Brave new world
The conference, which I now attended for the first time, was an opportunity for me to step in into a new world: the academic side of MT. Indeed, when it comes to Machine Translation, the concern for LSPs, in other words translation agencies, is mainly commercial. As far as my own work is concerned, my focus is very concrete (how to introduce MT into a production workflow?), practical (how to compensate post-editors and invoice customers?) and sometimes psychological (how to educate users on post-editing tasks, and how to convince translators that MT is not a threat?).
It is worth spending some time on that last point, because that’s the major difference between the academic world and the commercial world: in my world, there are still a lot of people who are afraid or openly against machine translation. At the conference, I met people who are passionate about MT, they work for MT, and I am sure many of them dream MT! As a matter of fact, I asked one of the developers I met there to go for a beer after dinner (after all, we were in Prague!), and he replied: “Ok, but just one, because I need to go back to my hotel and start training a new neural system”. I was impressed!
Maybe, maybe not
NMT (Neural Machine Translation) was undoubtedly a hot topic during the conference. I remember two researchers arguing over NMT during lunch. One researcher claimed that NMT is better than SMT (Statistical Machine Translation), while the other one, himself a top expert in NMT, claimed that SMT is better than NMT. Who was right? Anyway, what is “better” for a researcher may not mean the same for me when it comes to deciding the “best” engine for my customer. All the same, that anecdote clearly shows how passionate the people I met are about MT. I wish I would see more passionate people in my world too…
Another interesting topic was Quality Estimation (QE): how to predict the quality of the MT output to avoid sending garbage to post-editors. QE is a great concept! It would help translators come to grips with post-editing. But again, some say QE is a necessary step, while others have demonstrated that QE can be ignored altogether. The same applies to the BLEU score (automatic quality metrics to measure the quality of an MT output): some assert the BLEU score is meaningless if applied to NMT. Their claim makes sense. Yet, BLEU has been the quality metric most mentioned during the conference. This is something I noticed during these three days: no answer is ever definite. Very often answers to questions began with “It depends…”.
It’s all about passion and algorithms
There were a lot of interesting topics, but what I mainly got from the conference is that there are people out there who have time to live their passion. Researchers and developers usually do not have the commercial pressure. Therefore they have time. They have the luxury of doing all kinds of experiments to verify a theory or confirm a phenomenon. They can spend days, weeks, months testing a new language model, optimizing algorithms, comparing alignment methods. It’s all about morphology, n-grams, attention, tokenization, etc. All this is very advanced and in-depth, all very theoretical. One participant even told me “I am an expert in my field of specialization, but the sad truth is that there are probably only five persons in the world who understand my specialization”!). That said, one should not denigrate or laugh at the academic world (as some in my world are tempted to do). Instead, we should pay a lot of attention and above all respect to the researchers and developers. Without their passion and algorithms, Machine Translation would not be where it is today.