Transletters. International Journal of Translation and Interpreting, 9(2025), pp. 1-6 ISSN 2605-2954
Moorkens, J., Way, A., & Lankford, S.
(2024). Automating translation. Routledge. 270 pp.
ISBN: 978-10-03381-28-0
Book review
Raghad Alsulami
University College London
Received: 14/01/2025
Accepted: 19/07/2025
! "
Machine Translation (MT) technologies are evolving at an unprecedented pace,
with new advances constantly reshaping the translation landscape. Much of the
latest scholarship around MT has revolved around Neural Machine Translation
(NMT), but they have now extended to newer forms of Artificial Intelligence
(AI) such as Generative AI, broadening the spectrum of MT solutions. In
today’s AI-driven world, it is more important than ever for translators to be
part of the conversation around MT, to be technologically competent, to
adapt, and to critically engage with such inevitable advances. Nevertheless,
translators aspiring to venture into the world of MT may struggle in where to
begin and how to navigate the vast array of resources, many of which are
highly technical and not tailored to their specific backgrounds or needs. This is
where Automating Translation steps in. This comprehensive and compelling
book brings together the essential knowledge and practical guidance that
students of translation and practicing translators interested in MT need —all in
one place. From exploring the history of MT and handling data to building
NMT systems and Multilingual Large Language Models (MLLMs), as well as
evaluating MT performance, it covers it all and more.
The book is structured into 11 chapters. Each chapter opens with a set of
key questions to spark curiosity and ends with follow-up tasks (except for
Chapter 1) and curated resources to support readers in exploring the topics
Raghad Alsulami
2
further. Chapters are self-contained, allowing readers to start at any point
without disrupting their overall reading experience. However, I personally
recommend that readers interested in learning how to build their own MT
systems start with Chapter 2, which focuses on data, then proceed to Chapter 4
for insights into the inner workings of NMT, and later explore the more
advanced topics in Chapters 6, 7, and 10. Following this sequence will, I hope,
develop an incremental and comprehensive understanding of building, training,
and optimizing MT systems. As we see throughout the book, the role of
translators becomes more crucial than ever in this critical data-driven AI age.
The authors thus point to some open issues in MT in the afterword, where
human input is not just supplementary but rather indispensable.
In Chapter 1, the authors begin by taking us on a seventy-odd-year
journey through MT history, starting from Warren Weaver’s 1949
memorandum, moving to rule-based and statistical MT, and leading to the
more recent developments of NMT and Large Language Models (LLMs). The
authors touch on the shortcomings of previous MT paradigms and remind us
at the end of the chapter that even with recent breakthroughs in AI, MT is still
far from being a solved problem, as some commentators would claim. Both
NMT and LLMs are data-driven and in need of large-scale high-quality
resources, something that is not equally guaranteed for the roughly 7,000
languages spoken today. The paucity of resources is one reason for the authors’
belief that AI advances will not replace translators, a stance they firmly
maintain through the entire book.
Having introduced MT in the first chapter, the authors shift their focus in
Chapter 2 to data, the cornerstone of any MT system development. In a very
accessible language, the authors present various sources of data that could be
used for building MT systems including translation memories, open-access
repositories, data harvested from the web, and synthetic data. Data-related
issues such as alignment, toxicity, bias, ownership, and data insufficiency are
also discussed, along with a few data postprocessing steps. Dedicating an entire
chapter to data is a commendable aspect of the book, as it addresses key
questions that readers are likely pondering. Where can one find reliable data?
How much data is sufficient for building a well-performing MT system? Who
holds ownership of the data? And what options are available when large
volumes of high-quality data are simply out of reach? By tackling these
foundational points upfront, the book paves the way for a smoother transition
into the more advanced chapters on building MT systems.
Automating translation
3
Expanding on the discussions from the previous chapter, Chapter 3
revolves around Computer-Assisted Translation (CAT) tools and translation
memories which, as Chapter 2 shows, are a useful source of parallel data to be
leveraged in the development of MT systems and any other data-driven
technologies. Here, readers get to learn about the origins of CAT tools, some
of their primary functions, and their architectures. The authors aptly explain
the distinction between MT and translation memories, which to novice
translators could be quite perplexing. The use of Generative AI within CAT
tools makes an appearance in this chapter, a topic sure to captivate readers
eager to explore AI’s potential within CAT workflows. As with all chapters in
the book, the authors conclude by directing readers to a wealth of valuable
resources for further exploration.
Moving to the inner workings of NMT, Chapter 4 introduces neural
networks, guiding readers through the progression of their architectures —
from feed-forward to recurrent neural networks, culminating in the more
advanced Transformer architecture. The last one has revolutionized the natural
language processing field, including MT, as a result of its attention
mechanisms. The authors nicely illustrate this by showing readers how this
mechanism is used in practice with some examples, all without getting too
technical. Since NMT is still error-prone and not entirely reliable, readers are
presented with a number of persistent challenges in NMT, such as inconsistent
terminologies, hallucinations, and difficulties in moving beyond mirroring the
lexical items of the source text to capturing its cultural and idiomatic nuances.
Evaluation lies at the heart of any meaningful conversation about MT, and
it is Chapter 5 that charts some of the common automatic metrics and human
evaluation methods for assessing MT. For automatic evaluation, the chapter
draws on both the strengths and limitations of string-based and pre-trained
neural metrics. It then proceeds to a discussion on some human evaluation
methods, inter-annotator agreement, crowdsourcing evaluation, among other
relevant topics. The chapter does not leave readers there; in fact, it goes a step
further, introducing readers, who may not necessarily possess technical skills,
to some user-friendly platforms like MATEO1 as a practical starting point for
performing automatic MT evaluation. This can empower translators, the
intended audience of the book, by involving them in the evaluation process,
reducing any potential frustrations caused by technical challenges and fostering
a sense of agency.
1 https://mateo.ivdnt.org [Last accessed: 09/12/2025].
Raghad Alsulami
4
Chapter 6 focuses on the critical decision of whether to build or purchase
an NMT system. Readers get to learn about the three options available: using
freely available MT solutions, seeking bespoke offerings from specialized
companies, and building their own systems. Given that building an NMT
system is a challenging undertaking, the authors explore this option in detail
and offer readers an overview of various open-source NMT toolkits,
explaining their core focus and key features so readers can make informed
decisions and choose a toolkit best suited to their needs. Particular attention is
devoted to one such toolkit, adaptNMT,2 by exploring its architecture,
customization of models, and modes of operation, among other relevant
aspects. The chapter ends with a use case demonstrating the effectiveness of
the adaptNMT toolkit in training an NMT system for a low-resource language
pair. It is worth mentioning here that this toolkit is particularly well-suited for
newcomers to the field of MT as it was initially developed to streamline NMT
development processes; this may offer practical value to the book’s intended
audience, empowering them to navigate the world of NMT with hopefully
greater ease and confidence.
Building on the discussion of self-built NMT systems from the previous
chapter, Chapter 7 offers a step-by-step practical guide to developing MT
systems using Google Colab. The chapter begins with a description of Jupyter
Notebooks, highlighting their key role in machine learning tasks, including MT.
It then presents several cloud-based platforms that support writing and
executing code in Jupyter Notebooks, one of which is Google Colab, the
primary platform discussed in the chapter. The platform’s key features are
showcased with a clear comparison of its plans, making it easier for readers to
find the one that fits their needs best. Moreover, AI is increasingly being
integrated into numerous platforms, and Google Colab is no exception. The
authors thus seize this opportunity to touch upon the potential of AI coding
within the platform in, for example, code completion and natural language-to-
code generation. Having established the core functionalities of Google Colab,
readers are then guided through a step-by-step process for training MT systems
with different models in a way that is both clear and easy to follow. Carefully
curated links and resources are provided to readers as well, with the hope that
they make the most of these valuable tools.
Since human-machine parity has not been achieved yet, post-editing of
MT outputs remains necessary if they are to be used for dissemination
2 https://github.com/adaptNMT [Last accessed: 09/12/2025].
Automating translation
5
purposes. Chapter 8 therefore paints a broad picture of the key facets of post-
editing, one of the earliest forms of human-machine interaction. Some of the
topics covered in the chapter include post-editing guidelines, translators’
attitudes of the process, the effort exerted in post-editing, end-users’
perceptions of post-edited texts, and post-editing literature. One of the critical
points the authors highlight is the impracticality of the distinction of light and
full post-editing guidelines with the current state of MT which calls into
question the relevance of light post-editing. The chapter was focused on
(human) post-editing and could have featured a discussion on automatic post-
editing (by LLMs), but this might not have been feasible at the time of writing
the book as this budding area of research was —and still is— in its infancy.
Moving beyond text, Chapter 9 centers around the use of MT and other
forms of automation in multimedia translation and localization, offering a
glimpse into their application across video games, software, websites,
subtitling, and dubbing. Primarily embraced for its velocity and cost-cutting
potential, automation has undoubtedly made significant inroads into various
forms of media. Nevertheless, the authors remind us of the potential
repercussions on the very heart of this ecosystem: the consumers of the end
products, be they users, viewers, or game players —who, for obvious reasons,
seek a seamless and satisfying experience that careless use of, or overreliance
on, MT could compromise. The authors also give due consideration to the
workers themselves when automation technologies are imposed in their
workflows with no regard to their perspectives and preferences which could
potentially lead to reduced job satisfaction and motivation.
LLMs have emerged as a key topic in discussions about MT, and it is
Chapter 10 that highlights their potential alongside that of MLLMs, prompting
the question: is this the future of MT? Before embarking on their capabilities,
the authors first establish how such AI models are purely statistical in nature
and lack the cognitive abilities or comprehension inherent to human reasoning.
Then, similar to how they learn about building NMT systems in previous
chapters, readers here are guided step by step through the process of creating
translation models by either using a custom GPT or fine-tuning ChatGPT to
meet their personalized needs. The adaptMLLM3 toolkit is also introduced,
showcasing some of its key features and demonstrating a use case in which
fine-tuning an MLLM for a low-resource language pair delivers better
translation performance compared to a baseline MLLM. This is very promising
3 https://github.com/adaptNMT/adaptMLLM [Last accessed: 09/12/2025].
Raghad Alsulami
6
and could help readers experiment with other languages and domains for
which only limited data is available.
Building nicely on the previous chapters, Chapter 11, the final chapter of
the book, shifts focus to broader and more serious ethical issues involving MT
and LLMs, such as environmental and social sustainability, along with concerns
over copyright and the translation data fed into such systems. The authors
thoroughly explain the repercussions of building large models like GPT-4 on
the environment by emphasizing their high energy consumption and
substantial carbon dioxide emissions, which in turn exacerbate climate change
and pollution. Such critical risks could affect us all, and it was such a
commendable facet of the book to bring them to the readers’ attention, as they
are, sadly, underdiscussed in academic circles within the field of translation
studies.4 Briefly touched upon in Chapter 9, concerns over social sustainability
are further elaborated in this chapter, with a focus on translators who are often
tasked with fragmented texts and tedious work, as well as end users who could
be exposed to biased and less diverse translations. Finally, the chapter discusses
MT for good, where MT has the potential to empower vulnerable users such as
asylum seekers and save lives in crises, when employed with care.
Reflecting on this work as a whole, this book serves as a valuable resource
for readers interested in both the fundamental principles underpinning state-
of-the-art MT and the practical steps to building NMT systems and
LLMs/MLLMs. The book also contributes to MT literacy by educating readers
on the useful applications of MT, scenarios where its use is inadvisable, and the
broader repercussions of its use on various stakeholders and the environment.
This will not just inform readers about MT technologies but also encourage
them to become responsible and critical users of these tools. While primarily
intended for practicing translators and translation students, I believe a great
deal of the book could appeal to a wider audience, including translation
scholars and students from other humanities disciplines. Translation educators
who wish to integrate lessons on MT and LLMs/MLLMs into their teaching
may also find many of the materials and tasks to be of considerable value. It is
hoped that, with all the enriching discussions and up-to-date resources
provided throughout the book, readers ready to embark on their journey into
the fascinating world of MT find the process more accessible and enjoyable.
4 The natural language processing community has become increasingly aware of these risks,
and this awareness is reflected in Green AI-inspired initiatives such as the inclusion of
sustainability statements at conferences.
