through the use of speech-recognition. From the
webinar, cellular switchboard word-error rate from 2009 was approximately 40%. Microsoft’s
switchboard word-error rate in 2017 (using NMT technology) was 5.1%. This also
leads to a vastly improved Text-to-Speech system, allowing for greater
accessibility.
The current
process of localization requires the text to be auto-translated (usually a
direct translation, lacking any proper grammar), then having a person or team
look over the text to ensure quality control. With new NMT technology, the
process can be instant and include grammar translation. Below is a graphic from
Microsoft’s informational page:
Image source: https://www.microsoft.com/en-us/translator/business/machine-translation/
Here we can see
the process in the most basic outline. The cloud represents the electronic “cloud”
of data hosted on servers, which passes the text to the Web API (application
program interface), then back through the system with the translated text. The web
API is referring to a system like Microsoft’s own Azure web API.
The webinar also
discussed the formation of the technology and the challenges the team had to
overcome. As with any neural-network framework, the NMT took hundreds of
thousands of words, phrases, and sentences and “learned”. This learning occurs
through a process as described by the picture below.
Image source: https://towardsdatascience.com/how-to-build-your-own-neural-network-from-scratch-in-python-68998a08e4f6
This image details the process of a neural
network. It begins with an expected result (dog) and multiple inputs. The
framework then works through repeated tests (the “hidden” column) until the
outputs are narrowed, and the correct result is obtained.
Here is a more
specific NMT illustration:
Image source: https://www.microsoft.com/en-us/translator/business/machine-translation/
The uses of
similar AI technology are limitless for professional communication. For
instance, adapting this technology to correct grammar and punctuation for
documents would not be far off. Many text programs already have basic
auto-correct and word choice. Using neural networks would advance this to
enable a standalone proof editing.
Granted, while
this all an exciting new horizon, this is yet another step in automation that
will doubtless put many people out of current jobs. At least, that is the worry
of many that tend to overlook the factor of an industry paradigm shift.
This paradigm
shift is the progression to fewer formal positions, such as text localization
or proof editor, and more machine programmers. We’ve already seen this in other
career fields, such as accounting, why not in professional communication?
For more
information regarding Microsoft’s AI NMT technology, visit the links here:
If you’re
interested in building a neural network from scratch, check out this tutorial:
