The ATA Chronicle - November/December 2020 - 27

Yes, academics and developers do work closely in the case
of MT, but at the moment the academics are coming more
often from computer science departments than from translation
departments. But I agree with Sharon that things are moving in a
positive direction. So-called " action " or " participatory " research,
which originated in the population and public health domains, is
being adopted more widely now and is very relevant to MT
research. In a nutshell, participatory research means that
researchers are taking steps to include the communities that their
work is intended to help more fully in the research process. As
Sharon notes, the idea is that the community (e.g., MT users)
would not be an afterthought but would be more active
participants in the research design process as well. In this way, the
research and resulting tools would hopefully better meet the
needs of users.

L

So, you don't think that MT would benefit from
proactively working with linguists as part of their
development teams? I have admittedly asked developers this
question a number of times, both in the areas of statistical
and neural machine translation, and have typically received
strongly worded answers-which I will not disclose here-
but I would be interested in what you have to say about this.

J

That's an interesting question. I've worked with linguists on
several occasions during my career (mostly on natural
language processing problems other than MT), which has always
been an enriching experience for me. The subject of my PhD was
to explore whether syntax could improve statistical MT. The
problem is that, so far, the impact of linguistic structures has been
relatively limited, and outweighed by the gain from " more data "
or " bigger model. "
In my understanding, both linguists and algorithms try to
discover the regularities (and irregularities) of the language, or
several languages at a time in the context of MT applications. Some
of those regularities will be consistent across domains and others
will change when we switch from one domain to another (e.g.,
conversational language versus news articles). The algorithms have
far more capacities to adapt to the context/domain switch given
that those algorithms have access to the relevant data. Therefore,
linguists would have a hard time to compete with machines on
the tasks where the data is abandoned. However, when we switch
to lower resource tasks, including the translation from very low
resource languages, or some translation for very specific domains,
we would definitely benefit from the linguistic insights, which
could guide MT developers in the design of algorithms.
So, to answer your question, yes, I do believe that linguists
could help with MT development when the data is limited or
sometimes inexistent. If we go beyond MT tasks (which is pretty
well defined), in natural language processing in general, I do
believe that linguists' insights are precious in formulating new
challenging tasks for natural language processing. And this is
how progress is made.

V

The move to data-driven MT seemed to reduce the
importance of linguistics-and of linguists. The
improvements of MT output, thanks to neural MT, could be

S

www.atanet.org

seen as limiting the role of linguists even further. However, there
is another way of looking at it. To move neural MT output to the
next level, the issues that need to be resolved are linguistic
issues (e.g., gender in language, style, register, and cohesion,
etc.). I think it would be a big mistake for MT developers to
assume that this is just a machine learning problem that will be
solved by data.
I believe a study by Pierre Isabelle, who served as the
principal scientist and group leader of the interactive
language technologies group at the National Research Council of
Canada, and his colleagues is a good illustration of what Sharon
is saying.3 This study creates a challenging test set to evaluate
the capacities of various MT systems to handle various linguistic
phenomena. On the other hand, what this work and a follow-up
work4 show is that even if not perfect, current MT systems are
making progress in handling those phenomena.
I think it's not only about the data. The data itself is
multidimensional. Various factors are important, such as the
amount of data, the quality of data, and diversity of data. But it's
also about the algorithms, which evolve and are able to handle
more data and get more out of the data. For example, in 2019,
Naver, which was an early pioneer in the use of user-generated
content, released an update to Papago, its automated translation
app.5 The update allows the user to control the register of
produced translation, including rendering English into honorific
Korean. There is a combination of data and smarter algorithms
behind this feature. So, to a certain extent, some of the problems
cited by Sharon could be partially addressed by more/better data
and smarter algorithms. But we definitely need more challenging
datasets and better evaluation procedures to progress further.
Algorithms such as bilingual evaluation understudy (BLEU)
scores won't be able to trace this kind of progress.

V

J

Is there a threshold for women to get into this field, and
what is the path to become part of it?

Is the threshold specific to MT, or does it apply more
broadly to tech or even science, technology, engineering
and mathematics (STEM)? Are there genuinely fewer women in
the field, or do women just have a lower visibility? For
instance, in academia overall (across all disciplines), there
are nearly as many women as men, but when you get to the
senior positions (e.g., full professor), the men account for
about 70-75% of the posts, and the women just 25-30% (in
Canada, anyway).6
So, the imbalance is not in the total number but in the
distribution, with men all bunched up in the senior ranks
and women all bunched up in the lower ranks. And this is
across all disciplines, so not really tech-specific. Generalized
explanations that are given are that women are penalized by
taking maternity leave, by having or wanting to do more of the
caregiving (both for children and elderly parents), or that they
are paid significantly less and so don't feel motivated to work
twice as hard. But I've seen specific studies on this problem in
academia and specific studies7 on gender pay gap issues, but
I've never seen specific studies on the MT field.

L

American Translators Association

27
http://www.atanet.org
The ATA Chronicle - November/December 2020

Table of Contents for the Digital Edition of The ATA Chronicle - November/December 2020
