Close to 50 of Japan’s leading companies from all industries including retail, life sciences, automobile and finance gathered at our SDL Japan Customer Summit held in Tokyo on November 27 . Together we explored and discussed the many new exciting opportunities offered by Machine Translation, and how the latest developments can transform their business.
Ryutaro Kusumoto, Vice President of Sales at SDL Japan, opened the event by looking back at the progress made since December 2017, when SDL Japan held its first local Machine Translation seminar. He stressed the exponential growth achieved by Artificial Intelligence (AI) over the past year, with increasing coverage and adoption across industries. He also highlighted the fact that Machine Translation represents one of the most complicated uses of AI, especially in Asian languages – a subject that would later be explored in more detail.
Over the course of half a day, four sessions were conducted looking at the opportunities offered by Artificial Intelligence in the field of Machine Translation, exploring how it can address market-specific needs. Insight was provided on how SDL’s research and development helps customers fully leverage the potential of AI in Machine Translation, and how we have greatly expanded our portfolio of solutions to support customer demands in Machine Translation, localization of content management.
Welcome to the Future
The opening keynote was held by Jim Saunders, SDL’s Chief Product Officer, where he shared an inspirational message on how our environment sees an exponential growth in content while introducing SDL’s business, technology and future perspectives.
Jim looked back at SDL’s leading positioning in the industry:
- 25 years of business in the industry
- Working with 88 of the top 100 global brands, including Japanese leading companies such as ANA, Panasonic and Canon
- Support for more than 180 languages
- Acting as a partner for a broad spectrum of large corporations in their Digital Transformation
Jim highlighted that “we currently live in a world where more data has been created in the past two years that in the entire previous history of the human race,” so how are we to approach managing that ever-increasing amount of content?
He explained that the first step is to thoroughly structure content. The explosion of information, which is unlikely to stop, challenges us into moving from content confusion to relevance. With an estimated 600 billion words being translated daily via generic Machine Translation portals, diversified channels and a growth of social media that directly impacts customer decisions, relevance is key.
However, currently 80 to 90% of the content that enterprises must process and understand today is unstructured, making it hard to navigate and leverage in business. Jim stated that this is where SDL comes into play, as the only company able to master content management and content delivery. Furthermore, SDL’s expertise in Linguistic AI, translation and services puts us “in a good position to help our customers solve this problem.”
With that vision, SDL is actively putting forward the five future states of content (Create, Organize, Agile, Secure and Best Salesperson) and a Global Content Operating Model (GCOM), which are key for organizations to better manage their content on a global scale – while reducing costs.
So what does the enterprise need to prepare for this exciting future? Jim stressed the agility aspect required in across the content supply chain, which can be leveraged to effectively reuse content. He gave the example of a Swiss manufacturing company, SKF, who managed to deliver 57 localized sites in 37 languages in just under a year, and it achieved 50% translation reuse across localization – leading to 15% increase of customer inquiries year-on-year (and up to 70% more traffic to localized websites).
In the areas of creation and organization, Jim explained the importance of Machine Learning (which can both understand content and help write it), and of AI (which can assist in organizing content and tailoring it to the right channels).
After highlighting SDL’s commitment to security in its software, Jim also explained how “Sales organizations that effectively use social media on all channels are 78% more effective than those who are not,” illustrating how content is a best-seller, closing in on the fifth state of content.
The power of Linguistic AI
Our next session was led by Mihai Vlad, SDL’s VP of AI and Machine Learning, where he introduced Linguistic AI, and shed more light on the whole AI vs. Human debate – which is an engaging topic for many industries when it comes to translation.
How accurate has the technology become over this past year? 2018 was a year when we reached a magic number of 95% accuracy, a threshold reached by most industries in the fields of image recognition, speech recognition or transcripts*.
The great improvement in quality comes from a paradigm shift, changing from a rule-based system to a framework centered around Machine Learning that directly learns from mistakes, or Deep Learning, where even wider networks are used.
However, in the field of language, measuring accuracy is a much harder task – Mihai illustrated that by showing the audience a simple example of how we relate to a word just by looking at an image. This task requires computers to analyze millions of bytes of information, whereas our minds already have a stock of reference images which helps to directly link to language.
Exploring the base principles of machine translation, Mihai quoted the American scientist Warren Weaver, who in 1949 said: “When I look at an article in Russian, I say: “This is really written in English, but it has been coded in some strange symbols. I will now proceed to decode.” Close to 70 years later, we achieved this with computers by encoding words into numbers, which are then encoded into strings. Analogies are then extracted, which can link similar words together, or identify singularity and plurality, even going all the way in exposing sociological bias such as the word “doctor” (categorized in the “male” gender category, which can fortunately be corrected by math).
Linguistic distance between languages is also a factor in the accuracy of Machine Translation. Together with the English-Russian combination, which SDL has successfully cracked in its Neural Machine Translation (NMT) 2.0, English-Japanese also proves to be challenging.
To show the progress made with NMT 2.0, Mihai challenged the audience to take a brief Turing test, and judge from a few select translation examples in Japanese on whether the translation was made by a human or by a machine. The results were successful in making the audience doubt several times, even splitting opinions half the time, when the translation examples were indeed made by a machine.
This interactive example successfully showed the jump in accuracy achieved by NMT 2.0, which is estimated to be 25% more accurate than NMT 1.0, itself 30% more efficient than the original Statistical Machine Translation (SMT).
Nevertheless, evaluating language remains a complex task as it involves a subjective aspect, however our team at SDL remains committed to push the boundaries of Machine Translation.
In part two of this blog, we recap the final sessions, and look at practical examples of how Machine Learning can transform a business.