Multilingualism in Scholarly Communication: How Far Can Technology Take Us and What Else Can We Do?

Data-driven tools and their implications for less widely used languages

Tools such as Google Translate and ChatGPT use data-driven approaches, such as machine learning. For a task such as translation, the data consist of previously translated texts. This means gathering a corpus of texts in one language and their translations into another language. Preferably, these translations have been done by professional translators so that the computer will learn from good examples rather than from crummy ones. However, computers cannot learn from just a few examples. For machine learning tasks, computers need millions, or sometimes even billions, of examples (Pérez-Ortiz, Forcada, and Sánchez-Martínez 2022). For languages that are widely used and for which there is a lot of translation activity, it is relatively easy to find examples of previously translated texts. Taking English and French as examples, these are languages used in many countries around the world, and there is a lot of translation activity between them. In Canada alone, the federal government’s Translation Bureau translated 343 million words for governmental departments, agencies, and Parliament in 2019 to 2020 (Government of Canada 2021), while the City of Ottawa’s French language services translated more than 13 million words in 2021 (Willing 2022). Of course, translation between these two languages also takes place in other regions of Canada, in other bilingual countries (e.g., Cameroon), in the various bodies of the European Union, at organizations such as the United Nations and the World Health Organization, and more. Therefore, it is easy to see how tool developers could compile a suitably large training corpus for translation between English and French. In contrast, other languages have far fewer speakers (e.g., Cree, Welsh). Moreover, even if two languages independently have a substantial number of speakers (e.g., Russian and Urdu), there may not be a significant volume of translation activity between them, meaning that it would be hard to create a large training corpus of translated texts. Languages and language pairs for which many texts and translations are available are described as high resource, while those for which fewer texts and translations can be located are known as low resource. This concept of high- and low-resource languages can also apply to domains and text types. For instance, some subjects are very common, while others—like many research topics—are more specialized. In order for a computer to learn relevant terminology or a particular style, those features need to be included in sufficient numbers in the training corpus. This shrinks the pool of resources even more. Good luck finding millions of examples of texts translated between Tamil and Czech on hydrogeology! Another feature of these tools is that they tend to produce better quality translations between languages that are closely related. The more distant the languages, the poorer the translation quality.

So what does this mean when it comes to using translation tools for scholarly publishing? Essentially, it means that these tools will do a better job of supporting researchers who work with widely used languages, such as English and French, while those researchers who work with less common languages risk being poorly served or not served at all. Google Translate currently supports 133 languages, while Meta AI’s ambitious No Language Left Behind (NLLB) project (Costa-jussà et al. 2022) aims to support 200. These efforts are moving us in the right direction, but they are far from meeting the needs of all the speakers of the world’s more than 7,000 languages. Even so, translation tools could potentially help to inject a greater degree of linguistic diversity into scholarly publishing, if only for some of the more widely spoken languages. But are they doing so?

Supporting change or reinforcing bad habits?

Early evidence suggests that the most common use of language technologies in scholarly publishing is not to support multilingual publishing. For example, some journal editors (e.g., Thorp 2023) have sought to rule out the use of generative AI tools in any stage of manuscript production. In response, Mohamed Seghier (2023) and other scholars for whom English is an additional language counter that using free tools such as ChatGPT or DeepL Translator to help with editing and proofreading in English should be permitted. Violeta Berdejo-Espinola and Tatsuya Amano (2023, 991) go so far as to argue that “reducing the technical and financial burden of editing and proofreading papers for nonnative English speakers would be a substantial step toward achieving equity in science.” Indeed, in a survey of 1,600 researchers, Richard Van Noorden and Jeffrey Perkel (2023) asked respondents to identify what they saw as the biggest benefit of generative AI for research. For more than half of the respondents, “the clearest benefit … was that LLMs aided researchers whose first language is not English, by helping to improve the grammar and style of their research papers, or to summarize or translate other work” (Van Noorden and Perkel 2023, 674).

For their part, Mohammad Hosseini and Serge P. J. M. Horbach (2023) suggest that using LLMs could help editors to overcome reviewer shortages. Because these tools can support peer reviewers with the task of preparing their reports, Hosseini and Horbach suggest that editors could access a larger and more efficient pool of candidate reviewers: “LLMs can also increase the pool of reviewers by opening it up to non-native English speakers (some of whom might be able to use various translation services to read a paper) and feed their opinion/views in broken English to LLMs and ask them to write a more presentable review in English” (4).

Meanwhile, Lynne Bowker, Philips Ayeni, and Emanuel Kulczycki (2023) conducted a systematic review of the literature at the intersection of language technologies and scholarly communication. In all 40 of the studies included their review, English features as one of the languages in the translation pair, and just one-quarter of the studies include a low-resource language. Moreover, in nearly two-thirds of the studies, English is the target language, and the focus is on non-Anglophone scholars using translation tools as writing aids to produce texts for publication in English.

What is striking in all of these examples is that they suggest that translation tools are not necessarily helping to displace English as the key language of scholarly communication. Instead, non-Anglophone scholars are using these tools to reduce the burden of preparing English-language publications, but this is not necessarily creating a genuinely multilingual scholarly communication ecosystem. Instead, the responsibility for translation in scholarly publishing continues to rest on the shoulders of non-Anglophone scholars, while the English language and English-speaking scholars remain in a privileged position.

Other tools and resources to support multilingualism in scholarly publishing

What can we learn from this? One key observation is that technology alone is not enough to achieve or sustain a multilingual scholarly communication ecosystem. Current translation tools, while not perfect, can carry out translation in multiple directions (at least for high-resource languages) and can support tasks such as discovering and reading research that has been written in other languages. Yet these tools are mainly being used to translate out of other languages and into English, thus reinforcing rather than diversifying this largely monolingual ecosystem. Certainly more research is needed to find techniques for better supporting low-resource languages, but there is also a need for policies that shift the responsibility away from expecting speakers of other languages to use translation tools to produce texts in English and towards encouraging the use of these tools to access and engage with research that has been written in other languages.

Beyond technology, what can different actors in scholarly publishing do to better support multilingualism? The Helsinki Initiative on Multilingualism in Scholarly Communication (2019) undertakes advocacy work to encourage policy makers to value and support research in multiple languages. They raise awareness through their seminar series and other activities. Likewise, the UNESCO (2021) Recommendation on Open Science considers multilingualism to be one facet of openness.

Various researchers have begun to share recommendations based on their own lived experiences (e.g., Khelifa, Amano, and Nuñez 2022; Nolde-Lopez et al. 2023; Steigerwald et al. 2022), but these are currently scattered in the literature belonging to different disciplines. Bringing this information together in a single open resource is one of the goals of the Developing Institutional Open Access Publishing Models to Advance Scholarly Communication (DIAMAS) project. DIAMAS is developing an Extensible Quality Standard for Institutional Publishing (EQSIP) for Diamond Open Access, and multilingualism is a key element addressed in the equity, diversity, inclusion, and belonging (EDIB) component of the EQSIP. Team members have combed the literature and compiled an open toolsuite of recommendations for ways in which authors/researchers, peer reviewers, editors and editorial board members, librarians, and journal and book publishers can facilitate multilingualism in scholarly publishing (Bowker et al., 2024). Here is a small selection of recommendations from the toolsuite, where they are accompanied by links to examples and additional information:

• For authors: Practice citation diversity (e.g., citing research published in other languages) and consider including a citation diversity statement with your own articles, including any linguistic limits on literature searches (e.g., searches that have been conducted only in English).

• For peer reviewers: Identify to the editors the languages in which you are able to provide peer review feedback.

• For editors: Provide/recommend guidelines to help authors prepare manuscripts in a reader- and (machine) translation-friendly way. Well-crafted input can lead to better quality translation output from automatic translation tools.

• For librarians: Add multilingual metadata to items to facilitate multilingual searches in library catalogs.

• For journal publishers: Translate abstracts, summaries, and tables of contents into multiple languages.

Another group, Open Scholarly Communication in the European Research Area for Social Sciences and Humanities (OPERAS), has the mission to coordinate and federate resources in Europe to efficiently address scholarly communication needs. OPERAS has a special interest group on multilingualism, which has already contributed to the development of a multilingual platform called GoTriple to support search and discovery in nine languages (with more to be added). In addition, OPERAS is currently exploring the development of a scientific translation service combining open source technological tools and resources with human skills to support the translation process within scholarly publishing (Fiorini et al. 2020; Fiorini 2023). A similar proposal has been put forward by the French language commissioner of Quebec in Canada to better support the country’s French language researchers (Dubreuil, Tremblay-Faulkner, and Parent 2023).

Finally, the global Coalition for Advancing Research Assessment (CoARA) has set out a shared direction for changes in research assessment practices intended to maximize the quality and impact of research (CoARA 2022). To this end, CoARA has established a Working Group on Multilingualism and Language Biases in Research Assessment, which has the dual objectives of raising awareness about the importance of multilingualism in scholarly publishing and providing guidelines for recognizing, rewarding and incentivizing research published in all languages (Donahoe 2024).

Overall then, the picture is encouraging. As demonstrated by this special issue, along with other initiatives, there is clearly an appetite for a more multilingual scholarly communication ecosystem, which can include but not rely solely on technology to diversify its linguistic base. Although the way forward is not yet entirely clear, this challenge cuts across nearly all academic disciplines, so it is important to break down the silos and to learn from and support one another as we strive for improved linguistic equity in research. Let’s keep the momentum going.