Following our launch of KantanNeural™ engines as part of our KantanFleet™ repository of pre-built MT engines, we received a number of questions and interest around the product. To address these questions, we asked Tony O’Dowd, CEO and Chief Architect of a few questions about the Neural Machine Translation engines on KantanMT, the features and benefits of these engines and the impetus behind launching KantanNeural.

KantanMT: Before we begin, for our readers who missed our press release on KantanNeural, could you give us a gist of what it is all about?

Tony: KantanNeural is a new technology that we have introduced to the KantanMT platform, which will allow the KantanMT community to build, improve and deploy Neural MT in conjunction with Statistical MT solutions. While Statistical MT is currently the state-of-art for high quality MT solutions, Neural MT shows tremendous promise and possibilities. By incorporating KantanNeural into the KantanMT platform our clients can experience these benefits for themselves.

KantanMT: What was the push behind creating these KantanNeural engines? Do you think Neural Network technology in MT will be the way forward for us?

Tony: Our main motivation in the release of KantanNeural was to ensure that our clients are the first to experience the benefits of Neural MT within their production workflows. Additionally, we are addressing the huge potential of Neural MT systems across different verticals and markets.

Our initial tests on KantanNeural indicate that translation quality for highly complex languages has improved and we now want to deliver this benefit to our community so they can start working with Neural MT approaches themselves.

There is a lot of interest in the language industry surrounding the application of Neural Network technology in MT, and I believe our KantanNeural engines will help fill this gap.

August 6, 2016Why Neural Machine Translation (NMT) Might be the Next Big Step for the Industry.png

KantanMT: Is the quality of the KantanNeural engines better than the Statistical Machine Translation (SMT) engines?

Tony: In our preliminary testing, we are seeing improvement in translation fluency and adequacy, especially in complex languages such as Chinese, Japanese, German and Korean. In fact, we added a new Project Type to KantanLQR™, called A/B Testing to enable users to compare Statistical vs Neural MT engines.

Read More

We were pleasantly surprised to find that Neural engines performed very well compared to identical SMT engines. This A/B test of SMT and Neural MT was sufficiently good enough to refocus our development efforts and accelerate the roll-out of KantanNeural for our clients.

KantanMT: Are there any issues your team faced while building these engines?

Tony: As with any new approaches, there are always advantages and disadvantages to overcome and alleviate. For example, unknown words are still problematic, terminology management within a Neural network is challenging and not as accurate as traditional SMT systems and then the speed of training and decoding using Neural MT approaches is not as fast or agile as the more traditional and state-of-the-art SMT solutions.

The cost of running Neural systems on highly expensive GPU systems is a significant economic barrier and one in which we are making rapid progress on.

KantanMT: How soon can our clients begin building their own NMT engines?

Tony: Our plan is to launch a fully customisable KantanNeural solution during Q1 2017. This system will be fully featured and allow the KantanMT community to build, improve and deploy Neural engines. We are defining our current offering as a fully featured BETA release, which we shall use as a baseline for further releases later in the year.

KantanMT: What’s next in the horizon for KantanNeural?

Tony: Our focus throughout 2017 will be to roll-out more features within the KantanNeural solution to address needs such as automated scoring of test sets (akin to BLUE but tuned for Neural MT systems), extend KantanAnalytics to support Neural methods and to provide Quality Estimation Scores for Neural translations, incorporate better terminology and unknown word management and to continuously enhance KantanLQR to provide new and novel ways of comparing SMT and NMT translation outputs.

KantanMT: Thank you Tony!

If you would like to know more about KantanNeural,  watch the webinar recording with Tony and Dr Dimitar Shterionov, KantanLabs lead, where they provide an overview of KantanNeural and talk about the new A/B testing feature on KantanLQR.

Interested in knowing about where the language industry is heading? Watch our previous webinar with Tony and Kirti Vashee, here or download our white paper, ‘7 MT Trends in 2017,’ here.