Seeking business models in neural machine translation
I don’t know about you, but I got extremely tired of posts about neural machine translation, which is exactly why I decided to write another one (paradox #1). I promise: no talking about artificial intelligence, deep learning, neural networks or vectors this time. The aim of this article is to take the topic to another level and to shift the focus from technical specifications to possible business models of (neural) machine translation. What’s the business in MT after all?
This article is the conclusion of my very own experiences as person in charge of many MT projects on the one hand, and a set of research papers submitted by University of Antwerp’s Master in Translation and Interpreting students (2017-2018) as part of the module Introduction to Localization on the other. The students were asked to compare rule-based machine translation (oldest type of MT) and neural machine translation (newest type of MT) in detail, to think about their future from a business point of view and to come up with ideas to make money with MT. The MT engines that were used to carry out this research project were basic Systran engines provided by Yamagata Europe. Thank you Björn van Brunschot, Sarah Lacasse, Marlinde Van Loon, Darius de Mahieu, Lauren Van Noten and Jana Roothoofd for your contribution.
1. Will NMT rule out other types of MT?
NMT engines are capable of producing amazingly fluent and human-like translations, but does that mean that other types of MT are slowly disappearing from the MT landscape?Darius de Mahieu states that the choice of a specific MT technology still depends on the volume of training data one disposes of, the volume of data one wants to translate and the type of content that needs to be translated. In the case of lots of training data, a well-trained statistical engine might perform better than a neural engine. In the case of smaller volumes of translatable content that consist of short and predictable strings with fixed terminology, RBMT could still surpass NMT and provide consistent, accurate and readable translations. Björn van Brunschot also points out that translating a document with an NMT engine takes a lot longer than translating the very same document with an RBMT engine, which reveals an obvious weakness of NMT: processing performance.
Sarah Lacasse and Jana Roothoofd agree that the overall quality of NMT is probably the best, but the outcome is not flawless and might contain some serious problems. NMT tends to do crazy and funny things like repeating certain words several times or omitting entire text parts. This phenomenon is referred to as “neural babble” or “neuro-babble”. Furthermore, the fluency of neural machine translations sometimes camouflages genuine translation errors, so the biggest asset of NMT is actually one of its main dangers (paradox #2).
Despite NMT’s imperfections, we must acknowledge that research on RBMT has practically come to an end and that research on SMT is destined to end soon as well. Marlinde Van Loon and Lauren Van Noten question whether translation companies are still willing to invest time and money in training RBMT or SMT engines as the quality of NMT is increasing at an incredible pace.
So will NMT rule out other types of MT? PROBABLY.
2. What are the possible business models of NMT?
NMT solutions are all over the place, but how many LSPs are really making money with it? Not a lot I bet. Companies that want to sell MT have to take a lot of things into account. Apart from budget, hardware and software, MT demands technical knowledge and creativity in order to produce some added value behind it. When a company manages to meet all the necessary requirements, the actual implementation of a business model can take place. There are two directions according to me:
MT + post-editing:
MTPE is the most traditional way of integrating MT in translation workflows. Many LSPs and translation buyers opt for this strategy to tackle time and cost pressure as higher volumes of content can be processed in less time and at lower rates.
My favorite type of MT project. Many companies dispose of valuable data that can be used for internal business analysis on the condition that it gets translated – I’m talking about blogs, claim reports, client reviews, financial summaries, etc. If there’s no budget to translate this type of documents (which is practically always the case), then MT is considered to be the last resource (Björn van Brunschot). An LSP that manages to offer high quality machine translations, squeeze in process automation, implement some (semi-)automated pre- and post-processing (Marlinde Van Loon) and perhaps integrate translation memory technology, will definitely get some very challenging MT projects from enthusiastic clients.
So is it possible to implement profitable NMT business models in the current translation industry? DEFINITELY.
3. Is NMT threatening the translation profession?
I think of MTPE as a relatively new discipline that lives alongside traditional translation in the translation industry. More and more companies are training young graduated translators as post-editors. Freelance translators are calling for clear guidelines and specific post-editing training as well, which seems to be enough evidence to confirm my statement. The case of “no-choice-but-MT” projects is even clearer: no MT = no translation.
Some adventurous translators have already added an MT solution to their tool box. They use MT as a tool to retrieve context (if not available), obtain subject-related terminology (if not available) and solve complex translation issues. In other words, they use MT the same way as they use translation memory technology (Marlinde Van Loon & Lauren Van Noten). Allow me to end this chapter with an emphasis from Marlinde Van Loon:
NMT still does not form a threat for the traditional translation profession in 2018, but as MT systems keep evolving, it is very likely that some aspects of the translation process will eventually be taken over by MT and that creative writing will be the benefit that human translators have over MT engines.
So is MT threatening the translation profession? I DON’T THINK SO.
4. Is NMT disrupting the translation industry?
The only initiatives that are making big money with NMT have names like “Amazon”, “Facebook”, “Google” or “Microsoft”. These companies are responsible for the translation of > 99% of all the words on this planet, which means that the targeted business of every language service provider is reserved for companies that position themselves far away from the translation industry (paradox #3). Terrifying fact if you ask me. A company like Google has the budget, the hardware, the people and, more important, the data to destroy any other MT initiative. The admirable efforts from smaller companies to develop NMT systems are worth mentioning, but will never be able to compete with those of the tech giants.
From a “the glass is half empty” point of view, one might think that translation will end up being a niche market for the big boys – as a matter of fact the number of companies that use Amazon, Google or Microsoft for internal translation purposes is increasing at a very fast pace. So what is David supposed to do against Goliath? Do my beloved “no-choice-but-MT” projects offer enough added value to keep Google API keys at a safe distance? Or will tech companies wash away the translation industry just like they did with e.g. the video rental industry?
There is of course also a “the glass is half full” approach and I think it applies to NMT as well. Do tech companies intend to do serious business in translation? Of course not. Does Google feel threatened by the rise of e.g. DeepL, a company that positions itself in the heart of the translation industry and provides machine translations that often score better than those provided by Google? Definitely not. Google intends to serve its clients/users in the best possible way and MT is just a part of that strategy. Our clients need a better browser? We’ll give them Chrome. Our clients don’t know how to get to their destination? We’ll give them Google Maps. Our clients want to read our content in their mother tongue? We’ll give them Google Translate. MT is nothing more than a commercial steroid for Google and friends. Eliminating the translation industry is not on their agenda if you ask me.
So is NMT disrupting the translation industry? I invite you to share your opinion.