Dr. John Tinsley at the Translating Europe Forum 2018
Iconic’s CEO, Dr. John Tinsley, recently took part in a distinguished panel discussion as part of the Translating Europe Forum 2018 in Brussels. The Translating Europe Forum, hosted by the European Commission, provides two full days of useful and relevant information for those involved in the translation arena. The event took place November 8-9, with more than 500 participants and (last year) attracted more than 5,000 viewers via webstreaming.
The panel discussion had so much timely and important information that we have created three separate blog posts as part of our review. We have also included a short (two and a half minute) video snippet, of the significant points which John presented on the panel. Over the three posts we will discus; What are the key ingredients to build a successful Neural Machine Translation (MT) system, MT Use Case and the “Productivity” question, and finally, which Trends to look out for.
MT Background Information
Before we dive into discussing our key ingredients to building a successful Neural MT system though, it might be helpful to give some background information. A couple of years ago, the industry reached a point where machine translation technology had its limits. The capabilities of Statistical MT plateaued; We knew what it could do, we knew what it couldn’t do and we we worked within those parameters. Machine translation technology is a relatively new technology, originating back in the 1950s with “Rule-Based MT”. Not until the early 1990s was there a paradigm shift to “Statistical MT” and only in 2015 did the second paradigm shift to “Neural MT” take place.
Given that there are announcements almost on a daily basis of companies offering “Neural MT”, an important question to address is “What does it take to build a successful Neural MT system?”. We believe there are two key ingredients to build a successful Neural MT system.
Key Ingredients for successful MT:
- CLEAN data. There is more value out of data preparation and cleaning with Neural MT. It can really be a question of less is more where the less data is better quality. Noisy data, the opposite of clean data, can mean bad alignments, poor translations, misspellings, and other inconsistencies in the data used to train the system. Neural MT, compared to Statistical MT, is much more sensitive to noise, and therefore clean data is very important. We discussed this topic in detail as part of our Neural MT Weekly blog series, which you can read here.
- People. Smart people (really smart!), who know what to do with the data, know how to manipulate the data, and know how to manipulate the algorithms if they’re not producing the intended output. This describes Machine Translation Scientists at their best.
One of the big challenges with machine translation is that it is not a “one size fits all” technology. When developing a machine translation system, if the output doesn’t quite meet expectations, then you need the intellectual capability, the people, the experts, who can look at the data, look at the output, perhaps look inside the “blackbox” and work out what’s going on and how to improve that output.
You need to have the expertise to be able to go backwards from a problem and determine why it is happening.
When referring to the ‘black box’, it is not exactly a ‘black box’ of the unknown insofar as there are people developing MT algorithms and they know what they are developing. In the past, with Statistical MT, if your output was unusual, you were able to go back and trace why that happened. You could go into all of the models, and you could see the likelihood that the “unusual” translation was going to occur. We can’t necessarily do that, yet, with Neural MT, but we feel it is just a matter of time. Neural MT is so new, people aren’t working out how to look back retrospectively yet, because there is still so much to do going forward.
You can click on the video below to hear what Iconic’s CEO, Dr. John Tinsley, has to say about the aforementioned “black box” and the two Key Ingredients for building a successful Neural MT system.
In summary …
The landscape of Machine Translation is changing with the many developments of Neural MT happening in real time and in quick succession. We believe there are two key ingredients to building a successful Neural MT system, namely “Clean Data” and “Smart People”, and the quality of those two ingredients is indispensable.
In our next post, the second part of this three part review, we will cover Machine Translation: Use Case and Productivity, while in the final review post we will talk about Trends to look out for in Neural MT.
About Translating Europe
The workshops are part of the Translating Europe project, created to bring together translation stakeholders in Europe. They complement the yearly Forum organised in Brussels and consist of smaller scale events, targeted at specialised audiences within national translation communities. The Translating Europe workshops are organised by Directorate-General for Translation representatives in the EU Member States, often in cooperation with universities of the European Master’s in Translation.