ANALYSIS

Creative writing undoubtedly plays a key role in the translation process.

Following the COVID-19 pandemic, increased demand for translated creative content led to attempted applications of neural machine translation (NMT) to creative texts in an effort to cut cost and time required to translate creative content — such as movies, television series, and video games — into a vast number of languages. Research shows, however, that machine translation post-editing (MTPE) can limit translators’ creativity.

We must ask ourselves then, whether or not there’s a risk that writers’ creativity can also be affected by the extended exposure to machine-translated texts, considering:

A. That content creators are not isolated from their textual environments
B. The increasing percentage of machine-translated content that we consume on a daily basis

Here, I’ll explore the ways in which machine translation (MT) — and exposure to machine-translated content — can impact writers’ creative process. National languages are dynamic, meaning that they are constantly shifting, evolving, and being enriched as a result of their interaction with other languages and cultures. It’s possible that they could also be shrinking in terms of vocabulary, grammar, and syntax structure due to the extended use of machine translation in creative content. I’ll also discuss potential interventions that could be implemented in the education and training of translation professionals, as well as best practices to establish the role of creative writing in translating creative content.

Creativity and machine translation systems

Translation is a creative process that requires thorough understanding, interpretation, and re-creation of the source text by filtering it through the language and culture of the target text. Research suggests that it’s a process as creative as creative writing itself.

According to Randolph Quirk, translation is one of most difficult tasks that a writer can take up, while Horst Frenz places translation somewhere in between “creative arts and imitative arts.” Translation is a much more dynamic process than simply replacing source language words and grammatical structures into the equivalent ones for the target language. As readers of MultiLingual, you likely already know that translation often involves removing linguistic elements of the source text in order to achieve an “expressive identity” between the source language and target language texts.

Since the 1940s, MT has become a dominant research area with some of the first serious research projects taking place in 1951 at the Massachusetts Institute of Technology. The basic challenge for the creation of a reliable system of MT is the fact that “the computer … cannot develop the intellectual strategies required for indirect translation, from the superficial structures of the source text to the deep structures of the source language, from the deep structures of the source language to the deep structures of the target language and from the deep structures of the target language to the superficial structures of the target language,” as researchers Frideriki Batsalia and Elenia Sella-Mazi put it.

Current research focuses on machine-learning applications using natural language text corpora and more specifically on improving neural MT systems and natural language production systems, such as GPT-3 and Gopher, that produce natural language (with correct grammar and syntax, and even impressively idiomatic language) using algorithms and configuring vast datasets.

These are complex mathematical functions that identify patterns in the text corpora used to train them, enabling forecast patterns and producing natural language. The quality of the produced texts depends highly on the quality of the text corpora used, and they do not perform well with new texts (that are different than the text corpora they have been trained with).

According to an announcement made by Google in 2012, the volume of text being translated using Google Translate in one single day corresponds to the volume of text that would be found in one million books. According to more recent research, it is estimated that MT is responsible for 99.5% or more of all language translation done on the planet on any given day.

As unrealistic as this percentage may seem, it is actually not far from reality, especially if we take into consideration the unprecedented peak in the requests for localized creative content since the COVID-19 pandemic (streaming platforms, movies, TV shows, video games). This led to the use of neural machine learning systems to this type of creative content as well, with the goal of reducing the cost and time needed for localization efforts.

Language: static or dynamic? The case of NMT

Octavio Paz argues that “all texts are part of a literary system descended from and related to other systems. Every text is unique and at the same time it is the translation of another text. No text is entirely original because language itself, in its essence, is already a translation: firstly on the non-verbal world and secondly, since every sign and every phrase is the translation of another sign and another phrase. However, this argument can be turned around without losing any of its validity: All texts are original because every translation is distinctive. Every translation, up to a certain point, is an invention and as such it constitutes a unique text.”

There cannot be one single method according to which the translator should work but rather illimitable sets of different rules that constantly change, placing the text in a continuous dialectical relationship with these sets of rules. As a result, it is not possible to have the ideal translation, in the same way that it is not possible to have an ideal poem or an ideal novel. Every translation should be assessed taking into consideration the process of its creation as well as its contextual function.

Assuming that we assign 12 different translators to translate the exact same poem, the result will be 12 different translated versions. Through comparison between these 12 different translated versions we will be able to locate what Popovič defines as the invariable core of the original poem, which has not been affected by the 12 different ways of expression). At least this was a common place in the ’80s, when the MT technology was still at an early stage of development. It would be an interesting research topic today, in the era of the dominance of MT, to examine how many different translated versions would occur if we were to perform the same experiment.

National languages are considered to be dynamic, shifting, evolving, enriched through connections with other languages and cultures. The extended use of MT in creative content could also possibly cause them to limit in terms of vocabulary, grammar, and syntax structure. The translator connects two different language systems: one that is already expressed and static and another that is still dynamic and adjustable. The translator is facing a starting point and is processing intellectually a point of arrival.

The above relationship can be graphically represented in Figure 1.

Figure 1: A visual representation of a static to dynamic language system.

Figure 2: A visual representation of a static to static language system.

The act of processing, modifying and correcting a text that has been translated using MT — MT post-editing (MTPE) — has been found to significantly limit the translator’s creativity. This makes sense, as the creative process becomes severely limited when the translator chooses among predefined, already structured options and edits them in terms of grammar correctness and accuracy. In MTPE, the language system of the target text is already expressed, static, and predefined.

The above relationship can be graphically represented in Figure 2.

In 2022, at the European Commission’s Translating Europe Forum, a panel of speakers attempted to define translation quality in the era of NMT, MT post editing, and AI. One of the keynote speakers offered a pretty agile and inclusive definition, stating that “quality is giving the customer what the customer needs.”

Beyond this industry-based, customer-oriented definition, it’s important to take a step back and examine the bigger picture. Target language texts produced using MT systems are much more than products ordered by a customer and should not be treated as such. The fast-paced localization industry is oriented toward maintaining turnaround times as agile as possible and workflows as cost-effective as possible. The final product, however, has a much greater impact on consumers compared with those of other industries. 

Figure 3: The relationship among all factors involved in the transcreation process.

According to the Estonian Semoticist, Juri Lotman, “No language can exist unless it is steeped in the context of culture and no culture can exist which does not have at its center the structure of natural language.” Naturally, then, the question of how natural language is affected by MT rises.

The text corpus that developers use to train the MT system directly affects the quality of the produced natural language. If, in this relationship, we replace the MT system with the internet user/reader and we replace the text corpus with the machine-translated content that we consume on a daily basis, what will be the expected quality of the produced written and oral speech? Is it possible that the constant exposure to machine-translated content can also affect the natural language we produce as individuals?

To answer this question, let’s go back to the Holy Roman Empire in the 18th century. The writer and all-around polymath Goethe knew very well the importance of the translation process in promoting and establishing world literature, and he claimed that every language can be enriched through the translation process and its interaction with the foreign elements of other language communities.

In his poem Ein Gleichnis, Goethe narrates a story of cutting a bouquet of flowers from a field. On his way back home, the flowers began to wither, but once he carefully placed them in a glass of fresh water, a miracle happened: Their colors became as vivid as when they were sprouting in the field. Goethe compares this to the act of listening to one of his poems translated into a different language. The fresh water that gives life to the withering flowers is no other than the creativity needed to convey the meaning of the source text to the translated text.

Figure 4: A visual representation of the fluency trendline. Source: Re-defining Creativity in Localisation: Ways to Unleash Hidden Potential by Despina Pilou (2023)

Figure 5: A visual representation of the style trendline. Source: Re-defining Creativity in Localisation: Ways to Unleash Hidden Potential by Despina Pilou (2023)

With the modern translation practices that significantly limit the translator’s creativity and the industrialization of localization that neglects the creative aspect inherent in the translation process, is it possible that we are facing the grave danger of water scarcity? Is it possible that national languages become more limited as a result of MTPE and the extended exposure of average users to industrialized, machine-translated texts rather than being enriched?

Figure 3: Prompt design/engineering window of the OpenAI Translator plugin in Trados Studio 2022

Redefining localization: training interventions and best practices

Translation studies have benefited in the past through the interaction with relevant scientific fields, such as linguistics, narratology, grammatology, scientific research on bilingualism and multilingualism, and child language-learning.

In this interdisciplinary context, the field of neuroscience seeks to unravel the mysterious ways in which human translation works, as the neurological mechanisms involved in translating remain one of the chief known-unknowns. Research largely shows that the human brain does not work in the same way as a computer processing symbols. During the translation process, the body and the experiences acquired through it play a central role, our first-hand experience is being recalled and the brain-body system is activated.

Creative writing theory may also offer much to the field of translation studies and should be part of any translation undergraduate and postgraduate curricula. In modern translation industry practices, creative expression tends to become extinct and emphasis is put on standardized language, the use of specific repetitive syntax structures and terminology, any deviation from which is being penalized as an error.

Figure 6: Accuracy trendline. Source: Re-defining Creativity in Localisation: Ways to Unleash Hidden Potential by Despina Pilou (2023)

Figure 7: A visual representation of the kudos earned trendline. Source: Re-defining Creativity in Localisation: Ways to Unleash Hidden Potential by Despina Pilou (2023)

Copywriters who write source texts that will later go through the MTPE process typically have to follow strict guidelines ensuring that source texts will be compatible with the MT systems: That means simpler grammar and syntax structures, repetitive phrases, sentences that are as short as possible, repetition of nouns instead of using adjectives, avoiding conjunctions and secondary clauses, avoiding the use of passive voice, avoiding the use of idioms. Machine-translated texts therefore have similar grammar and syntax structure, as well as style and the process of translating ceases to be an original creation.

Through exposure to such texts it stands to reason that truly original and novel creation is difficult to achieve; balancing this lack is crucial and can be made possible by implementing creative writing theory in the localization workflow.

The importance of applying creative writing theory in literary translation is self-evident; the localization industry also seems to have grasped the importance of creative writing theory in the case of transcreation, used mainly in advertising and marketing, as it requires a deep knowledge of the cultural environment of both the source and target language — as well as strong creative skills.

Transcreation refers to a message that is adapted from a source to a target language without losing its style, tone, and context, and evoking at the same time the same emotions to the target-language readers as the ones intended by the author of the source text. In other words, a mirroring process. But in The Cambridge Introduction to Creative Writing, David Morley suggests that language is a shifting and evolving system, in which some words are charged with particular meanings in their host language, but that does not entail their carrying those associations into another language.

Figure 3 represents my attempt to represent the relationship among all factors involved in the transcreation process.

In terms of cardinality, the diagram suggests a dynamic, constantly shifting relationship of many-to-many (as opposed to a one-to-many or a one-to-one relationship).

The term transcreation is a portmanteau coined to combine the words “translation” and “creation” to highlight the significant degree of creativity required in the translation process. It’s clear that the localization industry does appreciate the importance of creativity in translation — some gaming localization companies include creative writing exercises when assessing new translator. Despite this understanding, however, many believe that creativity is an inherent talent — and by extension, that some translators are gifted, while others are not. Applying creative writing theory in universities and major localization companies is far from common practice, after all.

In the same way that MT systems are trained using text corpora, so should the translator be trained in active reading, literature, poetry, theater, and cinema scripts, and face the blank page that calls for the creation of new worlds, doubting standard practices, and suggesting new norms. Imagination, offering creative solutions to translation problems, and originality are indispensable skills that can be nurtured with style exercises, creative writing exercises, and researching (and possibly doubting) creative writing theory.

A pilot project conducted in 2022 at Orco S.A., a medium-sized language service provider, actually did show a clear correlation between creative writing exercises and the performance of translators. In this pilot, I combined the principles of translation theory and creative writing theory and applied them in practice to produce training workflows targeted at unlocking the creative potential of the company’s inhouse linguistic team. The pilot’s aim was to address an alarming tendency noticed among linguists who were involved in MTPE: They seemed to rely too much on the translation memory results, even when they were not working on MTPE, and the forms of expression they were using seemed to be very limited compared to linguists not involved in MTPE.

Linguists involved in MTPE also scored low in style and fluency and never received kudos for their work during linguistic quality assurance (LQA), so the pilot was also expected to offer insight on the correlation between creative writing tasks and translation quality. We monitored the linguists’ progress and evaluated them for a period of six months, examining LQA forms and looking out for noticeable errors.

Results are shown in Figures 4, 5, 6, and 7.

Significant improvement was reported in the fluency, style, and “kudos earned” trendlines. A dramatic drop in accuracy was also reported, that deteriorated at about four months. However, accuracy scores were restored back to normal at about six months after the project kick-off. The results of the pilot demonstrate not only that there is a strong correlation between creative writing exercises and the produced translation quality but also that creative skills are not inherent: Creativity in localization is an aptitude that can be nurtured.

Clearly, there is room for more research here. It is important to note, however, that such initiatives should be encouraged, both in academia and in the localization industry, to balance the fact that, “We tend to be mechanical when what is called for is creativity,” as David Bohm so eloquently put it.